AI-Generated Assets in Games: Copyright, Ownership, and the Risks Studios Are Walking Into

The Copyright Office Position: AI Outputs Are Not Copyrightable by Default

In January 2025, the U.S. Copyright Office issued its most direct statement yet on AI-generated content: "Copyright does not extend to purely AI-generated material, or material where there is insufficient human control over the expressive elements. Based on the functioning of current generally available technology, prompts do not alone provide sufficient control." That conclusion applies directly to every sprite, soundtrack, and line of generated code in your game.

The federal courts have reinforced this position. In Thaler v. Perlmutter, 687 F. Supp. 3d 140 (D.D.C. 2023), the U.S. District Court for the District of Columbia held that "human authorship is a bedrock requirement of copyright" and that "copyright has never stretched so far [as] to protect works generated by new forms of technology operating absent any guiding human hand." The D.C. Circuit affirmed that ruling in March 2025, making it the controlling appellate authority on the question.

The Copyright Office's 2023 Zarya of the Dawn decision sharpens the practical implication. The Office had registered a graphic novel that used Midjourney-generated images. On review, it cancelled protection for the individual AI-generated images — even though the creator had crafted detailed, iterative prompts — because the creator "lacked sufficient control over" the Midjourney outputs. What the Office did protect was the human-authored text and the creative selection and arrangement of the work's visual elements. For game studios, the translation is direct: your game's narrative, level design logic, and the creative assembly of assets may qualify for protection, but individual AI-generated character art, environments, and audio tracks do not.

The January 2025 Part 2 report does identify pathways to protection. Human authors retain copyright in their own expressive work that is perceptible in an AI output, in creative arrangements of AI material, and in sufficiently creative modifications to AI outputs. But the Office is explicit that each case must be analyzed individually — there is no blanket rule that AI-assisted work is protected simply because a human was involved in generating it. The threshold question is whether human authorship, specifically, is what produced the expressive content.

The Ownership Gap: What This Means for Game Studios

The copyright gap does not stay abstract for long. If your AI-generated assets lack copyright protection, competitors are legally free to copy or imitate them — leaving your studio without the primary tool used to prevent look-alike games. Intellectual property has always been central to game developers' competitive advantage, and the inability to enforce copyright on core visual assets fundamentally changes the leverage equation for studios building their first significant IP.

The deal risk is equally concrete. Publisher agreements and acquisition term sheets typically require the studio to represent that it owns all relevant IP and that no asset infringes third-party rights. When AI-generated content enters the picture, both representations become difficult to make cleanly. As games-specialist firm Brabners has observed, AI IP ownership issues discovered during technical due diligence create a predictable cascade: delays, renegotiation (including price reductions), and — in worst-case scenarios — abortive deals. For an indie studio that has spent three years building toward a publishing deal, an undisclosed AI asset stack is a material deal risk.

Acquirers have started building AI-specific language into their diligence frameworks to address this. Squire Patton Boggs has noted that the Copyright Office's position creates "a critical protection gap for acquirers, as AI outputs may lack enforceable IP rights regardless of the target's claims about ownership." Deal teams are now requesting training data provenance records, documentation of human creative modifications, and enhanced warranty provisions that standard acquisition agreements do not include. A studio that cannot produce this documentation during diligence will find itself negotiating from weakness.

There is also an insurance dimension worth noting, though game-specific policy data is limited. ISO has introduced AI-specific exclusions to commercial general liability policies that exclude coverage for personal and advertising injury arising from generative AI outputs — including copyright infringement exposure that traces to training data. Studios that assume their CGL policy covers AI-content claims may be operating with an uncovered exposure. The three risk vectors — competitive, transactional, and insurance — reinforce each other, and none of them resolves without addressing the underlying ownership question.

Midjourney, Stable Diffusion, DALL-E: What the Licenses Actually Say

A useful way to compare the major AI image platforms is by their IP indemnification posture — specifically, which platforms will stand behind their outputs if a third party sues you for copyright infringement. Ranked by that criterion, the hierarchy runs from Adobe Firefly at the top to open-source Stable Diffusion at the bottom, with OpenAI and Midjourney occupying distinct positions in between.

Adobe Firefly is currently the only major AI image platform that provides IP indemnification to paid subscribers. Adobe trains Firefly exclusively on licensed Adobe Stock content, openly licensed content, and public domain material — meaning its training pipeline is structured to avoid the infringement exposure that plagues other platforms. Qualifying paid Creative Cloud subscribers receive indemnification against copyright claims arising from Firefly outputs, with enterprise tiers providing higher coverage caps. This makes Firefly the default recommendation for any studio concerned about downstream IP risk.

OpenAI (DALL-E / GPT Image) takes the next-strongest contractual position: it explicitly assigns all right, title, and interest in outputs to the user and permits commercial use across all tiers. On Business, Team, and API tiers, OpenAI commits not to use content for training. The gap is indemnification — OpenAI does not offer IP protection against third-party infringement claims, meaning studios bear 100% of the infringement risk themselves. Studios should also note that DALL-E 3 is being retired on May 12, 2026, with GPT Image 1.5 as the replacement.

