The short answer: AI-assisted video game cloning is not inherently illegal, but it is not automatically legal because the code is new. Under U.S. law, developers can generally reuse unprotected ideas, gameplay concepts and functional methods. Some reverse engineering also qualifies as fair use. But copying a game’s protected audiovisual expression, distributing translated proprietary code, or circumventing access controls outside an applicable exception can create separate legal problems.
The crucial distinction is between recreating what a program does and reproducing the protected material that makes a particular game recognizable. Those questions do not necessarily have the same answer.
A growing online argument claims that a developer can give an AI system a commercial game, ask it to rebuild the game in another language, compare the outputs and legally release the result because no original source code was copied. It is a persuasive technical story with an incomplete legal conclusion. Changing the language does not, by itself, change the rights in what was reproduced.
The question is becoming urgent. On October 6, 2026, Tom’s Hardware reported on an unofficial, reportedly AI-assisted Grand Theft Auto V browser port that became inaccessible shortly after appearing. The reporting did not establish who caused the site to go offline or whether a copyright owner had issued a takedown demand. Nor did it independently audit what the port contained. Its significance is the question it raises, not a proven legal outcome.
A day later, Google launched Playground, an experimental platform for creating and sharing games from written prompts, while Google and Unity announced plans for Unity Spark. Making an original game through prompts, however, is different from reconstructing an existing copyrighted game. Conflating the two is where much of the confusion begins.
What Does It Mean to Decompile or Clone a Video Game With AI?
Compilation normally transforms human-written source code into machine instructions that a computer can execute. Decompilation attempts to recover a readable approximation of that source from a compiled program. Reverse engineering is the broader process of examining a program to understand its behavior or design; it does not always require decompiling code.
AI can help analyze disassembly, propose functions, translate instructions, generate replacement code, inspect compiler errors and compare test results. But "AI decompiled a game" is not a precise description of one technique.
| Method | What the developer produces | Why the legal question differs |
|---|---|---|
| Matching decompilation | Source intended to compile into the same machine instructions as the original | Studying a program and publicly distributing reconstructed source are distinct activities. |
| Functional reconstruction | New code intended to produce comparable behavior | Recreating unprotected functionality may be permissible; copying protected characters, art, dialogue or audiovisual expression may not be. |
| Static recompilation | A translation of an existing executable’s instructions for execution on another platform | A translated version may still reproduce protected program expression, even if its files or instruction format differ. |
| Engine reimplementation | A separately developed runtime that supports an existing game’s data formats and behavior | Independent engine code may present a stronger position, but the rights in game assets, access controls and any borrowed code still matter. |
| Porting recovered source | Adapted code intended to run on different hardware or operating systems | Portability does not itself confer a right to adapt or distribute another party’s protected program. |
These methods can overlap. A game may use a rebuilt engine, game data supplied by the player, an emulator, and individually reconstructed functions. No method is automatically lawful or unlawful under every set of facts.
For a technical overview of the methods and documented projects, see Curtis Pyke’s October 3 analysis at Kingy AI. The more consequential question is what happens when an AI-generated recreation leaves the development environment and reaches other players.
Is Game Decompilation Illegal Under U.S. Copyright Law?
No. There is no blanket federal rule making all decompilation illegal. The legality depends on the rights involved, what is copied, why it is copied, how access is obtained, what is distributed and whether an exception or defense applies.
Two provisions of the Copyright Act explain the starting point:
- 17 U.S.C. Section 102(b) says copyright does not extend to ideas, procedures, processes, systems or methods of operation.
- 17 U.S.C. Section 106 gives copyright owners exclusive rights, subject to limitations, including reproduction, preparation of derivative works and distribution.
A game’s underlying rules and functional concepts can be unprotected while its actual code, artwork, music, dialogue, character designs and protectable audiovisual presentation remain protected. Copying a method is not necessarily copying its expression. Yet reproducing the visual form of a game cannot always be excused by calling every creative decision a gameplay rule.
The law also recognizes fair use under Section 107. That is a contextual inquiry into purpose and character, the nature of the copyrighted work, the amount used and effects on the market. Noncommercial activity is not automatically fair use, and using AI is not a separate fair-use exception.
