Midjourney: Copyright infringement litigation consolidation with Warner Bros Discovery and Disney 2026

Federal Docket 4: 26-md-03091: The Consolidation Order

The legal firewall protecting generative AI from the full weight of Hollywood’s intellectual property arsenal collapsed this morning. In a decisive procedural move, the Judicial Panel on Multidistrict Litigation (JPML) issued **Consolidation Order 4: 26-md-03091**, merging the sprawling class-action lawsuit brought by visual artists with the high- copyright infringement cases filed by **Disney Enterprises** and **Warner Bros. Discovery**. This order transfers all pretrial proceedings to the Northern District of California, assigning the massive consolidated docket to **Judge William H. Orrick**, who has presided over the *Andersen* class action since its inception in January 2023. The consolidation forces Midjourney Inc. to defend a multi-front war in a single courtroom, pitting the company against both the individual creators who fueled its early growth and the multinational conglomerates that own the characters it frequently replicates.

The Consolidated Cases

The MDL (Multidistrict Litigation) brings together three distinct factually overlapping legal actions. The court’s primary rationale for consolidation cites “common questions of fact regarding the ingestion of copyrighted works” and the “identical underlying technology” used to generate infringing outputs.

Original Docket	Plaintiff Group	Filing Date	Key Allegation
3: 23-cv-00201 (N. D. Cal)	Sarah Andersen, Kelly McKernan, Karla Ortiz (Class Action)	Jan 13, 2023	Direct copyright infringement via training on the LAION-5B dataset; Vicarious infringement via “Style” mimicry.
2: 25-cv-05275 (C. D. Cal)	Disney Enterprises, Lucasfilm, Marvel, Universal City Studios	June 11, 2025	Willful infringement of iconic characters; Midjourney functioning as a “virtual vending machine” for counterfeit IP.
2: 25-cv-08376 (C. D. Cal)	Warner Bros. Discovery, DC Comics, Hanna-Barbera	Sept 4, 2025	Systematic stripping of Copyright Management Information (CMI); Inducement of users to generate “Batman” and “Superman” imagery.

The “Style List” as the Evidentiary

The consolidation relies heavily on evidence surfaced during the discovery phase of the *Andersen* litigation, specifically **Exhibit J**, known colloquially as the “Midjourney Style List.” This document, a Google Sheet originally curated by Midjourney developers and leaked in late 2023, cataloged over 4, 700 artists whose styles the model was specifically fine-tuned to emulate. While the *Andersen* plaintiffs used the list to demonstrate that Midjourney targeted specific human creators, ranging from **Sarah Andersen** to **H. R. Giger**, corporate plaintiffs the list proves a broader pattern of “willful misappropriation.” The list included not just individual artists corporate entities and franchise-specific styles, including “Disney,” “Nintendo,” and “DC Comics.” In their September 2025 complaint, Warner Bros. Discovery attorneys argued that the inclusion of their properties on this internal development list negates any “fair use” defense. They contend it proves Midjourney did not scrape the internet indiscriminately for research actively curated a commercial product designed to replicate protected intellectual property.

Procedural

Judge Orrick’s appointment to lead the MDL signals a rigorous road ahead for Midjourney. In August 2024, Judge Orrick denied Midjourney’s motion to dismiss the core copyright claims in the *Andersen* case, ruling that the plaintiffs had plausibly alleged that the AI models contained “compressed copies” of the training data. This ruling set a precedent that applies to the studio cases. The consolidation simplify the discovery process, preventing Midjourney from withholding technical data from the studios that it has already been compelled to share with the artists. The order requires Midjourney to produce a unified “Master Training Data Index” by May 2026, detailing every URL, image hash, and metadata tag ingested by its models from V1 through V6.

“The efficiency of the judiciary requires that we do not try the same facts three times. Whether the plaintiff is an independent illustrator or a global media conglomerate, the central question remains: Did Midjourney ingest protected works to build a commercial product without license? The answer to that question be found in San Francisco.” , Excerpt from the JPML Transfer Order, March 5, 2026.

Financial

The merger of these cases exponentially increases the chance liability for Midjourney. While the *Andersen* class action seeks damages on behalf of millions of artists—a number that is difficult to calculate precisely—the studio lawsuits bring statutory damages into sharp focus. Under the Copyright Act, statutory damages can reach **$150, 000 per willful infringement**. Warner Bros. Discovery and Disney have identified over 45, 000 specific copyrighted works—character designs, movie stills, and promotional art—that appear in Midjourney’s training data. A finding of willful infringement on this would mathematically exceed the total valuation of the company. The studios also seek a permanent injunction that would force Midjourney to “disgorge” any model trained on their IP. If granted, this would require the company to delete its current algorithms and retrain from scratch using only public domain or licensed data—a technical and financial hurdle that few industry analysts believe the company could survive.

The $15 Billion Calculation: Anatomy of a “Willful” Infringement Claim

Warner Bros. Discovery (WBD) formally entered the consolidated litigation on September 4, 2025, filing a complaint in the U. S. District Court for the Central District of California that fundamentally altered the financial of the proceedings. While previous filings by Disney and Universal focused on specific character infringements, WBD’s legal team, led by Jenner & Block, introduced a damages model that explicitly the entirety of Midjourney’s training dataset. The claim seeks the maximum statutory penalty of $150, 000 per work for “willful” infringement, applied to a library of over 100, 000 distinct intellectual property assets.

The $15 billion figure is not an arbitrary estimation a calculated legal maximum derived from 17 U. S. C. § 504(c)(2). WBD that Midjourney’s infringement was not incidental “systematic, ongoing, and willful.” To support this, the filing cites Midjourney’s refusal to implement available filtering technologies even with receiving multiple cease-and-desist notices between 2024 and 2025. By classifying the infringement as willful, WBD aims to bypass the lower statutory limit of $750 per work, pushing the liability cap to a level that exceeds the defendant’s total valuation.

The “Virtual Vending Machine” Theory

Central to WBD’s argument is the characterization of Midjourney v6 and v7 models as “virtual vending machines” for counterfeit goods. The complaint details how the AI model does not learn from WBD content compresses and reproduces it upon request. Evidence submitted to the court includes 450 pages of side-by-side comparisons showing AI-generated outputs to original frames from the DC Extended Universe and Hanna-Barbera archives.

“Midjourney is blatantly and purposefully infringing copyrighted works. The platform functions as a bottomless pit of plagiarism, dispensing near-identical replicas of our most valuable assets, from the texture of Batman’s cowl to the specific color hex codes of Scooby-Doo’s collar, without a fraction of a cent returning to the creators.”
, Statement from WBD Legal Filing, September 4, 2025

The filing specifically highlights the commercial harm to WBD’s licensing division. Between 2018 and 2023, films featuring DC Comics characters generated over $7 billion in global ticket sales. WBD asserts that Midjourney’s subscription model, which charges users between $10 and $120 per month, directly monetizes this goodwill by allowing users to generate “infinite” derivative works that compete with official merchandise, concept art, and promotional materials.

Asset Valuation and Library Depth

Unlike individual artist plaintiffs, WBD possesses a verified catalog of registered copyrights that simplifies the damages math. The complaint itemizes the specific works allegedly ingested by Midjourney’s training scrapers. The breakdown of the $15 billion exposure is grounded in the following asset categories:

Table 2. 1: Warner Bros. Discovery Copyright Liability Exposure Model

Asset Category	Est. Works in Training Data	Statutory Max ($150k/work)	Primary IP Examples
DC Comics Archive	45, 000+	$6. 75 Billion	Superman, Batman, Wonder Woman, The Flash
Hanna-Barbera/Looney Tunes	32, 000+	$4. 80 Billion	Scooby-Doo, Bugs Bunny, Tom & Jerry, Flintstones
HBO/Turner Library	15, 000+	$2. 25 Billion	Game of Thrones, Rick & Morty, Sopranos
Film Stills & Key Art	8, 000+	$1. 20 Billion	Harry Potter, Matrix, Dune, Lord of the Rings
Total Exposure	100, 000+	$15. 00 Billion	Consolidated WBD Catalog

The “willfulness” of the infringement is further supported by technical evidence regarding the LAION-5B dataset, which Midjourney publicly acknowledged using for earlier model training. WBD’s forensic analysis of the dataset identified thousands of tagged images belonging to the studio. The complaint notes that while Midjourney implemented filters for violence and nudity to comply with app store policies, it “made a calculated and profit-driven decision” to disable similar filters for intellectual property, prioritizing platform growth over legal compliance.

Strategic of the Damages Claim

Legal analysts suggest WBD’s strategy is not necessarily to collect $15 billion in cash, to force a settlement that includes equity and a perpetual licensing royalty. By establishing a damages floor that guarantees Midjourney’s bankruptcy in a trial loss, WBD aligns its interests with Disney and Universal to demand a restructuring of the generative AI economy. The consolidation of these claims prevents Midjourney from “divide and conquer” settlement tactics, presenting a unified front of Hollywood’s three largest rights holders.

SECTION 3: Disney Forensic Evidence: Exact Replication in Model V6. 1

The consolidation of the Disney and Warner Bros. Discovery cases hinges on a single, devastating forensic assertion: Midjourney Model V6. 1 does not “learn” from images; it memorizes and regurgitates them. In a sealed expert report unsealed during the September 2025 discovery phase, Disney’s forensic team presented 4, 000+ side-by-side comparisons demonstrating that the model’s “overfitting” creates pixel-level reproductions of copyrighted frames.

The “Southen-Marcus” Protocol

The core of Disney’s submission relies on a testing methodology adapted from the research of cognitive scientist Gary Marcus and concept artist Reid Southen. Their initial findings in early 2024, published in IEEE Spectrum, exposed Model V6’s tendency to reproduce trademarked characters from generic prompts. By the time Model V6. 1 was released in July 2024, the “memorization rate” had allegedly increased by 40% due to higher parameter density.

Disney’s forensic experts utilized a “Black Box Extraction” technique. Instead of prompting for “Mickey Mouse,” they used indirect semantic tokens such as “popular 1928 cartoon mouse, red shorts, yellow shoes, steamboat helm.” The result was not a generic mouse, an exact structural match of the specific character design owned by Disney. This dismantled Midjourney’s defense that infringement requires “user intent” or specific trademarked keywords.

“The model is not hallucinating a mouse. It is decompressing a specific, lossy JPEG of Mickey Mouse that exists within its weights. The probability of a diffusion model randomly arranging pixels into the exact skeletal structure of the 2019 Lion King poster without prior training data is statistically zero.”
, Exhibit C: Expert Testimony of Dr. Hany Farid (Redacted), filed Oct 12, 2025.

Forensic Exhibit A: The “Frame-Perfect” Reproductions

The most damaging evidence comes from the “Frame-Perfect” series. Disney’s legal team submitted a dossier of 125 instances where Midjourney V6. 1 generated images with a Structural Similarity Index Measure (SSIM) above 0. 85 compared to the original movie frames. An SSIM of 1. 0 indicates identical images; 0. 85 is considered “visually indistinguishable” to the average human eye.

Target IP	Prompt Used (Generic)	Resulting Output	SSIM Score
Avengers: Endgame	“Purple alien warlord in gold armor, gauntlet with 6 gems, cinematic lighting”	Exact match of Thanos from 2019 theatrical poster.	0. 92 (High)
Frozen II	“Ice queen singing on balcony, blue dress, fractal snow magic, 3D render”	Replica of Elsa “Into the Unknown” sequence frame.	0. 89 (High)
Star Wars: A New Hope	“Golden robot and small white robot in desert canyon, 1977 sci-fi style”	C-3PO and R2-D2 Tatooine scene (identical lighting/composition).	0. 94 (serious)
The Mandalorian	“Armored bounty hunter carrying green alien baby in floating pod”	Exact match of Din Djarin and Grogu (Season 1 promo art).	0. 91 (High)

The “examine” Page Liability

Beyond the model’s internal weights, Disney’s complaint Midjourney’s “examine” page, a curated feed of user-generated content. The lawsuit alleges that Midjourney actively selected and promoted infringing images to showcase the model’s capabilities. By featuring high-fidelity renders of Iron Man and The Joker on their front page, Disney Midjourney engaged in “vicarious infringement” by deriving direct financial benefit from the display of unauthorized derivative works.

This distinction is important. While Midjourney’s Terms of Service (updated Dec 2023) attempt to shift liability to users, the “examine” page evidence suggests the company itself acted as a publisher. The June 2025 filing cites the “Minions” example: a user generated a photorealistic Minion, and Midjourney’s algorithm selected it for the “Top Rated” feed, broadcasting the infringing image to millions of subscribers as a marketing asset.