Midjourney grants paid subscribers commercial ownership of outputs "to the fullest extent possible under applicable law" — but that limiting phrase matters. If U.S. copyright law does not protect the output (which is the Copyright Office's current position), Midjourney's contractual grant cannot manufacture copyright protection where none legally exists. Midjourney also retains a perpetual, irrevocable license to reproduce and sublicense user inputs and outputs, and requires a Pro or Mega plan for companies earning over $1,000,000 annually. No IP indemnification is offered, and Midjourney is a named defendant in the ongoing Andersen v. Stability AI training data litigation.

Stable Diffusion presents the highest-risk profile. The open-weights model gives studios maximum flexibility to run locally, but it carries no commercial indemnification, no IP warranty, and unresolved litigation risk. In August 2024, U.S. District Judge William Orrick refused to dismiss the core copyright claims against Stability AI, Midjourney, and related defendants in Andersen v. Stability AI, No. 3:2023cv00201 (N.D. Cal.), finding that artists may pursue claims based on the LAION dataset of 5 billion scraped images used to train the model. That risk follows the model wherever it runs — including on-premises.

Music and Audio: Suno, Udio, and the Training Data Problem

In June 2024, Sony Music Entertainment, UMG Recordings, and Warner Records filed copyright infringement lawsuits against Suno and Udio in federal courts in Boston and New York respectively. The RIAA's complaints alleged that both services had copied "decades worth of the world's most popular sound recordings" without permission or licensing to train their models, and described both companies as deliberately evasive about what they had copied. The labels sought injunctions that could have forced these platforms to halt operations.

The litigation has partially resolved, but not in a way that clears the risk for studios using these platforms today. UMG settled with Udio in October 2025, and Warner settled with Suno in November 2025 — but Sony has not settled with either company. Sony's fair-use cases against Suno (District of Massachusetts) and Udio (Southern District of New York) remain pending, with a Suno summary judgment hearing scheduled for July 2026. A ruling in that proceeding could set binding precedent for the AI music industry. Studios that read about the major-label settlements and concluded the litigation was over are operating on an incomplete picture.

The major-label settlements also leave a separate risk category untouched. Independent musician class actions — Nguyen v. Suno and Kim v. Uncharted Labs (the corporate entity behind Udio) — allege that UMG and Warner secured compensation for themselves but left independent artists without any recovery. If those class actions succeed, studios that have shipped games using Suno or Udio-generated audio could face secondary liability exposure that the major-label settlements do not insulate them from.

For game studios that need AI-generated music now, commercially licensed platforms such as Soundraw and Artlist AI provide explicit commercial licenses and IP representations for generated tracks, and neither is named in the RIAA training data litigation. Note that the specific current license terms for these platforms should be verified directly before use — licensing terms in this space shift frequently. The practical floor is to avoid Suno and Udio for any commercially shipped title until Sony's cases resolve, and to document the licensing status of every AI-generated audio asset currently in your build.

Code: GitHub Copilot, Training Data, and the Open-Source Contamination Risk

The most underappreciated risk in AI-assisted game development involves code, not art. When a developer accepts a GitHub Copilot suggestion, they may be incorporating verbatim or near-verbatim GPL-licensed code into a commercial codebase. The consequence is not just a copyright infringement claim — it is the possible obligation to release your entire commercial game's source code. GPL's copyleft provisions work on a taint basis: if GPL-licensed code is incorporated into a commercial product, the GPL's license terms require that the entire product be distributed under GPL, meaning open source. For a studio building a commercial game, that is an existential risk.

The Does v. GitHub class action (Nos. 4:22-cv-06823-JST & 4:22-cv-07074-JST, N.D. Cal.) established that this is a live legal question, not a theoretical one. Open-source programmers sued GitHub, Microsoft, and OpenAI, alleging that Copilot reproduces GPL-licensed code without attribution. The district court dismissed most DMCA claims in 2024, but breach-of-license and contract claims survived and are now on appeal to the Ninth Circuit (No. 24-6136). District court proceedings are stayed pending that appeal. The DMCA dismissals do not resolve the GPL contamination risk — the breach-of-license theory, which is the more structurally dangerous claim for commercial studios, is the one that is still alive.

GitHub itself has acknowledged the risk through its duplication detection filter. The filter detects and suppresses Copilot suggestions containing code blocks of 65 lexemes or more (roughly 150+ characters) that match public code on GitHub. When the filter is set to its "Block" mode, GitHub will defend commercial users against third-party copyright claims. That defense obligation is explicitly conditional: per GitHub's own documentation, "for commercial use, any GitHub defense obligations related to your use of GitHub Copilot do not apply if you have not set the Duplicate Detection filtering feature available in GitHub Copilot to its 'Block' setting." Studios using Copilot without the Block filter enabled have forfeited GitHub's contractual protection.

