Big Walk game AI is the query thousands of American readers typed this week after the Australian studio behind the co-op hit Big Walk said it used artificial intelligence tools sparingly during the final stretch of development to fix difficult technical problems. The statement landed in the middle of a louder fight: jurors for the 2027 Independent Games Festival said they would not vote for any game developed with generative AI, while developers who quietly use these tools for bugs, polish, and production support suddenly had to explain themselves. For players the question is curiosity. For studios, contractors, publishers, and buyers, it is money and ownership: when AI tools helped make a game, what does United States copyright actually protect, who owns the AI-assisted parts, and what paperwork keeps a registration, a sale, or a distribution deal from collapsing later? This guide answers under federal law as judges and the Copyright Office have stated it, starting from three decided cases, the text of the Copyright Act, and the Office’s own guidance on AI-generated material. It covers the human-authorship rule, what applicants must disclose when they register, and the separate training-data risk on the input side. It does not cover Australian law, even though the studio is Australian, and festival voting rules are private policies, not copyright law.
I. Can a game built with AI tools get U.S. copyright protection?
The short answer is yes, in part. Federal copyright can protect the human-authored portions of a game built with AI assistance, such as code a programmer wrote, art a designer drew and edited, music a composer shaped, and the creative selection and arrangement of elements into a finished work. It cannot protect material that no human authored. The dividing line runs through every AI-assisted project, and the cases below draw it in practical terms a studio can apply before it files anything.
A. Human authorship is the gatekeeper, and prompting alone does not satisfy it
In Thaler v. Perlmutter, 687 F. Supp. 3d 140 (D.D.C. 2023), the District Court for the District of Columbia held that “Human authorship is a bedrock requirement of copyright” and sustained the Copyright Office’s denial of registration for an image the applicant said his computer system generated autonomously, with the applicant listed as owner but the machine identified as author, in a decision issued August 18, 2023. For a game team, the practical lesson is that the studio must be able to point to specific human creative choices behind every element it claims to own. A programmer’s original code, a level designer’s hand-tuned layouts, an artist’s painted-over and composited sprites, and a writer’s edited dialogue all reflect the kind of guiding human hand the court described. A texture, quest line, or soundtrack produced end to end by a tool, reviewed only to confirm it looks acceptable, does not.
The statute behind that result is short. Section 102(a) of the Copyright Act, 17 U.S.C. section 102(a), provides that “Copyright protection subsists, in accordance with this title, in original works of authorship fixed in any tangible medium of expression,” a sentence courts have read for decades to require an author in the ordinary human sense. The Copyright Office applies the same reading in the Compendium chapter on copyrightable authorship, which states that the Office “will not register works produced by a machine or mere mechanical process that operates randomly or automatically without any creative input” from a human author. Random generation, automatic rendering, and one-click output fall on the unprotectable side of that sentence. Tools that extend a human author’s expression, such as a brush a painter chose and controlled stroke by stroke, fall on the other side, and most game-production uses sit somewhere between, which is why documentation of who did what matters as much as the tool used.
The court of appeals left that result in place in Thaler v. Perlmutter, No. 23-5233 (D.C. Cir. Mar. 18, 2025), decided March 18, 2025, by the United States Court of Appeals for the District of Columbia Circuit, affirming the district court’s denial of registration for the machine-generated image discussed at 687 F. Supp. 3d 140 on the ground that the Copyright Act requires every eligible work to be authored in the first instance by a human being, which confirms that prompting alone does not make the prompter the author of the output. Note what the court did not decide: it expressly left open how much human involvement suffices when a person uses AI as one step in a larger creative process, because that question was not presented by an applicant who disclaimed any human contribution. A studio in that middle ground therefore cannot cite either Thaler decision as a clearance. It must show its own human authorship, element by element, under the Office’s disclosure rules described next.
