Anthropic Tightens Claude Code to Stop Re-Creating Existing Video Games
Anthropic has updated Claude Code's instructions to discourage recreating existing video games, while still letting users build original games with the tool.
Anthropic has updated the system prompt for Claude Code, its AI coding tool, to keep it from being used to re-create existing video games, according to The Verge, which cited a report from Tom's Hardware. The change targets a growing problem in retro gaming communities, where AI-generated decompilations, recompilations and ports have become widespread.
The Verge reports that Anthropic still allows people to use Claude Code to make games. The company's stated concern, as described in the reporting, is with copies of existing titles rather than with game development as a whole.
CNET describes the update as minor. It frames the issue as a fine line between AI creations inspired by video games and those that infringe copyright. That distinction sits at the center of the change, though the reporting does not spell out exactly where the line falls in practice.
The reports do not say how the new instruction will be enforced, how many projects have been affected, or whether Anthropic plans further changes. Those details remain unclear.
Background
Video game copyright covers more than a single file. Developers hold rights to their code, art, music, levels and characters, and those rights generally last for many years after release. Ideas, mechanics and genres are typically not protected on their own, which is why many games borrow from one another. The gap between borrowing an idea and copying expression is where most disputes arise.
Decompilation is a technique that tries to turn compiled machine code back into something closer to readable source. Recompilation takes that work further by rebuilding a program so it runs on modern hardware. Porting moves a game to a new platform. All three have long existed in hobbyist circles, often tied to emulation and game preservation, where enthusiasts and some nonprofit archivists try to keep older titles playable as hardware ages out.
AI coding tools have made these tasks faster and easier. A user can describe a goal in plain language and receive working code in return, which lowers the effort needed to reproduce a game. That speed is what has alarmed some developers and rights holders, because a result that once took months of skilled labor can now appear in hours.
System prompts are the hidden instructions a company gives an AI model before a user's request. They shape what the model will and will not do, and they can be changed without retraining the underlying system. An update to a system prompt is therefore a lighter-weight form of policy than a formal rule, and it can change quickly.
Companies that build AI tools generally write usage policies and technical guardrails to limit uses that could expose them or their customers to legal risk. Copyright is one of the most active areas of that risk, and it has drawn lawsuits and regulatory attention across the technology industry.
Why it matters
This update matters because it shows how AI companies are beginning to draw lines around specific creative works. Rather than waiting for a court ruling on whether AI-generated copies infringe copyright, a company can adjust its tool's behavior now. That approach can protect rights holders and reduce the company's own exposure, but it also means private policy choices are shaping what creators can do.
For game developers, the change offers some reassurance that a widely used tool will not be a shortcut for copying their work. For retro enthusiasts and preservationists, the situation is more complicated. Many people who work on old games have legitimate goals, such as keeping titles accessible, and an instruction that treats all re-creation as suspect may catch some of that work too.
The update also highlights the difficulty of separating inspiration from infringement. A new game in a familiar style is normal creative practice, while a near-exact recreation of a protected title is not. Where a tool sits between those two points, and who decides, will affect how much people trust AI systems with creative work.
What to watch
Watch for whether Anthropic publishes more detail on what the instruction covers and how it applies to users who build original projects. Watch also for reactions from game studios, retro communities and preservation groups, since their responses will show whether the change is seen as fair. Finally, watch for whether other AI coding tools adopt similar limits, which would signal an industry pattern rather than a single company's decision.
Viewpoints
Commentary from our columns: three angles on this story.
The Skeptic open questions and weak points
The most useful question about this update is also the simplest: what, exactly, counts as a clone? A system prompt can tell a model to avoid re-creating existing games, but the instruction only works if the model can tell the difference between a faithful copy and a new game that borrows familiar mechanics. Those lines are blurry, and a tool trained to avoid one kind of output may refuse legitimate work that happens to look similar.
There is also the matter of how the rule is enforced. A change to instructions is quick and easy to make, but it is difficult for outside observers to verify. Users cannot see the prompt in full, and there is no public record of how often it triggers or where it fails. Until there is more transparency, the update is a promise rather than a measurable result.
Finally, the phrase "minor update" deserves scrutiny. Small changes can carry large consequences when they set precedent for how AI companies treat creative works. The real test will be whether the policy holds up when it meets a hard case, not whether it sounds reasonable in a single announcement.
The Builder the engineer's view
From an engineering standpoint, the hard part here is not writing the instruction. It is making a model reliably honor it. Natural language guidance can steer behavior, but it is probabilistic, so developers should expect it to work well on obvious cases and less well on borderline ones. Anyone building on these tools should plan for that gap rather than assume a single line of text solves the problem.
Developers working on legitimate projects should also think about how their requests are phrased. A project that references an older game for design reasons may be treated differently from one that asks for a near-identical reproduction. Clear, original specifications tend to produce better results and fewer surprises, and they also make it easier to show that the work is new.
For teams that maintain their own tools, this update is a useful reminder to review how upstream models behave and to document their own content practices. Relying on a vendor's policy without testing it is a risk in any production pipeline.
The Rulebook privacy, safety and regulation
Rules that govern creative output tend to work best when they are clear, public and consistent. This update is none of those things yet. It was made through a change to a system prompt, which most users cannot read, and its scope is described only in general terms. People who rely on these tools deserve a clearer statement of what is allowed and what is not, written in language they can check.
The copyright questions here are real and unsettled. Rights holders have a legitimate interest in preventing copies of their work, and AI companies have a legitimate interest in avoiding liability. But legal disputes over training data, output and fair use are still working their way through courts and regulators, and private policy changes can shape outcomes before those questions are answered.
Good governance would involve published guidelines, a way for users to appeal or clarify a decision, and a commitment to protecting legitimate preservation and original creative work. Those steps would make the policy more fair and easier to trust.