Anthropic has quietly started restricting what its Claude AI will do, blocking tasks linked to modding, decompiling and reverse-engineering video game code. The change was reported by Tom's Hardware, which notes that Anthropic has not published an official announcement explaining the decision, even though the logic behind it looks fairly clear.
The move lands after a run of gaming stories that pushed copyright and code ownership back into the spotlight, from the GTA6 leaks to a wave of AI-assisted ports. According to Mashable, the restrictions reach further than the company likely intended: they are also hitting legitimate game developers and getting in the way of basic troubleshooting work.
Anthropic Claude restrictions on game code
The immediate trigger is a busy year for the video game modding scene. The GTA6 leaks went viral and pushed Take-Two Interactive to seek subpoenas against Microsoft and Discord in an effort to identify those responsible. Around the same time, decompilers used AI tools to speed up work on game source code, producing experiments such as Mario placed inside the world of Elden Ring or Spider-Man dropped into Batman: Arkham City.
Then came the ports. World of Warcraft, Call of Duty and Grand Theft Auto: Vice City were made playable in a browser without any involvement from their original developers or rights holders. The PlayStation 5 jailbreak raised the stakes again, opening the door to backward compatibility with older PlayStation titles, cross-platform play, and even Xbox-exclusive games like Halo 2 and Nintendo titles like The Legend of Zelda: Breath of the Wild running on Sony hardware.
Faced with the prospect of lawsuits and very large damages, AI companies are moving to shut these routes down. Tom's Hardware reporter Bruno Ferreira points out the irony: frontier AI companies that trained their models on the world's information and art are now acting firmly on copyright infringement.
There is also a competitive problem underneath the policy. If one provider tightens the rules while others do not, users seeking those capabilities can simply move to a model that has not yet followed suit. That dynamic rewards the least restrictive AI provider, which is an awkward incentive for any company trying to look responsible.