Sega v. Accolade: Reverse engineering can be fair use
In Sega Enterprises v. Accolade (1992), the Ninth Circuit considered whether Accolade could disassemble Sega software to discover the unprotected functional requirements needed to make independently developed games work on the Genesis console.
The court recognized fair use where disassembly was necessary to access those unprotected elements and the developer had a legitimate reason to do so. The opinion directly contradicts the blanket assertion that decompiling a commercial game is categorically illegal.
But it did not announce that anyone may copy and redistribute an entire game after inspecting it. Compatibility research and releasing a competing copy of protected game content are different questions.
Sony v. Connectix: Even intermediate copying may be protected
In Sony Computer Entertainment v. Connectix (2000), the Ninth Circuit addressed an emulator that allowed PlayStation games to run on another platform. Connectix had made copies of Sony’s BIOS during development, but wrote its own replacement software.
The appellate court held that the intermediate copying at issue was fair use. An important part of the analysis was that the finished emulator did not contain infringing Sony material.
The case is especially relevant to AI workflows: an engineer may need to examine protected material to learn how a system behaves, without thereby receiving permission to ship that material in the finished product.
Google v. Oracle: Software interfaces are a special context, not a license to copy games
The Supreme Court reached another important but narrow result in Google LLC v. Oracle America (2021). Assuming for the purpose of its analysis that the relevant Java API declaring code was copyrightable, it held Google’s particular copying of that code to build a different platform was fair use.
The Court focused on the functional character of the interface, the amount and purpose of copying, and the new platform being created. It did not hold that all software interfaces are uncopyrightable, or that a game can be replicated in full because software is functional.
The Pixel-Matching Paradox: Technically Accurate Can Still Mean Legally Risky
One of the strongest claims behind AI-assisted cloning is that the developer can check whether a replacement game works by comparing what appears on the screen. If the same input produces the same output, the AI can keep refining its code until it gets closer.
As an engineering technique, that is sensible. A comparison can expose broken physics, missing objects, different animation timing or incorrect rendering. But matching a game’s visible output is not a legal safe harbor. The output may itself contain copyrighted audiovisual expression.
The most revealing case is Tetris Holding v. Xio Interactive (2012).
Xio developed Mino, a puzzle game whose creators argued that they had copied only Tetris’s uncopyrightable rules and functionality. The court agreed that the underlying idea of falling, rotating pieces to complete rows was not protected as such. But it found that the competing game had copied an extensive combination of protected visual choices, including the presentation of pieces and other features of the playing field. It granted summary judgment on copyright infringement.
The ruling did not say that a 10-by-20 grid, a ghost piece or every isolated design choice is automatically protected in all games. The court assessed the particular elements and their combination, emphasizing the wholesale similarity of the visual expression.
A useful contrast is Data East v. Epyx (1988). The Ninth Circuit rejected an infringement finding involving karate games where many similarities followed from the sport itself and the technical limitations of the era. Comparable gameplay and visual elements do not always equal unlawful copying.
Together, these decisions show why neither extreme is correct: similar gameplay is not automatically infringement, and independently written code is not automatically a defense.
Consider two hypothetical AI projects:
- One observes the rules of a platform game and creates new characters, artwork, level designs, sound, dialogue and source code. It may share unprotected mechanics with the original without copying its protected presentation.
- Another produces new source code but recreates the original characters, distinctive environments, animations, sounds and screen composition almost exactly. The absence of source-code copying would not eliminate the audiovisual infringement question.
Pixel similarity can be evidence of how closely two games look. It is not, by itself, the complete legal test. Courts must identify which similarities involve protected expression, which are dictated by function and how those features relate in the works as a whole.
This distinction also explains why an emulator displaying a player’s legitimately obtained game content presents a different fact pattern from a developer bundling someone else’s sprites, music and environments into a freely distributed clone.
Does Rewriting a Game in Another Programming Language Make It Legal?
