Game Contract Dispute Highlights Risks of AI-Generated Code

A dispute described by game-industry IP lawyer Haley MacLean offers a compact warning for studios, contractors and publishers weighing generative AI in development: delivering something that was produced with AI is not the same as delivering a working, supportable milestone.

MacLean, the video game practice head at Voyer Law, outlined an anonymized matter involving a programming contractor whose submitted milestone was reportedly broken and riddled with problems. After revisions were requested before acceptance, the contractor allegedly said they could not make them and sought full payment. MacLean said it emerged that the work was AI-generated and that the contractor did not understand what had been made. The result, she said, was no payment and lost time for the client.

“Just because you can order at a restaurant doesn’t mean you’re a chef,” MacLean said while describing the incident.

The sharpest issue in the account is not simply whether a generative tool was used. It is whether the contractor could stand behind the deliverable: explain its operation, repair defects, revise it to match requirements and safely work with the project materials needed to produce it. In game production, those capabilities are part of the practical value of programming work, especially when code must fit into a larger project.

What a milestone is supposed to establish

A milestone is a defined piece of work submitted at a particular stage of a contract. For programming, that might mean a feature, system or implementation that meets an agreed specification. The source material does not identify the project, the programming task or the contract language in this case, so it would be wrong to presume exactly what the contractor promised. But MacLean’s account makes clear that acceptance was not automatic: the client identified problems and requested revisions before agreeing the work was acceptable.

That distinction matters. A visible result can look like progress while still failing as a deliverable if it cannot meet the agreed requirements or be maintained. Code can compile in one circumstance yet break when connected to other systems. It can also fail to handle expected inputs, conflict with existing dependencies, or become difficult to change because nobody responsible can trace its logic.

Generative AI can produce code from prompts, often by assembling likely responses based on patterns in its training and the information supplied by a user. The phrase “vibe coding” has increasingly been used for a prompt-led approach where a person directs an AI tool without necessarily understanding the programming output in detail. The label is loose, and it should not be treated as a technical certification or a universal description of AI-assisted work. Still, MacLean’s example captures the risk at its most basic: a contractor who cannot diagnose or revise an output may be unable to complete the job they were hired to do.

Generated output is not a substitute for accountability

It is possible to distinguish between use of a tool and abdication of responsibility. A developer could use a tool in a workflow and still review the result, understand its behavior, test it against the project’s needs and maintain it afterward. Conversely, the reported dispute involved a contractor who, MacLean said, could not fix the material after defects were raised.

That difference is crucial because software is not a one-time artifact. A game is typically a connected collection of code, assets, tools and pipelines. A new system has to coexist with the rest of that work. When it fails, someone needs to determine what went wrong, make a change without creating a new issue, and verify the fix. This is the work commonly meant by debugging: identifying the source of a fault and correcting it.

In practical terms, a team evaluating outside programming work needs more than a screenshot, a brief demonstration or a pile of files. It needs confidence that the person delivering the work can answer basic questions about how it functions and can deal with issues discovered during review. AI does not remove those obligations merely because it made producing a first draft faster.

The dispute also points to a scheduling cost that can be easy to overlook. A defective milestone does not just delay that one contractor’s task. Internal staff may have to investigate the output, explain what is wrong, request changes, consider replacement work and re-plan dependent tasks. Even when a client withholds payment for unacceptable work, the time spent evaluating and untangling it is not automatically recovered.

The confidentiality issue may be separate from the code quality issue

MacLean added a second allegation with potentially broader implications: she said the contractor had put confidential game information into an AI tool in order to create the milestone. In her description, that was a breach of confidentiality.

Confidential information in a game project can be the non-public material a contractor receives to do their work. The supplied account does not specify what information was entered or the terms of the relevant agreement. That lack of detail matters. It means outsiders cannot determine the precise scope of the alleged breach or what contractual remedies might apply.

Even so, the underlying operational lesson is plain. Development materials should not be treated as casual prompt text. Teams must know which tools a contractor plans to use and what information may be submitted to those tools. A prompt can be more revealing than it initially appears if it contains details about an unreleased project, its systems, assets, narrative, build problems or technical architecture.

This makes AI policy a production and contracting issue, not merely a debate about creative preference. If a company needs project details kept within a defined circle, it needs expectations that address third-party tools and outsourced work. The relevant question is not only, “Can this task be automated?” It is also, “What project information would be exposed in doing so, and is that permitted?”

Why anti-AI terms are becoming routine

MacLean has said she is seeing clients increasingly include anti-AI language in agreements. She characterized such clauses as close to boilerplate in this area and recommended pushing for protections where a publisher is also creating assets, including in marketing, porting and quality assurance.

Boilerplate refers to standard contract language repeatedly used across agreements. Standard wording does not make a term unimportant; it often signals that a recurring risk has become important enough to address expressly rather than leave to assumption. The report does not provide example clause text, and no single clause can answer every production question. A contract still needs clarity about what work is required, how acceptance works, how revisions are handled and what confidential material can be used where.

But an explicit generative-AI provision can force a conversation that otherwise may not happen until a delivery goes wrong. Is AI use prohibited, allowed with approval, or limited to certain tasks? Does a contractor need to disclose it? Can non-public materials be entered into any external system? Who is expected to review and maintain generated output? Those are operational questions as much as legal ones.

They also matter across roles. The source material specifically raises marketing, porting and QA alongside programming. Porting generally means adapting a game for another platform. QA, or quality assurance, is the process of finding, documenting and helping verify fixes for defects. Each discipline can involve confidential builds or materials, and each can create downstream problems if work is treated as disposable output rather than something a responsible person can validate.

A live argument inside game development

The incident arrives amid a wider and unsettled argument about generative AI in games. MacLean has previously described concerns that include copyright risk, reputational damage and uncertainty around rights ownership. At the same time, high-profile figures in the industry have expressed support for the technology. Valve co-founder Gabe Newell said in 2025 that AI could create a “funny situation” in which less-experienced programmers who use the tools effectively could be more productive in value creation.

Those positions are not necessarily about the same thing. A claim that a tool may make some people more productive is different from a claim that any generated output is good enough to ship, that its user understands it, or that sensitive game information can be supplied to it without restriction. Productivity and accountability can coexist only when teams retain the ability to review, test, repair and take responsibility for the result.

The terminology itself has become part of the friction. Epic has said it “vibe coded” a new Epic Games Store launcher and briefly made generative AI a requirement for an Unreal Engine game jam. Meanwhile, the leaders of the RPCS3 PlayStation 3 emulator have criticized AI-reliant vibe coders and urged them to learn debugging and coding. The disagreement is therefore not limited to whether AI belongs in a workflow; it extends to what competence should mean when tools can generate a substantial amount of material.

Players are encountering the consumer-facing side of that debate, too, amid complaints about AI-generated material on storefronts such as Steam and in open-source communities working on ports and recompilations. That does not make every use of automation identical, and the term “AI slop” is a criticism rather than a neutral technical category. It does show why studios have a reputational stake in understanding what was made for them and how it was made.

For development teams, the practical standard suggested by MacLean’s account is refreshingly unglamorous: a contractor should be able to deliver working work, address requested revisions and protect the confidential information entrusted to them. The method used to get to a first draft cannot excuse a failure on those fundamentals.

And for players watching the wider AI discussion, the important point is that the debate is not only philosophical. It reaches into the reliability of features, the security of unreleased projects, the accountability of vendors and the amount of cleanup left for the people who must turn a rough output into a functioning game. For a separate look at how deeply players can engage with a game’s systems once the work is actually in their hands, see this Cyberpunk 2077 build guide.