What the video game lawyer got wrong about AI

His IP warning targets 19% of how game studios actually use these tools

May 07, 2026

Nick Allan’s London Games Festival warning to swerve generative AI targets a use case that is roughly one-fifth of how the industry actually uses it
GDC 2026 State of the Game Industry data shows 81% of devs using AI deploy it for research and brainstorming, only 19% for asset generation, only 5% for player-facing features
Allan’s claim of “no IP ownership” in AI-generated work is overstated even in his home jurisdiction, where Section 9(3) of the CDPA grants 50-year copyright to computer-generated works
The legal defensibility argument against vibe-coding ignores that game cloning lawsuits do not turn on source code copyright

Nick Allan, partner and head of video games at Mishcon de Reya, told an audience at London Games Festival to avoid using generative AI to produce key assets. Game Developer’s Chris Kerr summarised the tenor of the talk as advice to swerve genAI like the plague. The number to set against that advice is nineteen.

That is the percentage of game industry professionals using AI tools who are using them for asset generation, according to the GDC 2026 State of the Game Industry survey published in January. Eighty-one percent are using AI for research and brainstorming. Forty-seven for code assistance. Five percent for player-facing features. Allan’s warning targets the part of the surface area that is, by some distance, the smallest.

Before getting into where the framing comes apart, one part of his talk earns its place. Allan told indie devs that any IP created by people who aren’t employees, by which he means contractors, freelancers, and external developers, is probably not going to be owned by the studio. It will be owned by them. He is right, and this catches indie devs out repeatedly. The default position under both UK and US law is that contractors retain ownership of what they create unless there is a written assignment, regardless of whether the studio paid in full. Specialist game lawyers have been making this point year after year, and I have seen it bite small studios more than once. If Allan’s whole talk had been about that one issue, it would have been a more useful piece of journalism than the one Game Developer published.

But on generative AI specifically, the framing falls apart in three places.

The nineteen percent problem

The GDC survey is the closest thing the industry has to a representative snapshot. Over 2,300 professionals responded. Across the full population, thirty-six percent said they use generative AI as part of their job. That headline number is the one most coverage focused on. The interesting numbers are inside it.

Of those thirty-six percent, here is what they are doing with the tools. Eighty-one percent for research and brainstorming. Forty-seven for code assistance. Forty-seven for writing emails. Thirty-five for prototyping. Twenty-two for testing and debugging. Asset generation comes in at nineteen. Procedural generation at ten. Player-facing features at five.

Asset generation: nineteen. Procedural generation: ten. Player-facing features: five.

The IP question Allan is warning about applies most directly to the bottom three of those numbers. It applies barely at all to the top five.

There is a reason the breakdown looks like that. Generating a key character, a background environment, a piece of music, or a story bible with a prompt and shipping that artefact as the studio’s IP is the use case where the legal risk Allan flags is most relevant. Using a model to interrogate a design decision, scaffold a piece of code, write a meeting summary, or produce a test plan is a different activity. None of those outputs are claimed as the studio’s copyright in the first place. There is nothing to register, nothing to defend, nothing to lose.

Half the industry has decided generative AI is having a negative impact on game development, up from thirty percent the year before and eighteen the year before that. Seven percent think the impact is positive. Those numbers are real and worth taking seriously. But the same survey shows that the people most opposed to genAI on principle are not the people who have stopped using it. They are using it. They are using it cautiously, and for the things where they think the trade-off is worth it. Research, brainstorming, code assistance, prototyping. The legal advice to swerve the technology assumes a use pattern that the data does not support.

Two risk profiles

A lawyer’s job is to minimise litigation risk. That is what they are good at. That is what their clients pay them for. When a specialist video games partner at a Tier 1 firm tells a room of indie devs to avoid generative AI, he is making the right recommendation against the risk profile he is paid to optimise.

A studio’s job is different. Studios survive or die on shipping velocity, runway management, and parity with competitors who are using the same tools. The risk a studio is optimising against is whether it will still exist in twelve months. Litigation in three years is somewhere lower on the list.

These two risk profiles point in different directions when it comes to genAI tooling. A studio that swerves AI to preserve theoretical IP defensibility is paying a real shipping-velocity cost to mitigate a hypothetical legal cost. The first cost shows up on the burndown chart this sprint. The second cost might never arrive at all, and if it does, will arrive in a form the studio will not be in business to face.

Reporting after GDC noted that adaptability to AI tools is becoming a hiring prerequisite, with job listings in 2026 considerably more likely to mention AI proficiency than they were a year earlier. Thirty-five percent of studios are now relying primarily on self-funding. The market is tighter, the runway is shorter, and the studios still standing have the most reason to take the productivity gains AI tooling actually delivers, rather than worry about hypothetical IP exposure that may never crystallise.