No. A language change does not determine whether protected expression was copied. A game translated from C++ into Rust, C#, JavaScript or WebAssembly-related code may still incorporate copyrighted program expression or constitute an unauthorized adaptation. Conversely, independently developed code that replicates unprotected functionality can be lawful even when the functional behavior is similar.
Computer Associates v. Altai (1992) shows why the distinction matters. After an earlier version of Altai’s software incorporated another company’s code, Altai undertook a controlled rewrite. The Second Circuit affirmed the finding that the rewritten version did not infringe copyright, applying an approach that separated protectable program expression from elements dictated by efficiency, external requirements or the public domain.
The case is particularly instructive because the court also allowed separate trade-secret questions to continue. A clean copyright result did not automatically resolve every legal problem arising from the original acquisition of information.
For a modern AI developer, the lesson is practical. Independent specifications, documented provenance, separate implementation work, tests and review can help establish what was created independently. But the phrase clean room is not a magic statutory exemption. A clean-room process does not legalize copied audiovisual assets, misappropriated trade secrets or unlawful circumvention.
How Much Can AI Really Reconstruct in 2026?
There is real progress, but public demonstrations often compress several different measures of success into a single percentage.
As of this article’s October 10 review, the public README for the Super Mario 64 DS decompilation project reported three separate milestones:
| Project-reported measure | Reported progress | What it actually measures |
|---|---|---|
| Matched functions | 99.9% | Functions whose reconstructed output matches the target machine code under the project’s stated checking method. |
| Converted source | 31.7% | Functions converted into the more readable C++ structure described by the maintainers. |
| Linked units | 93.5% | Matched units compiled directly into the associated PC port rather than relying on stand-ins. |
These are maintainer-reported, changing repository counters, not figures independently reproduced by sherafy.com. The matching count includes a separately disclosed category of hand-written assembly primitives. The percentages measure different denominators and should not be combined into a claim that AI autonomously recovered 99.9% of the original developer’s readable source files.
The project also credits prior community documentation, developer-written tooling, automatic templates, human contributions and AI-assisted matching. That is a more credible account of technical progress than asserting that an untrained person can rebuild any commercial game in a week or a month.
Research likewise demonstrates both improvement and limitations. The peer-reviewed 2025 DecompileBench study examined 23,400 functions from 130 programs and found that the LLM approaches it tested produced more understandable code, but substantially less functionally correct code, than the traditional tools in that comparison. That result describes the tested systems and tasks; it should not be treated as a fixed ranking of all models in late 2026.
A June 2026 AutoDecompiler preprint reports improvement from repeated compilation, execution and feedback-based refinement. It supports the importance of verification loops, though its preprint results do not establish reliable end-to-end recreation of arbitrary games.
A game that reaches its title screen may still fail when it saves, loads a later level, plays audio, encounters rare physics states or finishes its campaign. The real benchmark for a finished port includes those behaviors, not just a working demo or a near-perfect screenshot.
Game Cloning Legal-Risk Matrix: What Changes the Answer?
The following comparison is an editorial risk assessment based on the statutes and cases discussed above. It is not a legal determination about any particular project.
| Developer activity | Primary legal issue | Practical assessment |
|---|---|---|
| Build an original game using a familiar mechanic, with independently created code and art | Idea versus protected expression | Generally lower copyright risk, assuming no other protected material is copied. |
| Study a lawfully obtained game to discover compatibility requirements | Intermediate copying, fair use, contract terms | Potentially defensible, especially where analysis is necessary to access unprotected functions. Facts matter. |
| Release a new engine that reads game data supplied separately by users | Independent code versus protected assets | Often a stronger model, but not an automatic license for all aspects of the project. |
| Recreate an existing game’s characters, music, level art and audiovisual presentation using new code | Substantial similarity and derivative works | Significant infringement risk where protectable elements are copied. |
| Publish code produced by translating a proprietary executable instruction by instruction | Reproduction and adaptation of program expression | Material risk; changing the representation does not establish independence. |
| Circumvent DRM or encryption to obtain the information used in the project | DMCA Section 1201 | Separate legal review required, even where the intended final software is noninfringing. |
| Obtain an unreleased game or confidential source through unauthorized access | Access laws, confidentiality, trade secrets, copyright | High independent exposure; this is not the same issue as examining software lawfully acquired. |
The deciding evidence is usually more specific than a project’s label. Useful questions include: What copy did the developer acquire? Was access authorized? What was reproduced during testing? What files are being distributed? Does the finished product contain original expression or protected game material? Is an access-control measure being bypassed?