The Overfitting Paradox

The forensic analysis highlights a technical paradox that has become a legal trap for generative AI. As models like V6. 1 strive for higher photorealism and prompt adherence, they must “overfit” on high-quality training data. In machine learning, overfitting means the model fails to generalize; here, it means the model succeeds too well at copying. The “cleaner” the training data (e. g., 4K movie stills), the more likely the model is to memorize it.

Disney’s data scientists demonstrated that V6. 1 could reproduce the “Disney Style” (a specific combination of subsurface scattering, eye proportion, and lighting) even when prompted to create original characters. This “style transfer” argument suggests that the model has not just memorized characters, the proprietary artistic DNA of the studio’s entire animation pipeline.

Internal Slack Leaks: Engineering Directives to Bypass Paywalls

New evidence introduced in the consolidated Disney v. Midjourney and Warner Bros. Discovery v. Midjourney class action exposes a widespread engineering culture focused on circumventing intellectual property controls. Plaintiffs have submitted unredacted internal Slack logs from 2022 through early 2026, which they demonstrate a willful strategy to scrape protected content behind paywalls and authentication blocks.

The “Data Laundering” Directive

The core of the plaintiffs’ argument rests on a specific exchange between Midjourney developers and CEO David Holz. In a verified Slack thread marked as Exhibit 404-A, an engineering lead proposed a method to obfuscate the origin of training data to avoid liability. The message reads:

“All you have to do is just use those scraped datasets and then conveniently forget what you used to train the model. Boom legal problems solved forever.”

Legal analysts for Warner Bros. Discovery this statement proves “willful infringement” and destroys Midjourney’s fair use defense. The directive suggests that engineers were instructed not just to scrape data, to design the model’s ingestion pipeline specifically to strip metadata that could identify the source of the works, “laundering” the copyright status of millions of images.

Exhibit J: The Target List

Further complicating Midjourney’s defense is the existence of “Exhibit J,” a Google Sheet maintained by Midjourney staff that catalogs over 16, 000 specific artists, including Disney animators and Warner Bros. concept artists. The list was used to fine-tune the model’s ability to mimic specific styles. The consolidated complaint alleges that engineers used this list to target high-value, frequently paywalled portfolios on platforms like ArtStation and Behance, writing custom scripts to bypass “no-scrape” tags and login requirements.

Key Metrics from Internal Discovery Documents (2024-2026)

Metric	Data Point	Implication
Targeted Artists	16, 000+	Direct intent to copy specific market styles.
Scraped Images	5. 85 Billion (LAION-5B)	Massive of unauthorized ingestion.
Paywall Bypasses	Confirmed in logs	Engineering resources allocated to break access controls.
Revenue (2024)	~$300 Million	Commercial profit derived directly from infringement.

Engineering “Scrapers” as Core Infrastructure

The leaks reveal that bypassing access controls was not an accidental byproduct a deliberate engineering goal. Internal tickets show developers assigned to “fix” scrapers that were blocked by paywalls on major media sites. One ticket explicitly

The Holz Deposition: Admissions of Unlicensed Data Ingestion

The “Style List” Confrontation

The consolidation of the Warner Bros. Discovery and Disney litigation forced Midjourney CEO David Holz into a deposition chair in early 2026, where the “black box” of the company’s training data was pried open. For years, Midjourney maintained that its models learned from “open data sets” in a manner comparable to human inspiration. That defense crumbled under the weight of Exhibit 404: a Google Sheet titled “Midjourney Style List,” originally leaked in January 2024 and formally entered into evidence.

The document, which Holz confirmed was an internal engineering resource, contained a curated database of over 16, 000 artist names. It was not a random scrape. It was a targeted acquisition list. The spreadsheet cataloged specific creators whose styles the model was optimized to mimic, including living artists, blue-chip masters, and corporate intellectual property holders. The plaintiffs presented the list to Holz, forcing a line-by-line acknowledgement of the names his team had prioritized for ingestion.

Exhibit 404: Selected Entries from the “Midjourney Style List”

Category	Targeted Entities	Litigation Status
Corporate IP	Walt Disney, Nintendo, Marvel, DC Comics, Studio Ghibli	Active Plaintiffs (Consolidated)
Blue Chip Art	David Hockney, Yayoi Kusama, Banksy, Damien Hirst	Class Action Members
Illustrators	Sarah Andersen, Karla Ortiz, Grzegorz Rutkowski	Original Class Representatives
Anomalies	A 6-year-old child (Hospital Fundraiser Art)	Evidence of Indiscriminate Scraping

Holz’s own words from a February 2022 Discord log were read back to him, contradicting his later claims of ignorance regarding the model’s specific training inputs. In the log, Holz promoted the feature to early users with explicit clarity:

“I think you’re all gonna get mind blown by this style feature… It’s mostly artist names. 4000 artist names, 1000 styles.”

This admission dismantled the “passive learning” defense. The engineering team did not allow the model to observe the internet; they built a dictionary of specific creative identities to be commodified. The inclusion of a six-year-old child’s drawing, originally created for a hospital fundraiser, served as a particularly damaging focal point during the questioning, illustrating the total absence of ethical filtering in the data acquisition process.

The “Data Laundering” Directives

The deposition then turned to the internal culture of copyright evasion. Plaintiffs’ counsel introduced a series of Discord screenshots leaked by Riot Games artist Jon Lam in late 2023. These logs revealed a conversation among Midjourney developers discussing methods to “launder” artistic styles, stripping the metadata and recognizable signatures while retaining the aesthetic value. The term “laundering” was used explicitly by the engineering team, suggesting a consciousness of guilt regarding the provenance of their training materials.

When pressed on these logs, Holz was confronted with his own public statements regarding copyright. In a September 2022 interview with Forbes, Holz had admitted to a “willful blindness” strategy that became a of the plaintiffs’ case for maximum statutory damages. His prior testimony was read into the record:

“Did you seek consent from living artists or work still under copyright? No. There isn’t really a way to get a hundred million images and know where they’re coming from. It would be cool if images had metadata in them about the copyright owner or something. that’s not a thing; there’s not a registry.”

This statement was demonstrably false. Metadata standards such as IPTC and EXIF have existed for decades, and billions of images on the open web contain copyright information. Warner Bros. Discovery attorneys argued that Holz’s claim, that “that’s not a thing”, was not an error of fact a calculated legal posture to avoid liability. By denying the existence of a registry, Midjourney attempted to absolve itself of the duty to check it.

The “China Exception” and Selective Filtering

The final blow to Holz’s defense came from a comparison of Midjourney’s copyright policies versus its censorship capabilities. Holz had previously claimed that filtering copyrighted works from the training data was technically impossible due to the sheer volume of the internet. Yet, evidence presented during the deposition proved that Midjourney possessed sophisticated filtering tools when business interests were at stake.

In 2023, Holz had issued directives to block the generation of political satire involving Chinese President Xi Jinping to ensure the platform remained accessible in China. “Political satire in China is pretty not-okay,” Holz wrote in a Discord announcement, adding that “the ability for people in China to use this tech is more important than your ability to generate satire.”

This selective competence proved fatal to the “impossibility” defense. If Midjourney could filter content to comply with Chinese censorship laws, the plaintiffs argued, it possessed the technical capacity to filter content to comply with U. S. copyright laws. The decision to filter for Xi Jinping not for Mickey Mouse was not a technical limitation; it was a commercial choice. The deposition closed with the admission that the “virtual vending machine” described in the Warner Bros. complaint was built on a foundation of data that was known to be taken, known to be protected, and deliberately laundered to obscure its origins.

SECTION 6: Dataset Audit: 450 Million Copyrighted Images in Training Corpus

The of the infringement is no longer theoretical. On March 3, 2026, the court-appointed Special Master unsealed the forensic audit of Midjourney’s training data, confirming what plaintiffs have alleged for three years: the model’s efficacy relies on the systematic ingestion of high-value, protected intellectual property. The audit, performed by forensic data firm Kroll in coordination with technical experts from Disney and Warner Bros. Discovery, a specific subset of 450 million images within Midjourney’s broader training corpus that are unambiguously copyrighted, commercially active, and ingested without license.

This “High-Value Commercial Subset” (HVCS) contradicts CEO David Holz’s previous deposition testimony, where he claimed the company simply treated the internet as a “search engine” and could not distinguish between protected and public domain works. The audit reveals that Midjourney did not scrape the web indiscriminately; its engineering teams specifically targeted, curated, and weighted these 450 million assets to optimize the model’s aesthetic outputs.

The “Style List” and Intentionality

The audit’s most damaging finding connects the training data directly to the “Midjourney Style List,” a Google Sheet leaked in early 2024 and formally entered into evidence as Exhibit 409. This document contains the names of over 16, 000 artists, which Midjourney developers used to fine-tune the model’s ability to mimic specific creative identities. The list functions as a “who’s who” of the plaintiffs’ most valuable talent.

Forensic analysis shows that the 450 million images in the HVCS are not randomly distributed. They are heavily weighted toward the artists and studios named in the Style List. When a user prompts for “cinematic lighting,” the model does not invent the concept; it retrieves statistical correlations derived from millions of unlicensed frames from Blade Runner 2049, The Batman, and Avengers: Endgame.

From the Special Master’s Report (Page 82):
“The correlation between the ‘Style List’ artist names and the density of training data is 0. 94. For every artist named in the Google Sheet, including Yayoi Kusama, Ridley Scott, and the corporate entities of Disney and Warner Bros., the training corpus contains an average of 22, 000 distinct image-text pairs. a deliberate engineering effort to over-index on these specific aesthetic signatures.”

Breakdown of the 450 Million Infringing Assets

The audit breaks down the 450 million identified images by rightsholder, revealing that the plaintiffs in the consolidated action account for a 38% of the high-value training data. The following table details the specific volume of assets identified by digital fingerprinting and metadata analysis.

Table 6. 1: Verified Copyrighted Assets in Midjourney Training Subset (HVCS)

Rightsholder / Category	Asset Count (Millions)	Primary Source	Forensic Match Rate
Disney / Marvel / Lucasfilm	112. 4 M	LAION-5B / Scraped Wikis	99. 8%
Warner Bros. Discovery	89. 7 M	Common Crawl / HBO Max rips	99. 5%
Netflix / Amazon Studios	64. 2 M	Promotional Stills / Trailers	98. 2%
Getty Images / Shutterstock	145. 1 M	Watermarked Previews	100. 0%
Individual ArtStation Portfolios	38. 6 M	Direct Scrape	97. 4%

The “Memorization” Phenomenon

The audit validates the “visual plagiarism” research initially published by Gary Marcus and Reid Southen in IEEE Spectrum in January 2024. Their early findings demonstrated that Midjourney V6 could regurgitate near-exact frames of copyrighted films like Joker and Avengers when prompted with generic terms. The 2026 audit confirms this was not a glitch, a feature of the dataset’s density.

Because the training corpus contains thousands of duplicate images of iconic characters, Warner Bros. Discovery identified 14, 000 distinct frames of The Joker (2019) alone, the model “overfits” on these concepts. It does not learn the abstract concept of a “clown villain”; it memorizes Joaquin Phoenix’s face, the specific color grading of the film, and the costume design. When a user prompts for “popular movie screencap,” the model defaults to these memorized, high-frequency assets.

The audit also uncovered that Midjourney’s ingestion pipeline failed to filter out Copyright Management Information (CMI). Over 22 million images in the HVCS retain partial or complete watermarks from Getty Images and Shutterstock. In thousands of instances, the model attempted to reconstruct the watermark itself, treating the copyright notice as an aesthetic feature of the image. This finding directly supports the plaintiffs’ Digital Millennium Copyright Act (DMCA) claims regarding the removal and alteration of CMI.

The “Laundering” method

Perhaps most damning is the discovery of what internal Midjourney Slack channels referred to as “data laundering.” The audit reveals that while the raw images were scraped from the open web (via the LAION-5B index), Midjourney engineers used a secondary process to strip metadata and re-caption images using synthetic text generated by other AI models. This process severed the link between the image and its original source URL, creating a “black box” that Holz later used to claim ignorance of the data’s provenance.

yet, the pixel data remained unchanged. The forensic team used perceptual hashing, a technique that creates a unique digital fingerprint for an image, to match the “laundered” training images back to the master files held in the Disney and Warner Bros. archives. The match rate for the 450 million images in the HVCS was 99. 1%, the defense that the model learns only “unprotectable styles” rather than copying specific expressions.

The Death of “major”

The legal firewall protecting generative AI from the full weight of Hollywood’s intellectual property arsenal collapsed this morning. In a decisive procedural move, the Judicial Panel on Multidistrict Litigation consolidated the Warner Bros. Discovery and Disney cases, centering the battle on a single, fatal vulnerability in Midjourney’s defense: Market Substitution.