This is what I mean when I say legal advice and studio survival are pointing in different directions. They address different risks. They cannot be optimised together by the same heuristic. A studio that takes Allan’s advice without filtering it for relevance is solving the wrong problem.

Allan said that “most jurisdictions will just say there’s no IP ownership” in AI-generated work. Kerr did not push back on this in the article. It is worth pushing back on. Allan and his audience were both based in the UK, and the UK is one of the few jurisdictions on earth that does grant copyright to computer-generated works without a human author.

Section 9(3) of the Copyright, Designs and Patents Act 1988 provides fifty-year copyright protection to literary, dramatic, musical or artistic works “generated by computer in circumstances such that there is no human author of the work.” Authorship is attributed to “the person by whom the arrangements necessary for the creation of the work are undertaken.” The UK government’s own consultation paper notes that for a general-purpose AI generating output from a prompt, “the author will usually be the person who inputted the prompt.” The UK is one of perhaps four major jurisdictions globally with this kind of provision, alongside Ireland, India, and Singapore. Most of the rest of the world, including the US, has no equivalent and treats fully AI-generated work as uncopyrightable.

The UK is one of the few countries on earth that does grant copyright to computer-generated works without a human author.

The provision is contested. The government consulted on it between December 2024 and February 2025 and received over eleven thousand responses. It may be repealed. But it is currently the law, and a London-based partner addressing a London audience telling them there is no IP ownership in AI-generated work is glossing over the bit of the law that contradicts him.

What cloning lawsuits actually turn on

Allan’s other genAI argument concerned source code. Using a tool like Claude to vibe-code a video game, he said, will not necessarily create an ownership issue, but if someone copies your source code, “you probably wouldn’t be able to sue them in the way that you would if it had been written by human beings.”

This is technically arguable. It is also practically irrelevant for the vast majority of game studios.

Game cloning lawsuits do not turn on source code copyright. They turn on audiovisual look-and-feel, trade dress, character likeness, and trademark. The leading authority is Tetris Holding LLC v Xio Interactive Inc (2012), where Tetris won on the protectable expression of the game’s visual elements and the trade dress of its piece colours, playfield ratio, and packaging. The court explicitly distinguished protectable expression from unprotectable rules and functionality. Xio’s source code was never the issue.

The same pattern holds across the canonical game cloning cases. Spry Fox v LOLApps (Triple Town v Yeti Town, 2012) followed Tetris reasoning before settling. Atari v Amusement World (1981). Capcom v Data East (1994). The Fortnite dance emote suits. All of them turned on expressive elements. None turned on source code.

There is a structural reason for this. Source code is rarely accessible to a cloner in the first place. Compiled game binaries do not include readable source. Cloners are looking at the running game from the outside, screenshotting it, prototyping a copy of what they can observe, and rebuilding the implementation from scratch. The thing being copied is the audiovisual surface, not the codebase. Source code copyright is the wrong tool for the job because the situation in which it applies almost never arises. A studio worried about clones is worried about the trade dress, the character art, the music, the iconic mechanics expressed as visible behaviour. None of those are at additional legal risk because the underlying code was scaffolded by Claude.

Here is the point. If your studio’s competitive moat depended on source code copyright, you would have a problem. But it does not. The moat for any small studio is execution speed, brand, and the audiovisual expression you ship around the code. The legal defensibility argument against vibe-coding solves for a category of risk that does not exist for small studios in the games space. It is solving the wrong problem twice over: once for being targeted at the wrong type of asset, and once for being relevant to a kind of dispute that does not happen.

Where the moat actually sits

What survives all this is a smaller, sharper version of Allan’s advice. The contractor IP point is right, and indie studios should fix it on Monday morning. Trademark in your core markets is cheap and obvious. Avoid Ferraris and real-world logos in scenes you can replace with originals. None of this is new. All of it is bedrock. All of it gets ignored.

Where the framing comes apart is the genAI specific advice. The legal exposure on AI-generated assets is real for the nineteen percent doing asset generation and the five percent doing player-facing features. Those teams should think carefully about which outputs they treat as protectable studio IP and which they ship behind the work of human creators who can take authorship of the final form.

I have been in this fight. On a recent cricket project for a client extremely averse to infringement risk, the generated outputs kept producing real-world brand logos because the models had been trained on millions of images of actual cricket. Logos ghosted into helmet stickers, shirt sponsors, sightscreens, anywhere a real broadcast frame would carry them. I spent more time than I would have liked correcting them out in post. None of it was even for production assets. It was material going in front of the client. The risk Allan describes is real for this slice of the work, and managing it means clean-up workflows, prompt discipline, and human authorship layered over the generated output. The fix lives in the workflow.

Discussion about this post

Ready for more?