For studios that want indemnification built into the product rather than contingent on a setting, Amazon Q Developer Pro ($19/user/month) offers IP indemnification against third-party copyright infringement claims arising from its code suggestions. AWS will defend Q Developer Pro subscribers against such claims so long as they allow AWS to control the defense and settlement. The free tier of Q Developer does not include indemnification. For commercial studios, the choice between Copilot (conditional defense obligations, active litigation on appeal) and Q Developer Pro (affirmative indemnification, $19/month) is a straightforward risk-adjusted decision.

Human Authorship Strategies: Getting Copyright Protection for AI-Assisted Work

The Copyright Office's January 2025 Part 2 report establishes three pathways to copyright protection for AI-assisted work, and it is useful to think of them as a spectrum from weakest to strongest. At the weakest end: generating an output from a prompt provides no copyright protection, regardless of how detailed or iteratively refined the prompt is. The Office stated directly that "prompts essentially function as instructions that convey unprotectible ideas." Selecting one output from a batch of generated images is also insufficient — the Office held that "selection of a single output is not itself a creative act."

The three actionable pathways are selection and arrangement, creative modification, and human-expressive inputs. Selection and arrangement — organizing AI-generated outputs in a creative, non-obvious structure — can produce thin copyright protection in the arrangement itself, similar to how a photograph of public domain art can be protected by the creative choices of framing and composition. For game studios, this means the creative assembly of a world, level, or narrative structure from AI-generated components may qualify for protection even when the individual components do not.

Creative modification provides stronger protection. When a human artist substantially edits an AI-generated character or environment — reworking proportions, adding expressive detail, painting over outputs — the resulting modifications are protectable human authorship. The scope of protection tracks the scope of the human modifications, similar to a derivative work. The key is that the modifications must themselves meet the originality threshold; mechanical adjustments to brightness or scale do not count.

The strongest pathway is using human-created expressive work as the AI input where that work is perceptible in the output. The Copyright Office's Part 2 report uses the Rose Enigma registration as its example: a human-drawn illustration used as an AI input, with the expressive elements of the drawing clearly perceptible in the generated output. The Office confirmed that the drawing's copyright extended to those perceptible elements in the output. For game artists, this translates to a concrete workflow: sketch or paint the character, environment, or design concept; feed that human-authored work into the AI model; and document the correspondence between the input and the output.

Whatever strategy you use, documentation is load-bearing. Odin Law's AI governance guide for game developers recommends maintaining detailed audit logs of all prompts, outputs, and human edits, and retaining video evidence of the generation process when possible. Platforms including Steam, Apple, and Google now require AI use declarations, and the same documentation that supports a copyright claim will also satisfy those disclosure requirements.

Practical Risk Management: What Studios Should Do Now

The analysis across sections 1–6 points to four discrete action items, organized by asset type. Each one can be acted on today without waiting for the law to stabilize further.

Art and visual assets. Audit every AI-generated image asset in your game: which tool generated it, what ToS version was in effect, what human modifications were made, and whether the generation process was documented. Going forward, rank your image tools by IP safety: Adobe Firefly first (IP indemnification, licensed training data), OpenAI GPT Image second (ownership assignment to user, no indemnification), Midjourney paid third (commercial license capped by applicable law, no IP warranty, active Andersen litigation), open-source Stable Diffusion last (no indemnification, live training data litigation). Firefly's indemnification is the clearest available protection for studios that cannot absorb infringement litigation costs.

Music and audio. Remove Suno and Udio from your production pipeline for any commercially shipped title until Sony's pending fair-use claims resolve — the July 2026 Suno summary judgment date is the minimum milestone to watch. If you have already shipped a title using Suno or Udio audio, document the specific tracks, assess your replacement options before your next update or DLC release, and monitor the Sony litigation closely. Safer alternatives include Soundraw and Artlist AI, both of which offer commercial licenses and IP representations — verify the current terms directly from each platform before committing.

Code. If your team uses GitHub Copilot, enable the duplication detection filter in Block mode immediately. As GitHub's own documentation states, Copilot's defense obligations for commercial users do not apply unless that setting is active. Consider whether the uncovered risk profile of Copilot (conditional defense obligations, active Ninth Circuit appeal in Does v. GitHub) justifies migrating to Amazon Q Developer Pro ($19/user/month), which provides affirmative IP indemnification. For most commercial studios, $19/user/month for indemnification is a straightforward cost-benefit calculation.

Publisher and acquisition readiness. Before any publisher deal or acquisition due diligence, prepare an AI asset disclosure package: a complete inventory of AI-generated assets organized by tool, applicable license terms, training data status of each tool, and documentation of human creative modifications. Acquirers are now requesting this specifically, and a studio that can produce it proactively — rather than in response to a diligence request — signals competence and reduces the risk of deal disruption. The financial consequences of IP infringement exposure discovered mid-diligence, as Brabners has documented in the games context, range from price reductions to deal collapse.