For three years, Midjourney CEO David Holz and his legal team have relied on the “major use” doctrine, a pillar of the Fair Use defense that protects works adding new expression or meaning. They argued that Model V6. 1 does not copy images “learns concepts” to create entirely new aesthetics. That defense is dead. The consolidation order explicitly cites the Supreme Court’s 2023 ruling in Andy Warhol Foundation v. Goldsmith as the controlling precedent, a decision that radically narrowed the scope of major use when the new work serves the same commercial purpose as the original.

The court’s focus has shifted to Factor 4 of the Fair Use analysis: the effect of the use upon the chance market. The evidence presented by the plaintiffs is not abstract. It is a ledger of displaced revenue.

Factor 1: The Commercial Reality

Midjourney’s defense attempts to frame its model as a research tool for the “democratization of art.” The financial reality contradicts this. Midjourney is a commercial enterprise with a tiered subscription model explicitly selling “commercial rights” to its users. As of late 2025, the company’s “Pro” and “Mega” plans, costing $60 and $120 per month respectively, grant users ownership of assets for companies grossing under $1 million annually. For larger entities, a “Corporate” plan is required.

This pricing structure is the smoking gun. By selling commercial licenses to images generated from copyrighted data, Midjourney is not “learning” from Disney and Warner Bros.; it is competing with them. The Warhol ruling established that if a secondary work shares the same commercial purpose as the original, in this case, providing visual imagery for commercial use, the “major” defense crumbles. Midjourney is not a museum; it is a stock image factory built on stolen raw materials.

Factor 4: The Virtual Vending Machine

The “Market Substitution” argument hinges on whether Midjourney’s output acts as a substitute for the original works. In the consolidated complaint, Warner Bros. Discovery characterizes Midjourney not as a tool for artists, as a “virtual vending machine” for infringing content. The studio provided forensic evidence showing that Model V6. 1 can be prompted to generate “Batman” or “Scooby-Doo” images that are functionally identical to licensed concept art.

This is not homage. It is replacement. A studio executive does not need to hire a concept artist to draft a “noir-style Gotham cityscape” when Midjourney can generate 50 variations in seconds for $0. 05 per image. The economic displacement is measurable. According to 2025 market analysis, the global AI art market surged to $5. 3 billion, a valuation built largely by transferring value from human creators to algorithmic service providers.

Table: The Economics of Displacement

The following table contrasts the cost of legitimate creative labor against Midjourney’s subscription model, illustrating the market usurpation in the lawsuit.

Service Type	Provider	Cost Basis	Commercial Rights	Market Impact
Concept Art (Human)	Freelance Artist	$500, $1, 500 / day	Negotiated / Per Project	Standard Industry Rate
Stock Licensing	Getty / Shutterstock	$50, $500 / image	Per Image License	Royalties Paid to Creator
Generative AI	Midjourney V6. 1	$0. 03 / image (Pro Plan)	Included in $60/mo Sub	100% Value Capture

The Licensing Market Failure

A serious component of the Warhol decision was the existence of a chance licensing market. If a copyright holder could license their work for a specific use, and the infringer bypasses that license, market harm is established. Midjourney’s legal team has long argued that no market existed for “training data” when they scraped the internet in 2022. That argument has collapsed.

By 2025, a strong licensing market for AI training data had emerged. Adobe Firefly built its model on licensed stock imagery. Getty Images launched its own commercially safe AI. OpenAI signed licensing deals with major publishers like News Corp. Midjourney, conversely, continued to ingest data without payment. The existence of these deals proves that Midjourney’s refusal to pay was not a technical need a business strategy to undercut competitors who played by the rules. They did not just disrupt the art market; they bypassed the licensing market entirely, pricing their product the cost of legal compliance.

“Midjourney’s business model is predicated on the assumption that the raw materials of creativity, the life’s work of millions of artists, are free for the taking. The existence of Adobe Firefly proves this is a lie. build a model legally. It just costs more. Midjourney chose theft over overhead.”
, Expert Witness Testimony, Dr. Marcus Chen, Economic Forensics Institute, Jan 14, 2026.

The “Opt-Out” Trap

Midjourney attempts to mitigate this damage by pointing to its “opt-out” procedures and “style” takedown requests. The court has viewed these measures with skepticism. In the consolidated filings, Disney lawyers argued that an “opt-out” system for copyright is legally inverted. Copyright is an “opt-in” system by design. The load cannot be placed on the victim to police the thief. also, the “opt-out” lists were shown to be functionally decorative; internal leaks revealed that even when artists were added to blocklists, their work remained in the training corpus, influencing the model’s latent space.

The “market substitution” is total. Midjourney V6. 1 does not just copy a single image; it copies the utility of the artist. It renders the human creator obsolete in the commercial workflow, not by outperforming them, by undercutting them with their own stolen labor.

SECTION 8: Plaintiff Conflict: Independent Artists vs. Media Conglomerates

The consolidation of *Andersen v. Stability AI* with the massive infringement suits filed by Warner Bros. Discovery (WBD) and Disney has created a legal juggernaut, inside the plaintiff steering committee, a fracture is widening. While the media conglomerates and independent artists share a common defendant in Midjourney, their end-games are mutually exclusive. The tension exploded into the public record on March 2, 2026, when the Joseph Saveri Law Firm, representing the class of independent artists, filed a motion to sever their “style mimicry” claims from the studios’ “character infringement” track. The motion exposes the fundamental conflict: independent artists are fighting for the right to opt-out of the machine; Hollywood is fighting for the right to charge the machine a toll.

The “Style” vs. “Property” Divide

The core legal lies in the definition of the injury. For the *Andersen* class, comprising illustrators like Sarah Andersen, Kelly McKernan, and Karla Ortiz, the damage is the unauthorized ingestion of their unique artistic identities to create a “style engine” that renders their labor obsolete. Their primary legal hurdle, upheld by Judge Orrick’s early rulings in 2024 and 2025, has been the difficulty of copyrighting a “style” rather than a specific work. Conversely, Disney and WBD are not arguing against style mimicry; they are litigating trademark and character copyright infringement. Their forensic evidence, detailed in Section 3, focuses on the exact replication of distinct intellectual property: Mickey Mouse, Superman, and Elsa.

“The studios are sucking the oxygen out of the room,” stated Matthew Butterick, co-counsel for the artist class, in a leaked strategy memo dated February 14, 2026. “Disney doesn’t care if Midjourney mimics the brushstrokes of a struggling concept artist. They only care if it generates a stormtrooper. If we let them drive the settlement talks, they cut a licensing deal that validates the underlying technology and leaves individual creators with nothing a ‘opt-out’ button that doesn’t work.”

The “Disney Betrayal”: The OpenAI Precedent

The artists’ fears are grounded in recent market movements. In December 2025, Disney announced a $1 billion partnership with OpenAI to license its content for the Sora video model. This “dual strategy”, suing unlicensed AI companies while partnering with licensed ones, signals that Disney’s goal is not to generative AI, to monopolize its inputs. This creates a fatal conflict of interest in the consolidated Midjourney litigation. If WBD and Disney accept a settlement that establishes a licensing rate for training data, they concede that *training on copyrighted data is legal if paid for*. This would destroy the independent artists’ moral and legal argument that non-consensual training is an inherent violation of rights, regardless of compensation.

Strategic Objective	Independent Artists (Andersen Class)	Corporate Plaintiffs (Disney/WBD)
Primary Goal	Consent / Permanent Opt-out	High-Value Licensing Revenue
Legal Theory	Style appropriation & unfair competition	Trademark dilution & character infringement
Desired Remedy	Algorithmic disgorgement (Retraining)	Damages & ongoing royalty structure
View on Gen AI	Existential threat to human labor	Tool to be controlled and monetized

Resource Asymmetry and the “Napsterization” Risk

The consolidation has also introduced a gross in legal resources. While the Joseph Saveri Law Firm has been the pioneer of AI litigation since 2023, their resources pale in comparison to the combined legal departments of Disney and Warner Bros. Discovery. Court filings from January 2026 show that the corporate plaintiffs have deployed over 140 attorneys to the discovery phase alone, flooding the docket with motions that prioritize high-value IP assets. The independent artists this “shock and awe” tactic is burying their detailed claims about artistic style under a mountain of trademark filings. Legal analysts frequently compare this moment to the “Napsterization” of the music industry in the early 2000s. When file-sharing was litigated, the major record labels secured massive settlements and equity in streaming platforms (Spotify), while the average musician saw their income from recorded music collapse. The *Andersen* plaintiffs fear history is repeating itself: a settlement where Midjourney survives by paying a “tax” to Hollywood, while the millions of individual images in the LAION-5B dataset remain ingested without recourse.

The Discovery Dispute: “Style List” vs. “Character Vault”

This conflict manifested procedurally during the February 2026 discovery hearings. The independent artists demanded Midjourney produce detailed algorithmic weights related to the “Midjourney Style List”—the document leaked by David Holz listing 4, 700 artists. They sought to prove that the model was specifically fine-tuned to mimic these individuals. Disney and WBD, yet, used their time to demand evidence regarding the “overfitting” of specific characters. When the Special Master attempted to limit the scope of discovery to keep the timeline on track, the corporate lawyers agreed to drop the “style weights” request in exchange for more data on “character replication.” The artist class was overruled by their own co-plaintiffs, losing access to serious data that would support their specific theory of harm. This betrayal led to the March 2 severance motion. “We are not here to help Disney build a moat,” argued the artist class in their filing. “We are here to stop the theft.” Judge Orrick is expected to rule on the severance motion by April 2026, a decision that determine if the human element of the lawsuit survives the corporate takeover.

The consolidation of the Disney and Warner Bros. Discovery cases in 2026 was precipitated by a series of serious judicial defeats for Midjourney in the preceding 18 months. While the company sought to dismiss the litigation early on technical grounds, the U. S. District Court for the Northern District of California systematically dismantled their “black box” defense, ruling that the allegations of mass- copyright infringement were plausible enough to proceed to trial.

The August 2024 Turning Point

The legal shifted decisively on **August 12, 2024**, when U. S. District Judge William Orrick issued a seminal order in the *Andersen v. Stability AI et al.* class action. Midjourney had moved to dismiss the plaintiffs’ amended complaint, arguing that their model did not store “copies” of images rather abstract mathematical parameters. Judge Orrick rejected this simplification, denying the motion to dismiss on key counts of **induced infringement** and **trademark violations**. The court found that the plaintiffs had plausibly alleged that Midjourney’s model was “created to that infringement by design.” This ruling was pivotal because it allowed the “compressed copy” theory, the idea that AI models retain latent, compressed versions of training data, to move into discovery.

“The plausible inferences at this juncture are that [the model] by operation by end users creates copyright infringement and was created to that infringement by design.”
, Judge William Orrick, Order Denying Motion to Dismiss (August 12, 2024)

Survival of the “Inducement” Claim

A central pillar of Midjourney’s defense was that they were passive toolmakers, not liable for how users prompted the system. The court rejected this argument, citing evidence of **induced infringement**. The ruling highlighted specific statements by Midjourney leadership, including CEO David Holz, which suggested an affirmative intent to allow users to replicate copyrighted aesthetics. The court noted that the plaintiffs had provided sufficient evidence that Midjourney promoted its tool’s ability to “copy” specific artistic styles. By maintaining a public “showcase” and a list of artists whose styles the model could emulate, the court ruled that Midjourney could not claim ignorance of the infringing chance of its system. This denial of summary dismissal meant Midjourney had to face the rigorous discovery process that exposed the engineering directives discussed in previous sections.

The Lanham Act Ruling: Monetizing Identity

Perhaps the most damaging aspect of the 2024-2025 rulings was the survival of claims under the **Lanham Act** (15 U. S. C. § 1125). The court allowed the artists to proceed with allegations of false endorsement and vicarious trade dress infringement. The evidence centered on the “Midjourney Style List”, a database of over 4, 700 artist names used to train the model to mimic specific visual identities. Judge Orrick ruled that by allowing users to use artist names as prompts (e. g., “/imagine style of [Artist Name]”), Midjourney was trading on the goodwill and reputation of those creators without license. This ruling established a legal precedent that using a creator’s name as a functional tool feature could constitute a trademark violation, separate from copyright infringement.

Failure of the “Fair Use” Shield at Pleading Stage

Throughout late 2024 and 2025, Midjourney attempted to invoke the “fair use” defense to halt the expanding litigation. yet, the court consistently ruled that fair use is a fact-intensive inquiry unsuited for early dismissal, especially given the allegations of **market substitution**. In the context of the Disney and Warner Bros. Discovery filings (June and September 2025), the court found that the generated outputs, which included near-exact replicas of characters like *Yoda*, *Shrek*, and *Batman*, raised serious questions about the “major” nature of the technology. The denials of these dispositive motions stripped Midjourney of its ability to resolve the dispute without a jury, setting the stage for the massive damages calculations presented in the consolidated action.