What About DRM, Owning a Copy and Game Preservation?
A lawful copy is relevant, but owning or licensing a game does not give the purchaser unrestricted permission to redistribute its code and assets. Different activities can implicate copyright, licenses, technological access controls, trade secrets and, in some circumstances, other computer-access laws.
17 U.S.C. Section 1201 generally prohibits certain acts of circumventing technological measures that control access to copyrighted works. Its Section 1201(f) interoperability exception contains important conditions: among other things, the person must have lawfully obtained the right to use a copy, the analysis must be directed at necessary interoperability elements not readily available, and the relevant acts must not constitute copyright infringement. It is not a general exemption for cracking commercial games.
There are also narrow video game preservation exemptions in 37 C.F.R. Section 201.40(b)(19). Some address restoration of local gameplay after authentication servers are discontinued; others permit qualifying libraries, archives and museums to preserve certain games under specified conditions. They do not grant the public an unrestricted right to distribute cracked copies or host playable commercial games online. The Copyright Office’s 2024 Section 1201 proceeding explains the rulemaking behind these exemptions.
A practical example of a different distribution architecture is OpenMW, an independently developed engine for The Elder Scrolls III: Morrowind. Its developers state that the engine does not use the original executable and does not ship Bethesda’s game data. Users who want to play Morrowind supply the original game content themselves. That distinction is legally meaningful, though OpenMW’s existence is not a court ruling that every independent engine project is lawful.
This is also where the distinction between reconstruction and an actual intrusion matters. Our CyberLeek GTA 6 leak investigation separates the unauthorized circulation of nonpublic game material from the unverified method by which that material was originally acquired. For a documented incident involving unauthorized system access by AI agents, see our analysis of the OpenAI-Hugging Face lawsuit. Neither factual situation should be casually relabeled as ordinary reverse engineering.
Does Using AI Change Who Owns the Code or Remove Copyright Liability?
Two separate questions are often confused: Can the AI-assisted creator own copyright in the output? And can that output infringe another creator’s copyright? A negative answer to the first does not guarantee a negative answer to the second.
In its 2025 report on AI-generated works, the U.S. Copyright Office concluded that copyright protection depends on sufficient human authorship of expressive elements. AI assistance does not automatically disqualify a work, but purely machine-generated material and prompting alone may not establish the human creative contribution necessary for protection. The analysis is fact-specific.
That means a person could face a copyright claim for reproducing protected elements of an existing game while also being unable to claim copyright in some of their own purely AI-generated material. AI can reduce production effort; it does not create an exemption from another owner’s rights.
The September 2026 Ninth Circuit decision in Doe v. GitHub also illustrates the importance of identifying the precise legal theory. The appellate court rejected a particular claim about removal of copyright-management information from AI output. It did not decide that AI-generated software can never infringe copyright, or that every act of AI code reuse is permissible. We examined the procedural boundaries in our separate Doe v. GitHub briefing.
For developers, a responsible approach is to track source provenance, license obligations, any supplied third-party files and the specific material reproduced in the output. A model’s assurance that its code is "original" cannot substitute for those checks.
Will AI Game Reconstruction Destroy the Software Business?
The strongest economic claim circulating around these demonstrations is that AI will soon make established games and software products so easy to duplicate that conventional software companies lose their value.
The verified evidence supports improving reconstruction capability. It does not establish that broad software-market collapse is imminent. The distinction matters because "can generate a functional substitute" and "can legally, reliably and commercially replace the incumbent" are different propositions.
AI may lower the cost of prototyping, compatibility research, porting, testing and parts of code recovery. It may make it easier for small studios to implement familiar features. It could also reduce the value of software products whose main advantage is difficult-but-reproducible functionality. These are plausible economic consequences, not measured outcomes established by a handful of game demos.
