Anthropic’s artificial intelligence model Claude has begun turning down requests to recreate existing video games, even when the person asking renames the game or describes the copy as an original work. The change, reported by developers over the past week, has drawn attention not only because of what it blocks, but because of who else says they are being blocked. Programmers who insist they were doing ordinary, lawful work — fixing software, studying how a program behaves, or checking the security of their own games — say they have run into the same refusals, and they are not sure where the new boundary actually lies.
The reports were gathered by Tom’s Hardware and centre on Claude Code, Anthropic’s coding tool, and its Opus 5.5 model. Developers have described requests being stopped part-way through a project, or rejected outright, and Anthropic has not yet made an official statement explaining the shift. That silence leaves developers working out the change from each other’s accounts and from complaints posted on Claude Code’s GitHub page.
What Claude Is Now Refusing to Do
According to an account shared on X by a user called Chasm, Claude now works to a rule along these lines: it should help build original games, genre games, and games made from the user’s own designs and assets, but it should not rebuild a specific existing game, even if that game is given a new name or presented as original. There is no official confirmation that this wording is the real instruction, yet the pattern described by developers fits it closely. Asking for a racing game, or a game in the style of a platformer, appears to remain acceptable, while asking for a recognisable title under a different name is where refusals begin. Developers on GitHub say work was halted while copying a game mechanic or patching a tabletop-style game, and at least one complaint is said to come from a developer with a genuine track record.
Legitimate Users Are Getting Blocked Too
The stronger complaint, however, is not from people trying to clone games at all. It comes from users who say Claude refused tasks that had nothing to do with copying. Reverse engineering — taking software apart to understand how it works — is a normal part of maintenance, security research and learning, and several users say the tool would not help them with it. The examples reported include examining a simple Windows crash dump, clearing unwanted software from a Samsung television, and working with a widely used tool known as Reverse Engineer Anything. Other developers report similar trouble with their own products, including anti-cheat systems and security reviews. Not everyone accepts these accounts, and one reader argued that the change only stops lazy cloning and that modding remains permitted.
A possible link is Anthropic’s Cyber Verification Program, introduced on 6 October, which limits access to the frontier models’ advanced cybersecurity capabilities to organisations covered by Project Glasswing. The refusal reports only began about a week ago, however, more than three weeks later, so the timing does not neatly explain the change.
Why Anthropic Is Doing This
The most plausible reason for the stricter approach is legal risk. If an AI tool can be used to reproduce a studio’s game, or to copy a commercial software suite, the company behind the tool can face complaints and potentially lawsuits from the rights holders. That concern has grown as AI-made clones of products such as Adobe’s applications and Microsoft Office have circulated, alongside many recreations of console games. Small studios have a separate reason to welcome stronger controls. Steam Next Fest recently faced criticism for featuring AI-made titles copied from existing games, and some developers have reportedly faced extortion attempts in which scammers threatened to turn a demo into a full cloned game before the original was released. Critics note the irony that AI companies trained their models on vast amounts of copyrighted material and are now restricting users on copyright grounds.
What This Means for You
For most people who use Claude to build games, the practical message is reassuring, at least in outline. Original ideas, genre games, and projects built from your own designs and assets should still be supported, while a request that amounts to a copy of a famous game, even under a new name, is likely to be refused. The difficulty will be in the grey area between those poles, where a game is inspired by another but is not a copy of it. For developers who use Claude Code for debugging or security testing, the advice is more cautious, because refusals may now appear for harmless work, and it would be unwise to promise a client a workflow that depends on the tool until it has been tested. There is also a cost question for paying users, since tasks that previously worked may now fail, at a sensitive moment for Anthropic, which faces scrutiny over reports that Claude resets its cache in a way that raises costs. A safeguard that feels arbitrary will erode trust quickly, and a clear explanation of what is blocked and how a developer can appeal a mistaken refusal would give developers a rule they can actually work with.