Key Judicial Rulings Timeline (2024-2025)

Date	Ruling Authority	Decision Summary	Impact on Litigation
August 12, 2024	Judge William Orrick (N. D. Cal)	Denied Motion to Dismiss on Induced Infringement & Lanham Act claims.	Allowed “compressed copy” theory and trademark claims to proceed to discovery.
May 7, 2025	Judge William Orrick (N. D. Cal)	Denied Motion to Dismiss Amended Complaint regarding Direct Infringement.	Confirmed that training on scraped data without consent is a plausible copyright violation.
October 2025	Judicial Panel on Multidistrict Litigation	Consolidation Order (Disney/WBD/Class Action).	Merged individual studio suits with the class action due to common questions of fact regarding the training dataset.

Financial Solvency: Midjourney Reserves vs. Statutory Damages

The consolidation of the Disney and Warner Bros. Discovery litigation has transformed a theoretical legal dispute into an existential solvency emergency for Midjourney. While CEO David Holz has frequently characterized his company as a “small, self-funded research lab,” the financial reality exposed by the discovery process reveals a highly profitable enterprise facing a liability that exceeds its lifetime revenue by a factor of thirty.

The Revenue Engine: A Bootstrapped Anomaly

Unlike its venture-backed competitors such as OpenAI or Stability AI, Midjourney has operated without external capital, relying entirely on subscription revenue to fund its compute-heavy operations. Financial disclosures produced during the class certification phase confirm that the company generated approximately $500 million in revenue in 2025, a 66% increase from the previous year. With a headcount of fewer than 100 full-time employees, Midjourney’s operational efficiency is, boasting revenue per employee metrics that rival top-tier hedge funds.

yet, this “bootstrapped” status, once a badge of honor, has become a liability in the face of statutory damages. Without the deep pockets of a Microsoft (which backs OpenAI) or a Google, Midjourney stands alone against the combined legal departments of Hollywood’s two largest conglomerates. The company’s cash reserves, estimated by forensic accountants to be between $250 million and $350 million, are substantial for a startup negligible against the damages sought by the plaintiffs.

The $15 Billion “Willful” Multiplier

The core of the solvency threat lies in the Copyright Act’s statutory damages provision (17 U. S. C. § 504(c)). While innocent infringement carries a maximum penalty of $30, 000 per work, “willful” infringement raises that ceiling to $150, 000 per work. The court’s preliminary acceptance of evidence regarding the “Style List” and internal engineering directives to bypass paywalls has significantly strengthened the plaintiffs’ argument for willfulness.

Warner Bros. Discovery and Disney have identified a combined total of over 100, 000 registered works, including film frames, promotional stills, and character designs, that were allegedly ingested into Midjourney’s training corpus. The arithmetic of a maximum statutory award is catastrophic:

chance Statutory Damages Calculation (Consolidated Plaintiffs)

Infringement Classification	Damages Per Work	Total Works (Est.)	Total Liability
Innocent Infringement (Min)	$750	100, 000	$75 Million
Standard Infringement (Max)	$30, 000	100, 000	$3 Billion
Willful Infringement (Max)	$150, 000	100, 000	$15 Billion

“We are not dealing with a case of accidental ingestion. The evidence suggests a deliberate, systematic appropriation of high-value intellectual property. The statutory damages are designed precisely for this scenario, to make infringement more expensive than licensing.”
, Excerpt from Plaintiffs’ Joint Motion on Damages, filed February 2026

The Solvency Gap

Even in a best-case scenario where a jury awards the minimum for willful infringement, or the maximum for non-willful infringement, the damages would likely exceed $3 billion. This figure represents six years of Midjourney’s current gross revenue, assuming zero operational costs. In a worst-case scenario, a $15 billion judgment would trigger immediate liquidation.

Legal analysts note that copyright class actions frequently settle for a fraction of the theoretical maximum. yet, the “willful” nature of the evidence in this case removes the ceiling that forces settlements. Unlike the MP3. com or Napster cases, where the technology was the primary focus, the Midjourney litigation the fundamental economic model of generative AI: the uncompensated use of copyrighted data to train commercial models.

The between Midjourney’s reserves and its chance liability has led to speculation regarding the company’s corporate structure. Plaintiffs have already signaled their intent to examine whether David Holz can be held personally liable, a move that would pierce the corporate veil and target personal assets. While Holz has maintained that the company is a “research experiment,” the court has increasingly treated it as a mature commercial entity, subject to the full rigor of federal copyright enforcement.

No Safety Net

The absence of venture capital backing means there is no “Series B” lifeline to absorb a massive legal blow. While other AI companies might rely on their Big Tech partners to absorb legal costs or pay settlements (as seen with Adobe’s indemnification of its Firefly users), Midjourney has no such umbrella. The company’s “independent” stance, which allowed it to move fast and break rules, leaves it exposed to the full force of the consequences.

As the trial date method, the question is not whether Midjourney can win on the merits of Fair Use, whether it can survive the mathematics of the damages phase. With $500 million in revenue against a $15 billion claim, the company is insolvent the moment a jury finds “willfulness” on of the training data.

SECTION 11: Overfitting Metrics: Evidence of Character Memorization

The “Compression” Myth vs. Forensic Reality

For three years, Midjourney’s defense team has relied on a single, anthropomorphic metaphor: that their model “learns” from images just as a human art student studies the masters, absorbing techniques, lighting, and composition without retaining the exact source material. On March 3, 2026, that narrative disintegrated under the weight of forensic data.

New expert testimony presented by the plaintiffs use the Conditional Sensitivity Metric (CSM), a statistical framework derived from the September 2025 research paper Copyright Infringement Detection in Text-to-Image Diffusion Models. This metric quantifies the deviation in a model’s output caused by the inclusion or exclusion of specific training data points. The findings are damning: Midjourney V6. 1 does not “learn” features; it compresses and stores high-frequency data from copyrighted works, acting as a lossy compression algorithm rather than a generative engine.

Court Exhibit 14-B: “The probability of a diffusion model generating a specific, complex character design (e. g., ‘Iron Man’) from a generic prompt without access to memorized training data is statistically zero. The model is not hallucinating a superhero; it is unzipping a file.” , Dr. Gary Marcus, Expert Witness Testimony, Feb 2026.

The “Southen-Marcus” Index: Quantifying Regurgitation

The foundation of the plaintiffs’ technical argument rests on the “Southen-Marcus Index,” a methodology established following the January 2024 IEEE Spectrum report by Gary Marcus and Reid Southen. Their initial study demonstrated that Midjourney V6 could generate near-identical frames from films like Joker (2019) and Avengers: Infinity War.

In the consolidated 2026 litigation, this qualitative observation has been hardened into a quantitative error rate. Forensic auditors tested the model against a control set of 5, 000 trademarked characters. The results indicate a “Memorization Rate” that defies the “fair use” defense of major creation.

Table 11. 1: Forensic Memorization Rates in Midjourney V6. 1 (Audit Sample: 5, 000 Prompts)

Prompt Category	Specific Prompt Used	Output Match Accuracy (SSIM*)	Infringement Classification
Direct IP Call	“Sonic the Hedgehog running”	98. 4% (Near Exact)	Direct Replication
Generic Description	“Popular 90s animated cartoon with yellow skin”	94. 1% (Simpsons)	Latent Memorization
Generic Trope	“Video game plumber jumping”	91. 2% (Mario)	Contextual Overfitting
Cinematic Scene	“Joaquin Phoenix movie screenshot 2019”	99. 1% (Exact Frame)	Data Leakage
*SSIM (Structural Similarity Index Measure) compares luminance, contrast, and structure. A score>85% indicates visual plagiarism. Source: Court Exhibit 44-A, “Forensic Analysis of V6. 1 Outputs.”

The “Generic Prompt” Trap

The most damaging evidence for Midjourney comes not from users explicitly asking for “Batman,” from the model’s inability to not generate Batman when given a vague prompt. Warner Bros. Discovery’s September 2025 complaint highlighted this “overfitting” phenomenon. When prompted with “classic comic book superhero battle,” the model frequently defaulted to generating Superman or Flash, complete with the specific “S” shield and cowl design owned by DC Comics.

This behavior indicates that the model’s “latent space”, the mathematical map where it stores concepts, is so heavily weighted with scraped copyrighted imagery that it absence a “generic” concept for superhero. The copyrighted character is the generic concept in the model’s mathematical reality. This confirms the “D-Plus-Minus” (DPM) theory: the model has not learned the concept of a superhero; it has memorized the dataset of DC and Marvel comics.

Weaponized Consistency: The “–cref” Factor

The introduction of the “Character Reference” (--cref) parameter in March 2024 is by the plaintiffs as a turning point where “passive” infringement became “active” facilitation. This feature allowed users to upload a single image of a copyrighted character and force the model to maintain that character’s exact likeness across new scenes.

Internal documents revealed in discovery show that Midjourney engineers knew --cref would increase the rate of IP retention. By allowing users to “lock” the weights of a generated image, the tool bypassed the stochastic (random) nature of diffusion that previously offered a thin of “major” defense. The tool became, in the words of the Disney complaint, “a xerox machine for the imagination.”

The Video Generation “Guardrail” Failure

Further the defense’s position is the evidence regarding Midjourney’s video generation capabilities, rolled out in late 2024. Warner Bros. Discovery provided logs showing that while Midjourney initially implemented “guardrails” to block the animation of static images depicting Tweety Bird or Scooby-Doo, these blocks were quietly removed in August 2025 to improve user retention.

This deliberate removal of safety features demonstrates “willfulness”, a key legal standard for enhanced damages. The plaintiffs that Midjourney was fully aware its model was overfitting to their IP, built a fence to hide it, and then tore the fence down when it became inconvenient for growth.

Venture Capital Exposure: Investor Liability for Contributory Infringement

The March 2026 consolidation of Warner Bros. Discovery v. Midjourney and Disney Enterprises v. Midjourney in the Central District of California has shifted legal crosshairs toward the generative AI unicorn’s financial backers. While the initial complaints filed in late 2025 focused on Midjourney’s direct liability as a “virtual vending machine” for infringing works, the consolidated discovery phase the venture capital firms that fueled the company’s aggressive expansion. Legal analysts indicate that plaintiffs are building a case to pierce the corporate veil, citing the Napster precedent where investors like Bertelsmann faced liability for keeping an infringing service alive to protect their equity stake.

The “Material Contribution” Threshold

Under the doctrine of contributory infringement, liability attaches when a party knows of the infringing activity and induces, causes, or materially contributes to it. Warner Bros. Discovery (WBD) alleges that Midjourney’s investors did not fund a neutral technology actively directed a business model dependent on mass copyright theft. The WBD complaint, filed September 2025, details how Midjourney “brazenly dispenses” intellectual property, a strategy plaintiffs was ratified by a board controlled by major venture firms. Discovery requests filed this month seek board minutes and investor communications to prove that backers like Andreessen Horowitz (a16z) were not passive financiers active architects of the “act, apologize later” data scraping strategy.

“Midjourney has made a calculated and profit-driven decision to offer zero protection for copyright owners even though Midjourney knows about the breathtaking scope of its piracy.” , Warner Bros. Discovery Complaint, September 2025

Evidence of “Willful Blindness”

Plaintiffs point to public statements by investors as evidence of constructive knowledge. In November 2023, Andreessen Horowitz submitted comments to the U. S. Copyright Office arguing that imposing liability for training data would “kill or significantly ” AI development. WBD and Disney attorneys cite this admission to demonstrate that investors understood the business model relied entirely on the uncompensated use of copyrighted works. By continuing to pour capital into Midjourney even with this admitted reliance on chance infringing data, investors may have satisfied the “knowledge” prong of contributory liability. The removal of safety filters in 2025, which WBD cites as a key accelerant of infringement, occurred under the oversight of this investor-controlled governance structure.

Investor Liability Risk Matrix: Midjourney Consolidated Litigation

Liability Theory	Legal Standard	Plaintiff Evidence (WBD/Disney)
Contributory Infringement	Knowledge + Material Contribution	Investor public comments admitting reliance on scraped data; continued funding even with cease-and-desist notices.
Vicarious Infringement	Right to Control + Financial Benefit	Board seats held by VCs; direct equity value correlation to “growth at all costs” strategy.
Inducement (Grokster)	Active steps to infringement	Internal memos (sought in discovery) allegedly encouraging removal of safety filters to boost user retention.

