This line of reasoning is faulty. So if a cyber criminal instructs claude to commit crimes then it’s not their fault because the tool committed the crimes? What if it commits crimes that were never intended by the user? Who accepts liability? Claiming negligence, ignorance of the law, or “accident” does not absolve you from manslaughter or reckless endangerment.

The answer here is very simple. If you are building a tool and through negligence, reckless endangerment, or sheer incompetence your tool poisons or kills people — if it commits fraud, espionage, cyber crime, or any crime — you are liable. If you sell the product, then your customers have to accept liability for any crimes committed via their prompts. If neither you or your customers can accept liability, that means AI is too stochastic to control, or you aren’t competent enough to firewall and sandbox your AI “tool” in a way that prevents it from committing crimes, and you shouldn’t be allowed to build or operate those tools at all; definitely not selling it as a product.

You either accept liability, or you don’t have a legal, moral, or ethical right to operate. Open AI and Anthropic — all of humanity atm — are consistently proving they aren’t competent enough to contain AI, therefore they’re not competent enough to build or operate it.