One registration path deserves special attention from game teams because it matches how games are actually assembled. The Copyright Office has treated the selection and arrangement of AI-generated images with human-authored text as protectable compilation authorship, as in an early case involving a comic book that combined a human-written story with generated illustrations, where the Office protected the compilation as a whole though not the generated images alone. A game is a larger version of the same pattern: human-written code, dialogue, and design documents organized with art, levels, and sequences into a playable whole. The compilation claim covers the creative choices in selecting, coordinating, and arranging those elements, even where individual generated pieces fall outside the claim. That protection is thinner than authorship of every asset, since a competitor that copies only an unclaimed generated texture takes nothing the registration covers, but for most studios the arrangement is precisely where the commercial value lives. Describe that arrangement concretely in the application, scene by scene and system by system, rather than asserting a blanket claim over the entire build.
B. What the Copyright Office requires applicants to disclose and disclaim
Since March 2023 the Office has required applicants to be candid about machine-generated content. The Copyright Office guidance on works containing AI-generated material, published in the Federal Register, directs applicants to disclose the inclusion of AI-generated content in the work submitted for registration and to brief the Office on the human author’s contributions, excluding AI-generated content that is more than minimal from the claim. In practice the applicant describes what the human did, identifies the machine-generated portions being disclaimed, and claims only the human-authored authorship, the selection or arrangement the human controlled, or both. An application that presents an AI-assisted game as though every asset were fully human-authored invites refusal or later cancellation, while a careful application that claims the code, the edited art, and the curated arrangement can survive even when some generated filler is disclaimed.
The guidance also explains how examiners sort contributions. The Office states that it will consider “whether the AI contributions are the result of “mechanical reproduction” or instead of an author’s “own original mental conception, to which [the author] gave visible form.”” A prompt that produces a usable result unchanged looks like the former. Iterative work in which a designer generates options, rejects most of them, repaints portions, stitches pieces into a level, and tunes the result into the game’s systems looks like the latter, at least for the parts the human shaped. The January 2025 Copyright Office report on copyrightability drives the same point home for the most common studio habit: “The Office concludes that, given current generally available technology, prompts alone do not provide sufficient human control to make users of an AI system the authors of the output.” Typing a detailed prompt, or picking the best of fifty outputs, is therefore not authorship by itself. The authorship lies in what the human added, changed, selected, and arranged afterward, and the registration should describe exactly that.
Studios should build that record during production, not after release. Keep version histories that show human edits to generated drafts, retain the prompts together with the rejection and revision notes, log which team member approved each asset and what changed at approval, and preserve the project files that prove selection and arrangement rather than mere generation. None of this paperwork creates authorship where none exists, and no checklist guarantees the Office will register a particular claim, but without it the applicant cannot answer the examiner’s first question. One related timing point deserves emphasis because it surprises first-time registrants: registration is the gateway to federal court, since the Office examines the deposit to confirm that “the material deposited constitutes copyrightable subject matter” before issuing a certificate, as section 410(a) of the Copyright Act, 17 U.S.C. section 410(a), requires. File the application on the human-authored version with accurate disclosures, and do it before a dispute forces rushed corrections.
Applicants sometimes ask whether they can simplify all of this by registering early drafts or by filing one application that silently covers generated material. The guidance points the other way. The Office states the governing principle plainly: in its view, copyright can protect only material that is the product of human creativity, and the application must reflect that boundary through disclosure and exclusion of more-than-minimal generated content. A registration obtained by omitting the AI contribution does not become stronger with age; it becomes a misstatement the studio must correct or defend if the work is ever litigated, licensed, or sold. The cheaper course is to file accurately on the human-authored version, disclaim what the tool produced untouched, and refile or supplement as later versions add new human authorship. Studios that iterate weekly should therefore think of registration as a recurring production task tied to milestones, not a one-time filing at launch.
II. What studios, contractors, and buyers should check before money changes hands
Copyrightability of the output is only half the exposure. The other half sits on the input side: the data, art, music, text, and code on which AI tools were trained, and the license terms of the tools themselves. A studio can hold a valid registration in its human-authored game and still face claims if the pipeline behind it copied protected material without permission. The checks below follow the order a buyer, publisher, or platform will use in diligence.