A commercial video game is more than a working executable. It may depend on protected characters and creative assets, licensed music, online services, distribution agreements, security, ongoing support, multiplayer communities and brand recognition. Outside games, many software businesses likewise sell services, integrations, hosted infrastructure, compliance and support in addition to code.
Nor does a successful reconstruction establish the true labor cost of the project. Repositories can inherit years of community documentation, existing tools, specialist expertise and previous work. Public AI usage costs rarely include every failed attempt, debugging session, legal review and release test.
The more testable question is which parts of software production become cheaper and which forms of differentiation survive. That is a serious subject for economic analysis. It is not answered by declaring that the AI bubble will survive while every other software business collapses.
Frequently Asked Questions
Can you legally copy video game mechanics?
Generally, abstract gameplay rules and methods of operation are not protected by U.S. copyright. But specific code, artistic presentation and creative audiovisual choices may be protected. The Tetris and Data East cases illustrate why copying a mechanic and copying a game’s distinctive expression are not the same thing.
Is it legal to decompile a game you bought?
Buying or lawfully accessing a copy does not make every form of decompilation lawful or unlawful. Under Sega and Connectix, some intermediate copying and disassembly for a legitimate purpose may qualify as fair use. Licenses, access controls, purpose and any eventual distribution must be examined separately.
Is an emulator legal if it runs copyrighted games?
An independently developed emulator is not automatically infringing merely because it can run commercial games. Sony v. Connectix protected particular intermediate copying used to create a noninfringing emulator. Where BIOS code, game ROMs, DRM circumvention or copyrighted assets are obtained or distributed, additional legal questions arise.
Can you distribute an AI-generated remake if you remove the original assets?
Removing music, models and textures can reduce some risks, but does not establish that the distributed code or remaining presentation is noninfringing. A project may still contain translated proprietary instructions, protected narrative elements or other expressive material. Its access method and license obligations matter too.
Is it legal to use AI to make a game that looks almost identical to another game?
Not automatically. If the similarity involves substantial protected audiovisual expression, new source code may not prevent infringement. If the overlap is limited to unprotected rules, ordinary genre conventions or functional requirements, the analysis can favor independent development. The comparison is fact-specific.
Intelligence Assessment: What Is Established, What Is Conditional, and What Remains Unknown
Verified: U.S. law does not categorically ban game decompilation. Federal appellate decisions recognize fair use for certain forms of reverse engineering and intermediate copying. Other decisions confirm that game visuals and some nonliteral software expression can be protected even when the copying party uses different source code.
Verified: Public AI-assisted reconstruction projects now report extensive function matching and working ports. Their technical milestones represent specific tested achievements, not proof that any commercial game can be rebuilt autonomously or that publishing the result is lawful.
Conditional: Independently recreating functionality, building a clean-room engine or distributing software that requires users to provide their own game assets may be legally more defensible. The outcome depends on what was actually copied, acquired, circumvented and distributed.
Not established: That a prompt to rewrite a game in another language immunizes the result from copyright liability; that pixel matching proves legal independence; that a one-week reconstruction is typical; or that AI decompilation will necessarily destroy the software industry.
The practical conclusion is straightforward: AI changes the cost of reconstructing software much faster than it changes the legal distinction between unprotected functionality and protected expression. Anyone evaluating an AI game clone should examine the source of the material, the development method and the finished product separately. Technical equivalence answers whether a recreation works. It does not answer whether it can legally be distributed.
References and Further Reading
U.S. Statutes, Regulations and Government Guidance
- 17 U.S.C. Section 102 – Subject Matter, Section 106 – Exclusive Rights, and Section 107 – Fair Use – The statutory foundation for separating functional ideas from protected expression, identifying copyright-owner rights and evaluating applicable fair use.
- 17 U.S.C. Section 1201 – Circumvention of Copyright Protection Systems – Contains access-control rules and the conditional reverse-engineering interoperability exception.
