# The Kid Is Going to Get Into the Kitchen > The goal was never to keep the kid out of the kitchen. The goal was to raise someone who knows what to do once they're in it. Sam Altman is right that useful AI will sometimes cause harm. The harder question is what responsibility comes with the agency we say we want. **Published:** 2026-10-06 **Tags:** AI, human agency, AI safety, responsibility, applied AI **Canonical URL:** https://blog.nikdesign.ca/posts/the-kid-is-going-to-get-into-the-kitchen --- Sam Altman said something in a recent POLITICO interview that has been getting a lot of attention. > "We believe that the world should accept some bad things happening for the benefits of this technology and people having the agency." I mostly agree with him. That is probably not the reaction people expect from someone who spends so much time thinking about human control, guardrails, provenance, approval gates, and where AI should stop. But I don't think safety means nothing bad can ever happen. If we want AI to become genuinely useful, genuinely capable, and genuinely available to ordinary people, then mistakes are part of the deal. Models will be wrong. People will misuse them. People will misunderstand them. Someone will trust an answer they should have checked. Someone will use a legitimate capability for a stupid purpose. Someone will do something none of the people who built the model anticipated. We should work very hard to reduce those failures. We should not pretend we can reduce them to zero. The interesting part of Altman's statement is not the acceptance of risk. It is the word **agency**. Because agency comes with responsibility. And that keeps making me think about my kitchen. ## I am learning to cook I am only now learning to cook properly. Naturally, this has made the kitchen a great place to cook analogies. Imagine you have a young kid. The kitchen has knives. Fire. Heavy things. Hot things. Chemicals. Electricity. Glass. A drawer full of mysterious implements designed to stab, slice, grate, crush, peel, or burn something. A kitchen is dangerous. So when the kid is very young, you keep them away from parts of it. You put up a baby gate. You lock a cabinet. You move the knives somewhere they cannot reach. But the kid is going to get into the kitchen. Eventually, they have to. That was always the point. The baby gate was never supposed to guarantee that this human being would go through life without encountering a knife. It existed because there was a period when the child did not yet have the capability to use the knife safely. Eventually they watch you cook. Then they help. Then they get something relatively safe to cut. Then they use a real knife while you stand beside them. Eventually you stop standing beside them. One day they make dinner for you. The knife did not become less sharp. **The human became more capable.** ## Safety cannot mean permanent childhood There are two absurd conclusions we could draw from the dangerous kitchen. One is that knives hurt people, therefore nobody should have knives. The other is that the kid will eventually encounter a knife anyway, therefore we should hand them one now. Nobody sensible does either. We build safeguards around capability, and we change those safeguards as capability develops. That is the part of the AI safety conversation I think we are going to have to get much better at. A system that can never produce anything wrong, dangerous, offensive, uncertain, or open to misuse is probably not going to be a very capable general-purpose system. But a system being capable does not mean every person should have every capability, in every context, with no friction, supervision, or consequence. **Capability is not permission. Access is not judgment.** This is not an argument against guardrails. It is an argument for understanding what guardrails are for. A good guardrail protects someone while preserving the possibility that one day they will not need it. If our definition of AI safety becomes "the human must never encounter anything requiring judgment," we have not created agency. We have created dependency. ## Every AI lab is a kitchen This analogy works at another level too. Every AI lab is a kitchen. Mine happens to be a little kitchen in an apartment. OpenAI is running a fucking catering service. Anthropic, Google, Meta and the other frontier labs are operating kitchens at industrial scale. That difference matters enormously. If I cook something badly in my apartment, the blast radius is small. I know who is eating. I can watch what happens. I can throw out the dish and change the recipe. If you are serving hundreds of millions of people, you need a different standard. You need the equivalent of health codes, allergy procedures, temperature controls, inspections, training, traceability, incident response, and a way to understand what happened when somebody gets sick. Scale changes the duty of care. It does not change the basic nature of the work. So this is not me saying OpenAI should throw every capability over the wall and tell society to figure it out. The companies building frontier models have responsibilities proportional to the power of the systems they build and the number of people affected by them. Children should have stronger protections than adults. High-consequence actions should have stronger controls than low-consequence ones. Known failure modes should be mitigated. Catastrophic risks belong in a different category from ordinary mistakes. And when a company discovers that its kitchen is poisoning people, "cooking has risks" is not a sufficient answer. But there is another side to this bargain. A catering company cannot teach your kid how to cook. ## The lab cannot carry all of our responsibility Parents still have to parent. Teachers still have to teach. Employers still have to decide where AI belongs in a workflow and where human judgment remains non-negotiable. Governments still have to decide what society will permit. Product designers still have to decide where friction is protective and where it merely gets in the way. And users still have to decide when to trust, when to verify, when to stop, and when to ask someone else. We cannot demand increasingly powerful tools while progressively transferring every responsibility for our use of those tools back onto the companies that made them. I have written before that frontier models need drivers. I have written that the genie is already out of the bottle and the skill we now need is learning how to make a wish. I have written that human-in-the-loop is not a decorative approval button. It is accountability. These are all versions of the same argument. The point of human-AI systems is not to engineer human judgment out of existence. It is to put human judgment where it matters. ## Eventually, the human has to become the safeguard This is where I land on Altman's statement. Yes. Some bad things are going to happen. That does not mean we stop building safeguards. It means safeguards cannot be our only answer. We also have to build capability in the people using the systems. AI literacy cannot mean memorizing a list of things a chatbot might get wrong. It should mean developing instincts. Knowing when an answer sounds too finished. Knowing when the source matters more than the summary. Knowing when the model has crossed from generating possibilities into making a consequential claim. Knowing when to check. Knowing when not to delegate. Knowing when a task is reversible. Knowing when another human needs to be involved. Eventually, the human has to become the safeguard. That is what growing up with powerful technology looks like. ## Agency comes with responsibility I am pro-AI precisely because I want people to have access to powerful tools. I do not want the most capable systems in the world locked inside a handful of institutions that decide what the rest of us are allowed to do with intelligence. But I also do not think "agency" is a magic word that dissolves responsibility. Agency is the opposite. Agency means you can act. And if you can act, eventually you have to answer for how you act. The frontier labs have to build better kitchens. Those of us building on their models have to build safer workflows. Institutions have to establish boundaries appropriate to their risks. Parents have to parent. Users have to learn. We should not close the kitchen. We should not throw away the knives. And we should not hand one to a four-year-old and congratulate ourselves for empowering them. We teach. We supervise. We put up the baby gate. We move it when the kid is ready. We accept that they will still cut themselves occasionally. And eventually we trust them to cook. Because the goal was never to keep the kid out of the kitchen. **The goal was to raise someone who knows what to do once they're in it.** --- *The Sam Altman remarks discussed here were made in an October 2026 interview with POLITICO's Decoded. Contemporary reporting on the interview includes [Fortune](https://fortune.com/2026/10/05/sam-altman-ai-risks-bad-things-benefits-trump-voluntary-safety-pact-openai-regulation/) and [Business Insider](https://www.businessinsider.com/sam-altman-says-ai-benefits-outweigh-some-bad-things-happening-2026-10).*