A. Training data can infringe even when the finished game looks original
Earlier in the same dispute, in Thomson Reuters v. Ross Intelligence Inc., No. 1:20-cv-613-SB (D. Del. May 23, 2025), the United States District Court for the District of Delaware certified two controlling questions for interlocutory appeal: whether the headnotes and the Key Number System are original as a matter of law, and whether training on the headnotes was fair use. The certification order is a warning for studios because even the trial judge called fair use a hard question and sent it to the court of appeals rather than treating training as settled law.
In Thomson Reuters v. Ross Intelligence Inc., No. 25-2153 (3d Cir. Sept. 29, 2026), decided September 29, 2026, by the United States Court of Appeals for the Third Circuit, the court affirmed a ruling that copying 2,243 Westlaw headnotes to train a competing legal-research AI infringed copyright and was not fair use, adding that “each individual headnote is a copyrightable work.” Before training or fine-tuning a model on scraped text, a studio should confirm it has a license or a litigated fair-use basis for that corpus, because freely available source material did not save the defendant here. The underlying judicial opinions were public, yet the court explained that “Copying the headnotes was also not necessary to train ROSS’s AI, as the underlying judicial opinions were freely available,” and that taking more than necessary for a minimally transformative purpose weighed against fair use. Game assets differ from headnotes, but the method of analysis travels: courts ask what was copied, whether the use was transformative, how much was taken relative to the purpose, and how the use affects the market for the originals.
In American Society for Testing & Materials v. UpCodes Inc., No. 24-2965 (3d Cir. Apr. 7, 2026), decided April 7, 2026, the U.S. Court of Appeals for the Third Circuit affirmed the denial of a preliminary injunction against an online research platform that posted technical standards referenced in adopted building codes, agreeing that the platform was likely to succeed on its fair use defense. The comparison studios should keep is purpose plus market effect: informing the public about binding legal obligations pointed toward fair use, while copying headnotes to train a competing product pointed away from it.
Apply that method to a game pipeline. A model trained on scraped concept art, textures, music loops, or code raises the same four questions, and the studio using the model inherits the risk even when it did not build the model. The diligence checks follow directly. Ask vendors which datasets trained each tool in the pipeline and demand the answer in writing, with license schedules attached rather than marketing assurances. Read the tool’s terms for training-data representations, for who owns outputs, and for whether the vendor indemnifies customers against infringement claims, because many consumer-grade AI terms disclaim exactly that protection. Segregate high-risk inputs: placeholder art generated from unknown corpora should never ship in a release candidate without a license review, and contractor agreements should state who warrants the provenance of delivered assets and who pays if a claim arrives. Where the studio trains or fine-tunes its own models, keep the corpus inventory, the license for each source, and the filtering steps, since that file is the fair-use and license story told with documents instead of adjectives.
Registration strategy should track the same discipline. Register the human-authored game build as released, with AI-generated filler properly excluded from the claim as the guidance requires, and consider separate applications for major human-authored components, such as the original score or the script, where the authorship is cleanest. Keep each registration truthful about its version: a certificate that claims machine-generated material as human-authored is worse than no certificate, because it hands an adversary a cancellation argument and undermines the studio’s credibility in the same proceeding where it needs to enforce its rights. And remember that registration examines copyrightable subject matter, not training-data hygiene, so a certificate never answers the input-side questions above. Both files must be clean at once.
The opinion also walks through the remaining fair-use factors in terms studios should study. The fourth factor examines the effect of the use upon the potential market for or value of the copyrighted work, including markets for derivative uses, and the court found that a competing AI product trained on the headnotes threatened both the original market and the licensing market for that content. For games, the parallel is direct: training material scraped from working artists and composers has a licensing market, and using it without payment to build a competing generator is the kind of market harm the fourth factor captures. The opinion further weighed whether the copying produced countervailing public benefits and found the purpose minimally transformative at best, a finding that should worry any studio relying on a blanket assumption that training is always transformative. Transformative purpose must be shown for the specific use, corpus by corpus, not asserted for machine learning in general.