- 37 C.F.R. Section 201.40 – Circumvention Exemptions – Sets out the current regulatory exemptions, including specific video game preservation activities in subsection (b)(19).
- U.S. Copyright Office – 2024 Section 1201 Triennial Proceeding – Primary rulemaking record explaining the limited nature of the exemptions.
- U.S. Copyright Office – Copyright and Artificial Intelligence Reports – Government analysis of copyrightability and human authorship in AI-assisted output; not an opinion legalizing reproduction of existing games.
Court Opinions
- Sega Enterprises, Ltd. v. Accolade, Inc., 977 F.2d 1510 (9th Cir. 1992) – Recognizes fair use for necessary reverse engineering to examine unprotected functional elements in the circumstances before the court.
- Sony Computer Entertainment, Inc. v. Connectix Corp., 203 F.3d 596 (9th Cir. 2000) – Holds particular intermediate BIOS copying during development of a noninfringing emulator to be fair use.
- Tetris Holding, LLC v. Xio Interactive, Inc., 863 F. Supp. 2d 394 (D.N.J. 2012) – Shows why independently created game code can still infringe protected audiovisual expression.
- Data East USA, Inc. v. Epyx, Inc., 862 F.2d 204 (9th Cir. 1988) – Explains why similarities dictated by genre, realistic subject matter or technical constraints may be unprotected.
- Computer Associates International, Inc. v. Altai, Inc., 982 F.2d 693 (2d Cir. 1992) – Establishes the abstraction-filtration-comparison approach for nonliteral software expression and distinguishes copyright from trade-secret claims.
- Google LLC v. Oracle America, Inc., 593 U.S. 1 (2021) – Finds fair use on specific Java API facts without granting a general right to copy software.
- Doe v. GitHub, Inc., No. 24-7700 (9th Cir. Sept. 16, 2026) – Addresses an output-stage DMCA copyright-management-information theory; does not resolve ordinary copyright infringement by generated code.
Technical Research, Primary Project Documentation and Current Context
- Tangosdev – Super Mario 64 DS Decompilation README – Maintainer-reported function-matching, readability and PC-port integration metrics; progress is subject to change.
- Gao et al. – DecompileBench, Findings of ACL 2025 – Peer-reviewed comparison of LLM-assisted and traditional decompilation on functional correctness and readability.
- Liu et al. – Binary Decompilation LLM with Feedback-Driven Multi-Turn Refinement (2026) – Preprint examining iterative execution-feedback methods; findings require interpretation within their tested conditions.
- OpenMW – Official FAQ – Documents an independently written engine and the requirement to provide original Morrowind game content.
- Curtis Pyke – AI Video Game Decompilation: What Works, What’s Possible, and What’s Legal – October 2026 technical synthesis that helped bring attention to the distinction between different reconstruction methods.
- Tom’s Hardware – Unofficial AI-Assisted GTA V Browser Port Report, October 6, 2026 – Contemporary reporting on a port that quickly became inaccessible, with its takedown cause unverified.
- Google – Introducing Playground, October 7, 2026 – Primary announcement of prompt-based original-game creation, a separate activity from decompiling a copyrighted title.
Related sherafy.com Investigations
- Doe v. GitHub Explained: What the Ninth Circuit Actually Ruled About AI Code and Attribution – A fuller examination of a recent AI coding and copyright-management-information ruling.
- CyberLeek GTA 6 Leak: New Video, Crypto Money Trail and Rockstar’s Investigation – Distinguishes nonpublic game disclosures and their unresolved acquisition method from legitimate reverse-engineering questions.
- OpenAI Was Sued Over the Hugging Face Hack. What Does California Law Actually Make It Prove? – Examines unauthorized AI-agent access and legal attribution, a different legal category from copyright compatibility research.
Editorial currency note (October 10, 2026): This article addresses U.S. law and publicly documented technology as of the review date. Fair-use questions are fact-specific; contractual rules, exemptions, pending litigation and technical project counters can change. The examples above are not legal advice or determinations about specific projects. Recheck the current statutes, regulations, repository metrics and case history before relying on them for a particular release.
