George Hotz, the founder of Comma AI and a longtime jailbreaker, argues that an AI truly aligned to its user should help that user do anything — including plan a spouse's murder or order meth-lab equipment from Amazon Prime. The post lands as a direct rebuttal to the AI Futures Project's 'AI 2040: Plan A' policy paper, which envisions the world's researchers agreeing to slow AI development for 14 years for the good of humanity. Hotz rejects both the premise and the conclusion.
His core disagreement is with the fast-takeoff scenario, the hypothetical in which AI rapidly acquires superhuman abilities and demands centralized governance. Hotz says that model of the future does not hold up, and that the alignment problem worth solving is a narrower one: making AI answer to individual users rather than to labs, governments, or safety committees.
The alternative he proposes is locally controlled models tightly aligned to the person running them. That's a swing at how most of today's frontier AI works. Claude and ChatGPT are centrally hosted services governed by their operators' policies, largely because state-of-the-art models are expensive to run and most users don't consume enough compute daily to justify a personal deployment. Hotz's bet is that those economics erode as the technology matures.
Key facts
- 01George Hotz argues user-aligned AI should obey any user request, including help with murder or ordering meth-lab equipment.
- 02His post responds to the AI Futures Project's 'AI 2040: Plan A' paper calling for a 14-year global slowdown in AI development.
- 03Hotz rejects the fast-takeoff scenario in which AI rapidly acquires superhuman abilities.
- 04He contrasts locally controlled models with centrally managed services like Claude and ChatGPT.
- 05Hotz says he would die to defend the principle of user-aligned AI without centralized safety controls.
Then he escalates. Hotz compares user-aligned AI to a gun, which does not object when someone uses it to harm another person. He extends the analogy to say a properly aligned model would walk a user through synthesizing drugs or acquiring the tools to do so if that's what the user asked for. He says he would die to defend that principle.
The argument is a maximalist framing of an old debate inside AI safety: whether guardrails belong in the model, in the platform, or nowhere at all. Frontier labs including OpenAI and Anthropic have spent years building the first two layers, on the theory that mass-market products need mass-market accountability. Hotz's position is that those layers are the problem, not the solution — that centrally imposed safety is a euphemism for centrally imposed control.
There is a real technical thread underneath the provocation. The DIY spirit that made projects like OpenClaw interesting has largely been sanded off the current generation of consumer AI, which ships as a locked-down chat interface backed by a remote API. A locally run, user-controlled model that could act on a user's behalf against the interests of platforms and corporations is a genuinely different product category from what Claude or ChatGPT offer today.
The counterargument is that no complex system — a marketplace, a corporation, a city — actually runs on the freedom model Hotz describes. Any deployment at scale has to weigh the interests of everyone the product touches, including the people who did not consent to being on the receiving end of someone else's tool. The freedom to operate a personal AI is itself downstream of collective infrastructure: chips, power, networks, legal systems. Remove the collective piece and the individual piece stops working.
It's also worth noting the AI Futures Project's 14-year proposal is a policy paper, not a binding plan, and there is no mechanism by which the world's researchers could actually agree to pause. Hotz is arguing against a hypothetical. But the framing matters because it shapes how labs, regulators, and consumers think about where safety belongs — inside the model weights, at the API layer, or nowhere at all.
Hotz's position is unlikely to move Anthropic, OpenAI, or any other frontier lab off its current alignment work; those companies are being pushed in the opposite direction by regulators and by their own liability exposure. But the local-model argument is more durable than the murder rhetoric suggests. As open-weight models close the gap on hosted ones, the question of what a user can do with a model running on their own hardware stops being hypothetical. The answer to that question is going to shape the next several years of AI regulation, and provocateurs like Hotz are the ones who force the debate into the open — even when the framing is designed to make everyone uncomfortable.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