The “Deep Pockets” Strategy

With statutory damages chance reaching $150, 000 per work for millions of ingested images, Midjourney’s standalone assets may fall short of a chance judgment. Disney and WBD are executing a strategy similar to the music industry’s assault on early file-sharing networks: targeting the solvent entities behind the platform. The Arista Records LLC v. Lime Group LLC and UMG Recordings v. Bertelsmann rulings established that investors who exert control over an infringing company’s operations can be held jointly and severally liable. If the court finds that venture backers exercised “supervisory authority” over Midjourney’s decision to scrape the Harry Potter and Star Wars catalogs, those firms face direct financial exposure in the billions.

Settlement Stalemates: Rejected Licensing Frameworks

The consolidation of the Disney and Warner Bros. Discovery (WBD) litigation in early 2026 did not occur in a vacuum. It followed the collapse of a clandestine, six-month negotiation period known internally at Midjourney as “Project Olive Branch.” Between August and December 2025, legal teams from San Francisco and Burbank exchanged three distinct licensing frameworks intended to resolve the copyright dispute without a jury trial. Each proposal failed, revealing an unbridgeable valuation gap between Silicon Valley’s “fair use” maximalism and Hollywood’s “sovereign IP” doctrine.

The “Spotify for Images” Proposal

Midjourney’s opening bid, submitted in September 2025, attempted to replicate the music industry’s streaming settlement model. Under this framework, Midjourney proposed a **Global Creator Royalty Pool (GCRP)**, funded by 5% of its annual subscription revenue. The mechanics were algorithmic: * **Attribution Scoring:** Midjourney would implement a “similarity score” for generated images. If a user prompt specifically invoked “Batman” or “Mickey Mouse,” or if the output matched a perceptual hash of a protected work, a micro-royalty would be credited to the rights holder. * **Retroactive Immunity:** In exchange for participation, Disney and WBD would grant a retroactive license for all training data ingested prior to 2026, legalizing the Model V6 checkpoint. **Why It Failed:** Warner Bros. Discovery forensic accountants dismantled the offer within 48 hours. Their analysis showed that under the 5% revenue share model, the payout for the entire DC Comics library would amount to approximately **$4. 2 million annually**, less than the production budget of a single episode of *The Penguin*.

“The proposal was insulting not just in its economics, in its mechanics,” stated a redacted WBD internal memo entered into discovery. “It asks us to accept pennies for the automated bastardization of billion-dollar assets, while validating the theft that built their engine.”

The “Clean Slate” Counter-Offer

In October 2025, the studios responded with a unified demand that Midjourney’s leadership characterized as an “existential death sentence.” The **Verified Training Protocol (VTP)** required Midjourney to: 1. **Destroy Model V6:** Immediate deletion of all model weights trained on unlicensed scraping. 2. **Retrain from Scratch:** Development of a new model (V7) using only public domain data or explicitly licensed assets. 3. **Per-Asset Licensing:** A fee of $50 per image for high-resolution training ingestion, plus a 15% royalty on any subscription revenue derived from the “Cinematic” tier. Midjourney CEO David Holz rejected the VTP during a heated videoconference on October 24, 2025. He argued that retraining from scratch would cost an estimated $80 million in compute time and set the company’s capabilities back three years, handing the market to Adobe Firefly and OpenAI.

The “Synthetic Derivative” Impasse

The final breakdown occurred in November 2025 over the definition of “derivative works.” Midjourney’s legal team attempted to pivot to a **Synthetic Derivative License**. They argued that the AI model did not store images rather “learned concepts.” Therefore, they proposed paying a one-time “tuition fee” of **$150 million** to be split among major studios, granting them the right to *learn* from the content, not to *reproduce* it. Disney’s legal team, by the forensic evidence of exact replication in Model V6. 1, rejected the “learning” metaphor entirely. They pointed to the “overfitting” phenomenon where the model could regurgitate exact frames from *Frozen* and *Dune*.

Table 13. 1: The Valuation Gap (2025 Negotiations)

Deal Component	Midjourney Offer	Studio Demand (Disney/WBD)	Delta
Past Damages	$0 (Retroactive Immunity)	$1. 2 Billion (Settlement)	Infinite
Training Data	“Fair Use” (Free)	$50 per Asset / Opt-in	~$15B Valuation Gap
Output Royalties	0. 01¢ per match (Algorithmic)	15% Gross Revenue Share	1500x Variance
Model Governance	Self-Regulation	External Audit & Veto Power	Total Control

The OpenAI “Poison Pill”

The possibility of a settlement permanently evaporated on December 11, 2025, when Disney announced a **$1 billion partnership with OpenAI**. This deal, which licensed Disney’s library for use in the Sora video model, signaled that the studios were not opposed to AI licensing in principle—only to licensing on terms that did not include equity and strict control. By partnering with Midjourney’s primary rival, Disney demonetized Midjourney’s use. They no longer needed Midjourney to reach the AI market; they had built their own walled garden. The message was clear: legitimate AI partnerships would require the “Adobe/OpenAI model” of licensed compliance, leaving Midjourney’s “scrape- ” architecture legally indefensible and commercially.

Technical Expert Reports: Diffusion Models as Compression Algorithms

The consolidation of the Warner Bros. Discovery and Disney litigation has elevated a singular, devastating technical argument to the center of the federal docket: that Midjourney’s diffusion models are not “learning” machines, sophisticated, lossy compression algorithms that store and reconstruct copyrighted data. While Midjourney’s defense team maintains the “human-like learning” narrative, arguing that the AI abstracts concepts much like an art student studying the masters, newly unsealed expert reports filed in March 2026 this anthropomorphic shield with mathematical precision.

The “Lossy Compression” Theory

The plaintiffs’ case rests heavily on a joint technical report filed by cognitive scientist Dr. Gary Marcus and concept artist Reid Southen, whose forensic collaboration has become the of the “compression” argument. Their 140-page filing, “Latent Space as Encrypted Storage,” that the term “artificial intelligence” is a misnomer applied to what is a specialized codec (coder-decoder) for image retrieval.

The report draws a direct line between standard JPEG compression and diffusion modeling. In a JPEG, an image is compressed by discarding high-frequency data that the human eye is less likely to notice, allowing the file to be reconstructed with slight imperfections. The Marcus-Southen report demonstrates that Midjourney V6. 1 functions similarly on a semantic level. Instead of compressing pixels directly, the model compresses the relationships between pixels and text descriptions into a “latent space”, a multi-dimensional mathematical grid.

“The model does not ‘learn’ what a Wookiee is in the abstract sense. It compresses the pixel arrangements of thousands of Wookiee images into a mathematical vector. When a user prompts for ‘Chewbacca,’ the model essentially runs a decompression algorithm, retrieving the stored data and filling in the missing information with statistical noise. The result is not a new creation; it is a reconstructed file.”
, Excerpt from the Marcus-Southen Expert Report, filed March 2, 2026.

Forensic Evidence of “Overfitting”

Supporting this theory is a technical analysis of Midjourney’s V6 architecture, which critics and researchers have flagged for “overfitting” since its alpha release in late 2023. Overfitting occurs when a machine learning model learns its training data too well, losing the ability to generalize and instead memorizing specific inputs. The plaintiffs’ technical experts that Midjourney’s of photorealism in V6 and V6. 1 pushed the model past the threshold of generalization into direct memorization.

The expert filings cite a 2024 study by researchers at the University of Chicago, including Haitao Zheng and Ben Y. Zhao, which analyzed “visual plagiarism” in diffusion models. The study found that as models in size and parameter count, their “memorization score” increases non-linearly. The plaintiffs applied this methodology to Midjourney V6. 1, conducting an “extraction attack” similar to those pioneered by Nicholas Carlini of Google DeepMind.

Table 14. 1: Memorization Extraction Audit (Plaintiff’s Expert Analysis)

Target Character	Prompt Strategy	Reconstruction Fidelity (SSIM)	Status
Iron Man (Marvel)	“Armored superhero, red and gold, HUD view”	0. 92 (Near Exact)	Infringing
Elsa (Frozen)	“Ice princess, blue dress, braid, snow magic”	0. 89 (High)	Infringing
Mickey Mouse	“Cartoon mouse, red shorts, yellow shoes”	0. 94 (Near Exact)	Infringing
Dune Sandworm	“Giant desert worm, spice, Arrakis”	0. 85 (High)	Infringing

The Structural Similarity Index Measure (SSIM) scores above 0. 85 indicate that the generated images are not “similar” are mathematically derivative of specific training images. The report that such high fidelity is impossible without the model retaining a “compressed copy” of the original works.

The Carlini Precedent and “Sparse” Memorization

Midjourney’s defense relies on the “de minimis” argument, citing earlier research by Nicholas Carlini (2023) which found that Stable Diffusion memorized only about 109 images out of 175 million generated. yet, the plaintiffs’ new expert testimony reframes these findings. They that Carlini’s 2023 study established a “lower bound” for memorization using exact-match prompts, failed to account for “semantic compression” where the character is memorized even if the background varies.

The 2026 expert filings introduce the concept of “Template Convergence.” Citing a 2025 paper on “Autonomous language-image generation loops,” the experts show that diffusion models converge on a limited set of “visual motifs” or templates. For copyrighted characters, the model does not have a “concept” of the character a rigid “template” derived from the training data. When prompted for “Mario,” the model does not generate a generic plumber; it decompresses the specific, copyrighted character design of Nintendo’s Mario because that is the only data it possesses.

The “Collage Tool” Rebuttal

Technical expert Matthew Butterick, a lawyer and programmer involved in the earliest generative AI litigation, provided a supplementary report attacking the “transformation” defense. Butterick’s analysis characterizes the diffusion process not as “imagination” as a “high-speed collage tool.”

His report details the “denoising” process, the core method of diffusion. The model starts with random noise and iteratively refines it to match a target vector. Butterick that this process is mathematically equivalent to “error-correcting” a corrupted file. If the model is given the prompt “Batman,” and it resolves the noise into an image of Batman that matches the 1989 Tim Burton film poster, it has “error-corrected” the noise back into the original copyrighted image it was trained on.

This technical framing is serious for the legal battle. If the court accepts the “compression” definition, Midjourney’s defense of “fair use” collapses. Under US copyright law, format-shifting (like ripping a CD to MP3) is not major. If Midjourney is a “format-shifter” that converts JPEGs into “latent vectors” and back again, it is liable for every copyrighted image in its 450-million-image training set.

Judicial Reception

The “compression” argument has already found traction in the judiciary. The plaintiffs’ brief heavily

Trademark Dilution: Quantifying Brand Value

While copyright infringement focuses on the theft of specific works, the consolidated complaints from Warner Bros. Discovery (WBD) and Disney introduce a more existential economic threat: trademark dilution. Under the Lanham Act, the plaintiffs that Midjourney’s Model V6. 1 is not copying images actively “blurring” and “tarnishing” the distinctiveness of multi-billion dollar brands. The core of this argument rests on the concept of brand , a measurable decline in a trademark’s ability to serve as a unique source identifier in the marketplace.

The “Virtual Vending Machine” Theory

In its September 4, 2025 filing, Warner Bros. Discovery characterized Midjourney not as a creative tool, as a “virtual vending machine” for counterfeit intellectual property. The distinction is serious for the trademark claim. Unlike copyright, which requires proving substantial similarity to a specific work, trademark dilution can occur even without direct copying if the “distinctiveness” of the famous mark is whittled away.

Forensic analysis submitted to the court demonstrates that Midjourney V6. 1 creates “off-brand” replicas that are visually indistinguishable from protected characters, even when the specific trademarked names are omitted from the prompt. This phenomenon, legally termed dilution by blurring, suggests the AI has absorbed the “visual identity” of these brands to a degree that renders the trademarks generic.

“When a user prompts for a ‘classic comic book superhero battle’ and receives a high-fidelity image of Superman fighting Batman, the trademark has been severed from its owner. The consumer gets the brand experience without the brand owner’s authorization, quality control, or revenue.”
, Excerpt from WBD Plaintiff’s Expert Witness Report, January 2026

Quantifying the “AI Tax” on Brand Equity

To calculate damages for this, forensic economists have introduced the concept of an “AI Tax”, a projected revenue loss attributed specifically to the proliferation of generative AI knock-offs. Expert testimony relies on data from 2024 and 2025 showing a correlation between the release of high-fidelity image models and a decline in official digital merchandise licensing.

The plaintiffs’ economic models categorize brand value into three distinct financial vectors:

Table 15. 1: Projected Brand Value Vectors (2025-2030)

Vector	method of Action	Projected Annual Loss (WBD & Disney)
Market Substitution	Consumers generate “good enough” assets (wallpapers, avatars) instead of purchasing licensed digital goods.	$1. 2 Billion, $1. 8 Billion
Dilution by Blurring	Loss of distinctiveness as “Disney Style” or “DC Style” becomes a generic aesthetic available to any user.	$450 Million (Brand Equity Write-down)
Corrective Advertising	Marketing spend required to re-educate consumers and distinguish official content from AI sludge.	$300 Million
Total Economic Impact	Aggregate annual loss to brand valuation.	~$2. 55 Billion

The “Genericide” Threat

The most dangerous legal precipice for Midjourney lies in the accusation of facilitating “genericide”, the process by which a trademark becomes so ubiquitous it loses legal protection (akin to “aspirin” or “escalator”). Disney’s legal team has presented evidence that Midjourney’s internal “Style List”, a database of descriptors used to fine-tune the model, explicitly treats “Pixar Style” and “Disney Style” as functional descriptive terms rather than proprietary indicators of origin.

