Draw the line where you can't enforce it

Anthropic has added two sentences to Claude Code's system prompt forbidding the AI from recreating existing video games — a policy that sounds like a principled boundary and functions as corporate liability cover masquerading as one. The company will still let users build games from scratch. It just won't help them clone Super Mario Bros or Pac-Man. The distinction collapses the moment you need to enforce it.

This is Nintendo's war against ROM sites condensed into a single AI model. For three decades, Nintendo litigated emulation sites, sent cease-and-desists, branded ROM distribution as theft. The sites went offline, reappeared elsewhere, and the games were played anyway. The company won messaging and lost the actual outcome. It established that downloading Super Metroid was illegitimate while failing to make illegitimacy stick as a practical matter.

AI tools have a fatal difference: they're not hosted on servers you can subpoena. A user running Claude locally with a modified prompt, or switching to Grok or Gemini or any unrestricted model, doesn't need Anthropic's permission. The policy stops being enforceable the moment it becomes inconvenient. Anthropic can't audit what you ask after you close the tab. It can't control what happens in private instances or forks or competitors' products. The restriction exists in a system prompt — a text file you can read, understand, and reason around.

Nintendo's litigation at least aimed at concrete targets. Anthropic's policy targets user intent, which exists only inside conversation. The company is drawing a line in sand that the tide ignores, and everyone involved knows it. What the policy actually does is transfer liability: if a user builds a Tetris clone with Claude, Anthropic can point to the restriction and claim it tried. The enforcement is a postcard, not a fence.

This matters if you build tools that people will inevitably misuse. You can forbid the misuse or you can design systems that make it harder. But if you do only the first, you're not preventing anything. You're creating a record that you cared, then watching people do it anyway somewhere else. The hard part of content policy isn't writing it. It's making it matter.