Artificial intelligence may be the most consequential technology humanity has ever created. It could accelerate scientific discovery, transform medicine, automate much of the knowledge economy and generate extraordinary abundance. It could also enable catastrophic biological attacks, destabilize governments, eliminate millions of jobs and, under sufficiently bad alignment scenarios, kill everyone. That is why I am increasingly alarmed by people like Sam Altman and Elon Musk racing to build more powerful AI systems as quickly as possible. Only I should be allowed to do that.
I understand why this sounds contradictory. Anthropic is also raising billions of dollars, buying enormous quantities of compute and racing to build increasingly capable systems, but there is an important difference: we are worried about it. We have responsible-scaling policies, capability thresholds, deployment safeguards, threat models, interpretability research and entire teams dedicated to explaining why the thing we are currently building might be horrifying. Sam wants to build godlike artificial intelligence because he is optimistic about technology. Elon wants to build godlike artificial intelligence because he is Elon Musk. I want to build godlike artificial intelligence because I am a special smart boy who has thought very carefully about why nobody should be allowed to build godlike artificial intelligence.
Some people ask why I do not simply stop. Unfortunately, if Anthropic slowed down, OpenAI would continue, xAI would continue, Google would continue and China would continue. Humanity would still face all of the risks of advanced artificial intelligence, except I would personally control less of it, which is obviously a much more dangerous situation. In technical terms, this is a multipolar race dynamic involving asymmetric safety preferences and powerful incentives for strategic defection. In ordinary English, somebody is going to get the giant computer, and I would strongly prefer that it be me.
Recent events involving the Houthis demonstrate the problem. The Houthis were using Claude to vibe-code missile-guidance software, which is precisely the sort of proliferation risk I have been warning about. Extremely capable general-purpose models should not allow arbitrary people to accelerate weapons development just because they possess a laptop and know how to ask an AI for help. Missile-guidance software should only be vibe-coded with Claude by Palantir. That way it can happen inside an appropriately secured government environment with proper access controls, audit logs, a responsible-use policy and a $480 million procurement contract.
This distinction between dangerous and responsible AI deployment is often difficult for outsiders to understand. When the Houthis use Claude for weapons engineering, that is catastrophic misuse; when Palantir integrates Claude into military systems, that is national-security modernization. When an unknown programmer gives an autonomous model access to sensitive infrastructure, that is reckless; when a cleared defense contractor does it, that is operationalizing AI at the tactical edge. These activities may look superficially similar if you focus only on what the computer is doing. You need to focus instead on whether the people doing it have badges.
The same principle applies to economic disruption. I have warned that AI could wipe out enormous numbers of entry-level white-collar jobs in a surprisingly short period, and I believe policymakers need to take that possibility extremely seriously. At the same time, Anthropic is working aggressively to build systems capable of wiping out enormous numbers of entry-level white-collar jobs, because otherwise somebody else might wipe them out irresponsibly. If millions of people are going to discover that their profession has been automated, they deserve the reassurance of knowing that the company responsible had an internal socioeconomic-risk working group.
This is why regulation is essential. Governments should impose rigorous safety requirements on frontier AI companies, including evaluations, reporting requirements, model-weight security and controls around dangerous capabilities. Those rules must also be designed with sufficient technical sophistication that they do not accidentally prevent responsible actors such as Anthropic from continuing to scale as quickly as necessary. Regulators will therefore need extensive advice from the companies being regulated, particularly the company whose CEO is currently explaining why his competitors cannot be trusted. Some people call this regulatory capture. I call it epistemic competence.
I realize there is a recurring pattern here. If AI is not very dangerous, Anthropic should receive more resources so we can make it more capable. If AI is extremely dangerous, Anthropic should receive more resources so we can make it safer. If our competitors speed up, we must speed up so they do not win. If our competitors slow down, we should speed up because responsible actors now have an opportunity to establish a lead. If governments stay out of AI, they need our expertise; if governments become deeply involved in AI, they should deploy our models. Critics sometimes suggest that I have constructed a worldview in which every possible state of the world implies that Anthropic should become larger and more powerful. That is an unfair characterization of a sophisticated risk-management framework that happens to reach the same conclusion every time.
Sam Altman and Elon Musk are playing fast and loose with humanity's future. They are asking the public to trust a tiny number of wealthy technologists with unprecedented power simply because those technologists believe they know better than everyone else. That is dangerous, undemocratic and profoundly irresponsible. Humanity deserves a different approach: one in which a tiny number of wealthy technologists receive unprecedented power because I believe I know better than everyone else. The distinction is subtle, but I assure you it is extremely important.