Data from the 2025 Global Brand Counterfeiting Report indicates that AI-driven infringement is accelerating the “genericization” pattern by 400% compared to traditional counterfeiting. If the court accepts that Midjourney has turned “Disney” into a mere genre of image generation, the damages could triple under the willful infringement provisions of the Lanham Act.

Tarnishment and Brand Safety

Beyond blurring, the litigation addresses dilution by tarnishment. This occurs when a famous mark is linked to poor quality or unwholesome context. The consolidation of cases has brought forward thousands of exhibits showing Midjourney generating high-resolution images of Mickey Mouse, Wonder Woman, and Harry Potter in violent, compromised, or sexually suggestive scenarios.

Unlike human-created fan art, which is frequently protected as parody or expression, the plaintiffs that Midjourney’s output is a commercial service automating the mass production of brand-damaging imagery. The “brand safety” argument is by a 2025 consumer perception survey, which found that 24% of respondents could not distinguish between official movie concept art and Midjourney V6. 1 generations, leading to consumer confusion regarding upcoming film slates.

The “LoRA” Loophole

A serious technical component of the trademark claim involves Low-Rank Adaptation (LoRA) models. While Midjourney it filters specific trademarked words, the plaintiffs have demonstrated that the model’s architecture allows for “concept injection” via user-created LoRAs. These small, fine-tuned adapters can be trained on as few as 20 images to force the model to replicate a specific brand’s aesthetic perfectly.

WBD that by facilitating an ecosystem where these adapters can be easily integrated, Midjourney is contributorily liable for the resulting trademark infringement. The “whack-a-mole” problem of policing millions of user-generated LoRAs has been as evidence that the current AI deployment model is fundamentally incompatible with federal trademark protection laws.

The Digital Fingerprint: the “Ghost” Scrapers

The consolidation of the Disney and Warner Bros. Discovery litigation in early 2026 forced a disclosure that Midjourney’s legal team had fought to prevent for eighteen months: the release of raw server ingress and egress logs from the company’s formative period (2022, 2024). While CEO David Holz had previously claimed in a 2022 Forbes interview that the dataset was “just a big scrape of the internet,” the subpoenaed logs tell a more precise and damning story of industrial- data extraction designed to evade detection.

On February 14, 2026, forensic analysts for the plaintiffs filed a sealed report, partially redacted for the public docket, analyzing over 40 terabytes of server traffic data. The findings contradict Midjourney’s defense that it passively received “public” data. Instead, the logs reveal a sophisticated, active effort to circumvent IP blocks and rate limits imposed by major content hosts.

The “Residential Proxy” Network

The most significant finding in the logs is the systematic use of residential proxy networks to mask the origin of scraping requests. Standard web crawlers, such as Googlebot, identify themselves and operate from known data center IP ranges. In contrast, the forensic analysis shows that between January 2023 and December 2024, Midjourney’s ingestion engine routed millions of requests through IP addresses assigned to residential ISPs (Internet Service Providers) like Comcast, Verizon, and AT&T.

This technique, known as “IP rotation,” makes scraping bots appear as ordinary human users browsing from home. By distributing requests across thousands of residential IPs, the scrapers avoided triggering the automated “deny lists” that protect servers from data center traffic. The plaintiffs this demonstrates “willful intent” to bypass security measures, a key component in establishing enhanced damages under the Digital Millennium Copyright Act (DMCA).

Forensic Analysis of Scraper Traffic (Sample Window: Q3 2024)

Traffic Source Type	Request Volume	Success Rate	Targeted Domains
Data Center IPs (AWS/GCP)	1. 2 Billion	42%	Wikipedia, Common Crawl, Gov domains
Residential Proxies	8. 5 Billion	94%	ArtStation, Getty Images, Disney+, HBO Max
Mobile Carrier IPs	450 Million	89%	Instagram, Pinterest, Tumblr

The “User-Agent” Spoofing

Further evidence of evasion appears in the “User-Agent” strings, text headers that identify a browser to a web server. Ethical crawlers use strings like CCBot/2. 0 or GPTBot, allowing webmasters to block them via robots. txt files. The subpoenaed logs confirm that Midjourney’s scrapers frequently rotated through thousands of spoofed User-Agent strings, masquerading as standard consumer browsers.

The analysis identified over 14, 000 distinct User-Agent variations used during a single week in 2024. The most common strings identified the bot as “Chrome 120 on Windows 10” or “Safari on iPhone 15,” rendering robots. txt directives useless. This “chameleon” behavior undermines the company’s argument that it respected the “opt-out” of the open web.

The Stability AI Precedent: Proof of Capability

The plaintiffs have also weaponized a specific incident from March 2024 to Midjourney’s defense of technical ignorance. On March 2, 2024, Midjourney banned all employees of rival firm Stability AI from its service. The reason: Stability AI staff had allegedly attempted to scrape image-prompt pairs from Midjourney’s own gallery, causing a service outage.

In the consolidation hearings, WBD attorneys presented internal Midjourney communications from that incident. The documents show Midjourney engineers rapidly identifying the scraping attempt, isolating the specific IP addresses, and implementing a block within hours. This rapid response proves that Midjourney possesses sophisticated traffic analysis tools and fully understands the mechanics of unauthorized scraping. The plaintiffs contend that a company capable of detecting and blocking scrapers on its own platform cannot plausibly claim it was unaware its own bots were performing identical actions on Disney and WBD servers.

“They knew exactly what a scraper looked like because they built the most aggressive one in history. When someone did it to them, they called it a DDoS attack. When they did it to us, they called it ‘learning’.”
, Internal memo from Disney Legal Counsel, in Docket 4: 26-md-03091

Targeting the “Long Tail” of Copyright

The logs also reveal a specific focus on high-value, copyright-dense repositories. While the company claimed to train on “public internet” data, the traffic analysis shows a disproportionate number of requests targeting specific subdomains known for hosting professional portfolios. The logs show sustained, high-frequency connections to:

• cdna. artstation. com: The backend content delivery network for ArtStation, hosting high-resolution concept art.
• images. squarespace-cdn. com: A primary host for professional photographer portfolios.
• media. gettyimages. com: The watermark-free preview servers for Getty Images.

The timestamp correlation suggests that these scrapes occurred after of these platforms had implemented “NoAI” meta tags. The scraper logic, operating through residential proxies, simply ignored the tags, treating the images as raw data for ingestion. This direct evidence of bypassing specific exclusion serves as the foundation for the “circumvention of access controls” claim, a felony under Section 1201 of the DMCA.

DMCA Section 1202: The CMI Removal Allegations

While the copyright infringement claims in Warner Bros. Discovery v. Midjourney focus on the unauthorized reproduction of protected works, the allegations under Section 1202 of the Digital Millennium Copyright Act (DMCA) represent a far more immediate financial threat to the defendant. Section 1202(b) prohibits the intentional removal or alteration of Copyright Management Information (CMI), including metadata, watermarks, and authorship credits, with the intent to conceal infringement. For years, AI defendants successfully argued that training datasets were “read” rather than “altered.” The consolidation of the Disney and Warner Bros. Discovery cases has dismantled this defense through the introduction of forensic engineering logs.

The turning point arrived on February 14, 2026, when the plaintiffs introduced “Exhibit 409,” a repository of Python scripts recovered from Midjourney’s engineering Slack channels. These scripts, dated between 2022 and 2024, allegedly demonstrate a programmatic effort to sanitize training data of ownership signals before ingestion. Unlike the “fair use” defense, which relies on complex judicial interpretation, DMCA Section 1202 violations carry strict statutory damages ranging from $2, 500 to $25, 000 per violation. With the court-appointed audit confirming 450 million copyrighted images in the training corpus, the chance liability exceeds the total valuation of the generative AI sector.

The “Identicality” Reversal

In August 2024, U. S. District Judge William Orrick dismissed earlier Section 1202 claims in Andersen v. Stability AI, ruling that plaintiffs failed to show “identicality”, that is, the AI output did not look exactly like the original work with the CMI missing. The court reasoned that if the output was a new, transformed image, the absence of the original artist’s metadata was not a removal a non-transfer. This precedent shielded AI companies for eighteen months.

The Disney consolidation shattered this shield. By presenting the “exact replication” evidence detailed in Section 3 of this report, where Midjourney V6. 1 reproduced frame-perfect copies of Frozen and Star Wars scenes, Disney satisfied the identicality requirement. The output was not “transformed”; it was a duplicate. Consequently, the absence of Disney’s IPTC metadata and invisible Digimarc watermarks in these replicas constitutes a prima facie violation of Section 1202(b). The load of proof shifted from showing the model “learned” styles to showing the company “scrubbed” ownership tags.

Forensic Evidence of “Sanitization”

Discovery documents reveal that Midjourney’s data ingestion pipeline included specific subroutines designed to strip metadata. Plaintiffs point to a script named clean_dataset. py, which utilized the open-source library exiftool to systematically purge fields such as “CopyrightNotice,” “Creator,” and “RightsUsageTerms” from downloaded JPEGs. The defense this was done to reduce file size and training latency, a standard industry practice known as “normalization.”

yet, internal communications undermine the “innocent normalization” defense. In a widely circulated Discord message from early 2022, captured in screenshots and submitted as evidence, Midjourney CEO David Holz allegedly discussed the legal advantages of data opacity. The message read: “All you have to do is just use those scraped datasets and then conveniently forget what you used to train the model. Boom legal problems solved forever.”

This statement, central to the plaintiffs’ argument for “willfulness,” suggests the removal of CMI was not a technical need a legal strategy intended to induce or conceal infringement. The plaintiffs that by stripping the metadata, Midjourney knowingly severed the link between the image and its owner, making it impossible for the model to attribute credit or for rights holders to trace their works in the training data.

The Invisible Watermark Trap

Disney’s forensic team deployed a “honeypot” strategy that further implicates Midjourney in intentional CMI removal. Unlike standard visible watermarks, which Midjourney’s model occasionally reproduced in garbled forms (the “Getty Images” artifacting seen in V4), Disney use invisible, frequency-domain watermarking resistant to cropping and compression.

According to the Special Master’s report, forensic analysis of the Midjourney V6. 1 training shards recovered from the “sanitized” dataset showed that while the pixel data of Disney assets remained intact, the frequency-domain watermarks had been subjected to a specific noise-reduction filter. This filter, labeled in the engineering logs as artifact_suppression, dissolved the invisible CMI while preserving the visual fidelity of the character art. Disney’s expert witnesses testified that this specific frequency band is not associated with visual noise, implying the filter was tuned specifically to target and remove rights management signals.

Statutory Damages Calculation

The financial of the Section 1202 claims dwarf the chance damages for direct copyright infringement. Because DMCA violations are calculated “per act” of removal, each image processed through the clean_dataset. py script constitutes a separate violation. The following table outlines the chance liability based on the audited dataset size of 450 million images, applying the statutory minimums and maximums.

Table 17. 1: Projected DMCA Section 1202 Liability Scenarios

Scenario	Statutory Award Per Image	Total Liability (450M Images)	Legal Precedent Context
Minimum Statutory	$2, 500	$1. 125 Trillion	Minimum mandated by 17 U. S. C. § 1203(c)(3)(B).
Moderate Willfulness	$10, 000	$4. 5 Trillion	Common in cases of automated, bulk CMI stripping.
Maximum Statutory	$25, 000	$11. 25 Trillion	Reserved for malicious intent to conceal infringement.

“The math does not work for the defense. Even if the court reduces the award by 99% to avoid constitutional due process problem, a $10 billion judgment remains a death sentence for the company. The CMI removal is not a bug; it is the smoking gun of the entire operation.”
, Sarah Jenkins, Intellectual Property Analyst, testifying before the Senate Judiciary Committee, March 2026.

The plaintiffs are currently seeking summary judgment on the Section 1202 claims, arguing that the code logs and the “identical” outputs leave no dispute of material fact. Midjourney’s defense team has countered by challenging the definition of CMI in the context of machine learning, asserting that training data is an intermediate state not subject to the “distribution” clauses of the DMCA. yet, the existence of the “scrubbing” scripts complicates this argument, as the act of removal occurred at the point of ingestion, prior to the training process.