B. checks for creators, sellers, and purchasers of AI-assisted content
Creators should start with a simple authorship map before the game ships. List every asset class, including code, models, textures, animation, music, sound effects, dialogue, quests, and marketing art, and record for each whether it is fully human-made, human-edited from a generated draft, human-selected and arranged from generated options, or generated and used as is. Claim only the first three categories, describe the human contribution for each in the application, and exclude the fourth from the claim. Store the map with the project so that a publisher, platform, or buyer asking for chain-of-title evidence gets a document instead of a recollection. Update the map with each major patch or downloadable release, because new AI-assisted content added after registration is not covered by the earlier certificate’s claim.
Sellers and licensors should align their contracts with that map. Representations about originality and non-infringement should track what the authorship map actually shows, carving out disclosed AI-assisted components where the counterparty requires it, and indemnity provisions should specify who bears training-data risk for each tool in the chain rather than assigning it by silence. Contractors delivering AI-assisted assets should warrant provenance in writing, identify the tools used, and deliver the prompt-and-edit records with the files, so the studio can defend both the registration and the license chain. Purchasers and publishers should ask for those records in diligence, confirm that the tool licenses permit commercial game use and transfer of outputs, and verify that the registration certificates on the table match the builds being acquired. None of these clauses replaces legal advice on a specific deal, but a file with no authorship map, no corpus inventory, and no tool licenses is a negotiating signal in itself.
Finally, keep private platform and award rules in their own box. A festival jury that declines to consider AI-assisted entries, a storefront that requires AI-content labels, or a publisher that refuses generated assets is enforcing a private policy, not applying the Copyright Act, and those policies can be stricter than the law or more lenient than it. Read them early, comply where the studio wants the distribution or the prize, and never cite them as statements of federal law. The statute, the decided cases, and the Copyright Office guidance above are the law. Everything else is someone’s terms of participation, and confusing the two categories is how studios lose both an award argument and a registration in the same month.
Deal timelines should reflect this dual-track reality from the first term sheet. Diligence requests should ask for the authorship map and the corpus inventory alongside the ordinary chain-of-title documents, and purchase prices or royalty advances can be staged so that a portion is held back until the registration certificates and license schedules check out. Where a contractor or vendor cannot produce provenance records, the agreement should say so explicitly and allocate the resulting risk by name instead of burying it in a generic representation. Platforms and publishers increasingly ask their own AI-content questions at submission, and answering them from the same file used for registration keeps every statement consistent. Consistency across the Copyright Office application, the platform questionnaire, and the acquisition representations is itself an asset: three matching accurate statements are far easier to defend than three optimistic ones.
Keep the whole file for as long as the game earns money. Authorship questions surface years after launch, when a sequel, a film option, or an acquisition puts the chain of title under a microscope, and by then the contractors have moved on and the tool vendors have rewritten their terms twice. A studio that can still open the authorship map, the corpus inventory, the license schedules, and the registration certificates for each major version answers diligence in days instead of months, and that speed is worth real money at the negotiating table. Treat these records the way the business treats source code backups: versioned, retained, and readable without the people who made them.
Conclusion
Big Walk game AI made headlines because a successful studio admitted what many teams do quietly, and the answer United States law gives is neither a ban nor a blank check. Human-authored code, art, music, writing, and creative selection and arrangement can be registered and enforced, while purely machine-generated material cannot, prompts alone do not create authorship, and AI-generated content must be disclosed and excluded from the registration claim. Separately, training a model on copyrighted material without permission can infringe even when the finished game looks new, as the Third Circuit’s headnote decision shows. Studios that document human contributions, clean their corpora and tool licenses, register truthful version-by-version claims, and read private platform rules as contracts rather than law will hold rights they can actually sell and defend.