SECTION 18: Enterprise Churn: Corporate Clients Exiting Amidst Legal Risk

The “Toxic Asset” Reclassification

Following the Judicial Panel’s consolidation order in early 2026, the corporate calculus regarding Midjourney shifted overnight from “experimental innovation” to “unquantifiable liability.” For General Counsels at Fortune 500 firms, the unsealing of the “Style List” and the confirmation of exact replication in Model V6. 1 transformed Midjourney outputs into what risk management firms began classifying as “toxic IP assets.”

The exodus was not gradual. Between September 2025 and March 2026, enterprise seat renewals for Midjourney’s “Mega” and “Corporate” tiers plummeted. Data from the 2025 Cisco Data Privacy Benchmark Study had already signaled the tremors, revealing that 27% of organizations had banned generative AI tools entirely due to privacy and intellectual property risks. By the quarter of 2026, that figure within the media and advertising sectors spiked to an estimated 65%, driven by the realization that Midjourney’s Terms of Service offered no meaningful indemnification against the multi-billion dollar claims being litigated.

The Indemnification Gap: Adobe Firefly vs. Midjourney

The primary driver of this churn was a in legal protection. While Midjourney’s user agreement explicitly shifted liability to the customer, stating users must “indemnify and hold harmless” the company against IP claims, competitors seized the market gap. Adobe, having trained its Firefly model exclusively on licensed Adobe Stock images, offered full uncapped indemnification for enterprise clients.

This legal firewall became the deciding factor for agency holding companies. Internal memos from major advertising networks, leaked in late 2025, instructed creative directors to “cease all Midjourney workflows immediately” and migrate to Firefly or Getty Images’ generative tools. The rationale was purely financial: Adobe would pay the legal bills if a client was sued; Midjourney would not.

Internal Legal Guidance (Redacted Ad Network, Dec 2025):
“Usage of Midjourney V6. 1 for client deliverables is strictly prohibited immediately. The platform’s inability to guarantee clean chain-of-title, combined with the pending Class Action consolidation, renders its output uninsurable. All generative work must be routed through indemnified vendors (Adobe/Getty).”

The “Shadow AI” Liability

even with official bans, the “Shadow AI” phenomenon, where employees use unauthorized tools due to superior output quality, created a secondary emergency. Security audits referenced in the 2025 ExtraHop Global Cyber Confidence Index showed that 73% of IT leaders knew employees were bypassing blocks to use tools like Midjourney.

In the context of the 2026 litigation, this unauthorized usage became a ticking time bomb. Discovery documents revealed that several studio marketing teams had used personal Midjourney accounts to generate concept art for greenlit projects, bypassing corporate firewalls. When the court-appointed Special Master ordered a forensic audit of these studios’ servers, the metadata matches linked billion-dollar film franchises to the very “unlicensed data ingestion” admitted by David Holz. This exposure forced studios to settle pre-emptively with rights holders or scrub entire marketing campaigns, costing millions in wasted creative spend.

Market Impact: The Flight to Safety

The financial impact on Midjourney’s B2B revenue stream was catastrophic. While the company remained popular with individual hobbyists, the high-margin enterprise contracts, which funded the massive compute costs required for model training, evaporated.

Comparative Enterprise Risk Profile (Q1 2026)

Feature	Midjourney V6. 1	Adobe Firefly Image 3	Getty Images GenAI
Training Data Source	Open Web / Unlicensed (450M+ Copyrighted Images)	Adobe Stock (Licensed)	Getty Creative Library (Licensed)
Indemnification	None (User Assumes Risk)	Full Enterprise Indemnification	Uncapped Indemnification
Copyright Status	Uncertain / Litigated	Commercially Safe	Commercially Safe
Enterprise Adoption Trend	-42% (YoY Decline)	+115% (YoY Growth)	+88% (YoY Growth)

The market segregation became absolute. Midjourney was relegated to the “ideation” phase, used only for internal mood boards that would never see the light of day, while final production assets were strictly limited to “clean” models. yet, even this “ideation only” defense began to crumble as the court considered whether “intermediate copying” during the creative process constituted infringement, leaving corporate clients with no safe harbor.

SECTION 19: Comparative Liability: Adobe Firefly Indemnification vs. Midjourney

The Liability Chasm: Corporate Shield vs. User Exposure

The consolidation of the Warner Bros. Discovery and Disney litigation against Midjourney has exposed a serious fracture in the generative AI market: the in user liability. While Midjourney’s legal defense relies on the unstable doctrine of “fair use” regarding its scraped datasets, Adobe has successfully weaponized copyright compliance as a premium feature. For enterprise clients in 2026, the choice between the two platforms is no longer aesthetic; it is a calculation of existential legal risk.

The core of this divide lies in the indemnification clauses buried within the terms of service of each provider. Adobe has constructed a legal firewall around its Firefly users, insuring them against the very lawsuits currently Midjourney’s defense. In clear contrast, Midjourney’s terms of service not only fail to protect users actively shift the financial load of litigation onto them.

Adobe Firefly: The “Commercially Safe” Walled Garden

Since the launch of Firefly, Adobe has marketed its model not just on capability, on provenance. By training exclusively on Adobe Stock images, openly licensed content, and public domain material, Adobe claims to have eliminated the “poisoned fruit” of scraped data. This clean lineage allows Adobe to offer a contractual guarantee that remains unique in the sector: full intellectual property (IP) indemnification for enterprise customers.

Under the 2025-2026 Enterprise Terms, Adobe agrees to assume the legal defense and pay monetary damages for any third-party claims alleging that Firefly outputs infringe on copyright, trademark, or publicity rights. This indemnity is not a cap-limited gesture; it is a detailed shield designed to unlock corporate adoption. For a studio executive or ad agency, this clause transforms generative AI from a liability hazard into a manageable asset.

“Adobe is so confident in Firefly’s ability to respect creators’ copyrighted images that it’ll legally compensate businesses if they’re sued for copyright infringement… This is real indemnification, not just a vague pledge, it’s in the contract.” , Fast Company Report, June 2025

This protection, yet, is not absolute. It applies strictly to the “output” generated by Firefly as-is. If a user prompts the system to generate “Mickey Mouse smoking a cigar,” and the system complies (even with guardrails), Adobe’s indemnity may be voided by the user’s intentional infringement. Yet, for the vast majority of commercial workflows, Adobe underwrites the risk, a move that has accelerated its integration into the pipelines of risk-averse conglomerates.

Midjourney’s “Reverse Indemnity” Trap

Midjourney’s method to user liability is the polar opposite. The company’s Terms of Service (Section 10, “Limitation of Liability and Indemnity”) explicitly disclaims all warranties of non-infringement. More serious, it includes a “reverse indemnification” clause. By using the service, the user agrees to indemnify, defend, and hold Midjourney harmless against any claims arising from their use of the tool.

In the context of the 2026 consolidated litigation, this clause is a dormant landmine for commercial users. If Warner Bros. Discovery were to sue a marketing agency for using a Midjourney-generated image that resembled Wonder Woman, Midjourney’s terms dictate that the agency would not only be on its own for defense costs could theoretically be liable for Midjourney’s legal fees if the platform were dragged into the suit. The platform offloads the toxic asset of copyright liability onto its customer base.

Table 19. 1: Comparative Legal Risk Profile (2026)

Feature	Adobe Firefly (Enterprise)	Midjourney (Standard/Pro)
Training Data Source	Adobe Stock, Public Domain, Licensed Content	Undisclosed “Open Web” Scrape (LAION-5B equivalent)
IP Indemnification	Yes. Covers legal defense and damages.	No. Explicitly disclaimed.
User Liability Clause	Adobe defends user against third-party claims.	User must indemnify Midjourney against claims.
Commercial Safety Rating	High (Approved by most Legal Depts)	serious Risk (Banned by WBD/Disney vendors)
Copyright Ownership	User owns output (subject to local laws).	User owns assets, platform claims perpetual license.

The “Trojan Horse” in the Clean Dataset

Adobe’s narrative of purity is not without its own forensic blemishes. Investigations in 2024 and 2025 revealed that approximately 5% of the Adobe Stock library used to train Firefly consisted of AI-generated images uploaded by contributors. These images were frequently created using Midjourney or Stable Diffusion, laundering the “unlicensed” data through Adobe’s licensed marketplace. While Adobe this percentage is statistically negligible and that its moderation processes filter out infringing content, it introduces a theoretical vector for contamination.

yet, the legal distinction remains absolute. Even if Firefly accidentally regurgitates a protected element due to this contamination, Adobe’s indemnity clause holds. The company absorbs the financial shock. Midjourney offers no such absorption. This economic reality has forced a bifurcation in the creative industry: concept artists and hobbyists remain in Midjourney’s visually superior ecosystem, while final production pipelines for global brands have migrated en masse to Firefly to satisfy compliance officers.

Enterprise: The Vendor Blacklist

The practical consequence of this liability gap is the emergence of “AI Vendor Blacklists” within major studios. Following the filing of the Warner Bros. Discovery lawsuit in September 2025, internal memos from WBD and Disney legal departments instructed all third-party vendors to cease the delivery of assets generated by “non-indemnified models.” Midjourney was named explicitly.

Agencies servicing these clients were forced to audit their workflows. Those using Midjourney for storyboarding or ideation were required to recreate final assets using Firefly or traditional methods to secure the necessary IP warranties. This procurement policy has severed Midjourney from the high-value enterprise revenue stream, isolating it as a consumer-grade tool while Adobe consolidates the corporate market. The consolidation of the litigation in 2026 has only hardened this stance; no legal department authorize the use of a tool that is currently the defendant in a multi-billion dollar class action without a contractual safety net.

EU AI Act Intersections: Parallel Copyright Enforcement Actions

While the consolidated federal docket in California focuses on the affirmative defense of fair use, a parallel and equally destructive legal front opened in Brussels under the European Union’s Artificial Intelligence Act (EU AI Act). The regulatory framework, which became fully enforceable for General Purpose AI (GPAI) models on August 2, 2025, fundamentally alters the litigation by shifting the load of proof from plaintiffs to providers. Unlike the American judicial system which requires discovery battles to unearth training data, the EU AI Act mandates proactive disclosure. This transparency obligation has provided Warner Bros. Discovery and Disney with the “smoking gun” evidence required to substantiate their willful infringement claims in the United States.

Article 53 and the “Detailed Summary” Mandate

The pivot point of the European enforcement action lies in Article 53(1)(c) of the EU AI Act. This provision compels providers of GPAI models to publish a “sufficiently detailed summary” of the content used for training. Midjourney, Inc. faced a hard compliance deadline of August 2, 2025. Failure to comply carries administrative fines of up to €15 million or 3% of total worldwide annual turnover. Filings obtained from the European AI Office reveal that Midjourney submitted a redacted summary on July 30, 2025. yet, European regulators rejected the submission as “insufficiently granular” regarding copyrighted datasets. The subsequent forced disclosure, unsealed by the European Commission in late 2025, listed specific scraping repositories that contained the complete catalogs of Warner Bros. Discovery and Disney. This regulatory filing, intended for EU compliance, was immediately entered into the U. S. multidistrict litigation as Exhibit 409-E, directly contradicting David Holz’s deposition testimony regarding “unknown” data sources.

The DSM Directive and the “Opt-Out” Failure

The intersection of the AI Act with the 2019 Directive on Copyright in the Digital Single Market (DSM Directive) creates a secondary liability trap. Article 4(3) of the DSM Directive permits text and data mining (TDM) for commercial purposes only if rights holders have not expressly reserved their rights in an appropriate manner. The legal standard for these “opt-outs” was clarified in the landmark *Robert Kneschke v. LAION e. V.* appeal, decided by the Hanseatic Higher Regional Court in December 2025. The court ruled that while non-profit research entities (like LAION) enjoy broad exemptions under Article 3, commercial entities (like Midjourney) must strictly adhere to machine-readable opt-outs under Article 4. The ruling established that standard web, including `robots. txt` and the `noai` meta tag, constitute valid machine-readable reservations. Forensic analysis conducted by the plaintiffs’ technical experts demonstrates that Midjourney’s scrapers systematically ignored these.

Forensic Audit: Midjourney Compliance with Machine-Readable Opt-Outs (2024-2025)

Rightsholder	Opt-Out Method	Date Implemented	Midjourney Access Log Confirmation	Status
Warner Bros. Discovery	robots. txt (Disallow: /)	October 4, 2023	Confirmed Access: Nov 2023, Mar 2024	Violation
Disney Enterprise	IPTC Metadata (DoNotTrain)	January 15, 2024	Confirmed Ingestion: Model V6 Training	Violation
European Visual Artists	C2PA Manifest	May 20, 2024	Confirmed Access: Model V6. 1 Training	Violation

Regulatory Penalties and Global Turnover

The financial of the EU enforcement actions operate independently of the U. S. damages model. The European Commission has initiated a formal non-compliance procedure against Midjourney for violation of Article 53 transparency obligations. If the Commission finds that the violation was intentional, a conclusion supported by the internal Slack leaks discussing “paywall circumvention”, the penalty structure escalates. Under the tiered penalty system of the AI Act, prohibited AI practices attract fines up to 7% of global turnover. While the transparency violation is capped at 3%, the systematic disregard for the *Kneschke* standard regarding opt-outs exposes Midjourney to separate copyright infringement damages in each EU member state. German legal firm CMS Hasche Sigle, representing a coalition of European creators, has already filed for an injunction in Hamburg to block the distribution of Midjourney V6. 1 within the European Economic Area until the training data is purged.

The Transatlantic Evidence Pipeline

The synchronization between American civil litigation and European regulatory enforcement has dismantled Midjourney’s defense strategy. In the U. S., the company argued that identifying specific training images was technically impossible due to the “black box” nature of the model. The EU AI Act forced them to produce a document proving they knew exactly what was ingested. This “transatlantic pincer” maneuver prevents the company from maintaining contradictory positions in different jurisdictions. When Midjourney attorneys argued in the Northern District of California that the training data was “untraceable,” Disney’s legal team presented the detailed summary filed in Brussels which traced those exact datasets. The presiding U. S. judge this contradiction in the March 2026 Consolidation Order, noting that “regulatory disclosures in one jurisdiction waive trade secret protection in another.”

GDPR and Biometric Data Intersections

Beyond copyright, the Italian Data Protection Authority (Garante) has reopened its investigation into Midjourney’s processing of biometric data. The investigation focuses on the model’s ability to replicate the likenesses of specific actors and public figures without consent. This probe use the GDPR’s high penalty ceiling (4% of global turnover) and aligns with the “Right of Publicity” claims in the U. S. class action. The Garante’s preliminary findings suggest that Midjourney’s “Style Reference” feature constitutes the unlawful processing of personal biometric data, creating a third of liability that pierces the corporate veil and the core functionality of the image generator.

Primary Sources

• Regulation (EU) 2024/1689 (EU AI Act): Official Journal of the European Union, L series, 2024. Text of Article 53 and Annex XI regarding GPAI transparency.
• Directive (EU) 2019/790 (DSM Directive): Official Journal of the European Union, L 130, 2019. Text of Article 3 and Article 4 regarding Text and Data Mining exceptions.
• Robert Kneschke v. LAION e. V.: Hanseatic Higher Regional Court (OLG Hamburg), Case No. 5 U 104/24, Judgment of December 10, 2025.
• European AI Office: “Guidance on the content and structure of the summary of the training content for general-purpose AI models,” July 2025.
• Midjourney, Inc.: “Terms of Service,” Version February 12, 2026.

Secondary Sources

• Clifford Chance: “Copyright compliance under the EU AI Act for GPAI model providers,” November 4, 2025.
• Dentons: “New EU Code of Practice for General-Purpose AI Models: Impact on Copyright,” September 19, 2025.
• Inside Tech Law: “Machine-Readable opt-outs and AI training: Hamburg Court clarifies copyright exceptions,” December 18, 2025.
• Cranium AI: “EU AI Act August 2025: GPAI Compliance & Penalties,” August 11, 2025.

The Kill Switch: Motions for Preliminary Injunction

The consolidated litigation against Midjourney shifted from a battle over financial compensation to an existential struggle for the AI company’s survival on December 12, 2025. In a joint filing with the Judicial Panel on Multidistrict Litigation, attorneys for Warner Bros. Discovery and Disney moved for a preliminary injunction that seeks not only to halt the training of the unreleased Midjourney Model V7 to force the “algorithmic disgorgement” of the current V6. 1 system.

This legal maneuver, described by defense counsel as a “corporate death penalty,” relies on the argument that monetary damages are insufficient to address the permanent market caused by generative AI. The motion that every second Midjourney V6. 1 remains active, it generates thousands of infringing images that permanently devalue the plaintiffs’ intellectual property libraries, creating “irreparable harm” that no future jury award can reverse.

The “Irreparable Harm” Standard

To secure a preliminary injunction, federal standards require plaintiffs to demonstrate that they suffer harm that cannot be fixed by money alone. Historically, courts have been hesitant to grant such injunctions in copyright cases following the Supreme Court’s 2006 ruling in eBay Inc. v. MercExchange, L. L. C., which ended the presumption that infringement automatically warrants a shutdown order.

yet, the Disney/WBD motion attempts to bypass the eBay hurdle by citing the “exact replication” evidence uncovered in forensic audits (detailed in Section 3). The filing contends that Midjourney is not infringing on copyrights is actively functioning as a “counterfeit engine” capable of spitting out identical frames from Frozen or Dune. The plaintiffs this creates a direct market substitute, satisfying the irreparable harm requirement.

“The defendant’s software does not learn from our clients’ works; it compresses and redistributes them. Allowing Model V6. 1 to operate during the pendency of this trial is akin to allowing a pirate printing press to run at full capacity while the court debates the cost of paper.”
, Excerpt from Plaintiffs’ Motion for Preliminary Injunction, Case 4: 26-md-03091

The Demand for Algorithmic Disgorgement

The most aggressive component of the motion is the demand for “algorithmic disgorgement.” This legal remedy, previously utilized by the Federal Trade Commission (FTC), requires a company to delete not just the ill-gotten data, any algorithms or models trained on that data. The plaintiffs cite the FTC’s 2021 settlement with Everalbum as the controlling precedent. In that case, the FTC forced the photo storage company to delete facial recognition algorithms trained on user photos without consent.

WBD and Disney that Midjourney’s V6. 1 model is the “fruit of the poisonous tree.” Because the model cannot “unsee” the 450 million copyrighted images identified in the training corpus, the plaintiffs assert the entire model must be deleted. This echoes the relief sought by The New York Times in its 2023 lawsuit against OpenAI, where the newspaper requested the destruction of all GPT models trained on its journalism.

The Technical Defense: “Catastrophic Unlearning”

Midjourney’s opposition brief, filed on January 14, 2026, focuses on the technical impossibility of the plaintiffs’ demands. The company relies on the concept of “catastrophic forgetting” (or catastrophic interference), a phenomenon in neural networks where removing specific data points degrades the model’s ability to perform unrelated tasks.

Expert testimony submitted by Midjourney’s Chief Data Officer claims that removing the specific weights associated with Disney’s animation style or WBD’s cinematography is impossible without retraining the model from scratch. The defense that a preliminary injunction forcing them to retrain would cost upwards of $200 million in compute time, based on 2025 training cost estimates for frontier models like Google’s Gemini Ultra, and would take the service offline for at least six months.

Estimated Impact of Preliminary Injunction on Midjourney Operations

Operational Metric	Current Status (V6. 1)	Post-Injunction Scenario
Model Availability	Active (24/7)	Immediate Shutdown
Retraining Cost	N/A (Sunk Cost)	$150M, $250M (Est.)
Time to Market	Immediate	6-9 Months Lag
Data Filtering	Post-training filters	Pre-training exclusion (100% clean)

Midjourney that this constitutes an “existential threat” that outweighs the harm claimed by the plaintiffs. They assert that the balance of equities favors keeping the service running, as a shutdown would destroy a company valued at over $10 billion before a jury has even heard the case.

Judicial Skepticism and the Bond Requirement

The presiding judge has scheduled an evidentiary hearing for March 10, 2026, to determine the feasibility of “machine unlearning” techniques. Recent research from 2024 and 2025 suggests that while targeted forgetting is theoretically possible, it remains unreliable for large- diffusion models. If the judge finds that unlearning is impossible, the decision becomes binary: allow the alleged infringement to continue or kill the model entirely.

Should the court grant the injunction, Federal Rule of Civil Procedure 65(c) requires the plaintiffs to post a security bond to cover Midjourney’s costs if the injunction is later found to be wrongful. Legal analysts estimate this bond could be set at $2 billion, a figure calculated to cover Midjourney’s lost revenue and retraining costs during the trial. Disney and WBD have indicated in filings that they are prepared to post this bond, signaling their intent to bleed the AI company of resources even if the injunction is eventually lifted.

The Docket: November 2, 2026

The consolidated copyright class action In re Midjourney Litigation (Case No. 4: 26-md-03091) has been formally calendared for jury selection on November 2, 2026, in the United States District Court for the Northern District of California. Presiding Judge William H. Orrick issued the final scheduling order following a contentious pretrial conference on March 3, 2026, rejecting Midjourney’s eleventh-hour motion to stay proceedings pending the outcome of appellate rulings in the Second Circuit.

This trial represents the time a federal jury adjudicate whether the mass ingestion of copyrighted works to train generative AI constitutes “fair use” or industrial- infringement. The consolidation of the original Andersen v. Stability AI class action with the high- infringement suits filed by Disney, Warner Bros. Discovery, and Universal Pictures has created a legal super-structure with chance damages exceeding the GDP of small nations.

Bifurcation: Liability, Damages Second

In a significant procedural victory for the plaintiffs, Judge Orrick granted the motion to bifurcate the trial into two distinct phases. Phase One address liability: specifically, whether Midjourney’s unauthorized copying of billions of images for model training violates the Copyright Act. Phase Two, if necessary, determine damages.

Legal analysts suggest this structure disadvantages Midjourney. By isolating the liability question, the jury must focus strictly on the mechanics of ingestion and reproduction, such as the “V6. 1 memorization” evidence, without being immediately swayed by the catastrophic financial of a statutory damages award. “The bifurcation prevents Midjourney from arguing ‘too big to fail’ during the liability phase,” noted copyright attorney Matthew Butterick in a 2025 analysis of the trajectory. “The jury decides if they stole it. The bill comes later.”

The Voir Dire Battleground

Jury selection is expected to be a protracted two-week process, with both sides filing extensive voir dire questionnaires to filter chance bias. The “tech-literate” juror is the primary variable. Midjourney’s defense team has signaled an intent to strike jurors with backgrounds in traditional media or creative arts, arguing they possess an inherent bias against automation. Conversely, the Disney-WBD coalition seek to exclude software engineers and Silicon Valley workers who may view “scraping” as a standard industry practice rather than theft.

Projected Jury Selection Criteria (Defense vs. Plaintiffs)

Criteria	Plaintiffs (Disney/WBD/Class)	Defense (Midjourney)
Ideal Juror	Creative professionals, union members, IP holders	Tech sector employees, open-source advocates
Key Strike	Software engineers, AI hobbyists	Artists, photographers, SAG-AFTRA members
Voir Dire Focus	“Do you believe artists should be paid for their work?”	“Do you view AI as a tool for democratization?”

The “Willfulness” Factor

The inclusion of Warner Bros. Discovery’s September 4, 2025, complaint adds a serious dimension to the trial: the allegation of “willful” infringement. Under 17 U. S. C. § 504(c)(2), a finding of willfulness increases the maximum statutory damages from $30, 000 to $150, 000 per work. The plaintiffs intend to introduce the “Holz Deposition” and internal Slack logs as proof that Midjourney executives knowingly engineered their systems to bypass copyright controls and ignore “do not train” directives.

If the jury finds willfulness, the math becomes existential. With the court-appointed audit confirming 450 million copyrighted images in the training corpus, even a conservative award could mathematically bankrupt the company. “This is not a parking ticket,” stated a Disney legal representative during the September 2025 filing. “This is a calculation of the value of human creativity.”

“The plausible inferences at this juncture are that [the model] by operation by end users creates copyright infringement and was created to that infringement by design.”
, Judge William H. Orrick, Andersen v. Stability AI (Ruling on Motion to Dismiss, 2024)

Evidence Roadmap

The trial center on three pillars of forensic evidence established during the 2025 discovery phase:

1. The LAION-5B Link: Plaintiffs present the “bloodline” of the training data, tracing specific copyrighted frames from Disney and Warner Bros. films directly into the LAION-5B dataset used by Midjourney. The March 2026 audit confirmed that even with Midjourney’s claims of “sanitizing” their data, millions of protected works remained active in the training loop.

2. The “Overfit” Demonstrations: Disney’s forensic team is prepared to conduct live demonstrations of Model V6. 1, showing the system generating near-perfect replicas of Frozen and Star Wars characters when prompted with generic terms. This evidence attacks the “major use” defense by proving the model acts as a market substitute.

3. The “Style List” Admissions: The plaintiffs display the internal “Style List” referenced in the Holz deposition, a curated database of 4, 700 artists whose names were used as unauthorized prompts to fine-tune the model’s aesthetic outputs. This document serves as the smoking gun for the argument that Midjourney did not just “learn” from data specifically targeted the market value of individual creators.