Why AI Safety Matters to Me

AI safety may be the hardest and most important problem in Silicon Valley—and perhaps in the world today.

Some of the smartest people alive are working on it. The leading AI labs are investing billions of dollars. Researchers publish papers about alignment, interpretability, model behavior, and catastrophic risk. Yet after reading those papers and blogs, and listening to people such as Ilya Sutskever and Dario Amodei discuss the future of AI, I keep returning to two much more personal reasons why safety matters to me: education and biology.

My mother is a biology professor who has spent her career teaching and researching Alzheimer's disease. I grew up around her lab. Now I am also the father of a two-year-old.

Those two parts of my life shape how I think about AI: it's something that will influence how the next generation learns and how humanity understands life itself.

AI Will Help Raise the Next Generation

I care about education because I was born in a family of teachers and professors. My mom is a biology professor, my grandma high school physics teacher, and my aunt high school computer science. I'm also a dad of a two year old and the biggest question in my head now is "What does education mean to my son now with AI? What knowledge, qualities are still worth teaching?"

When I think about my own education, I do not remember only textbooks, equations, or exam results. I remember teachers.

A few great teachers made me who I am today. They did more than transfer knowledge. They shaped how I think, how I treat other people, how I decide what is right, and how I see the world. Their influence came not only from what they taught, but from how they taught it.

Children growing up today will have a very different experience. They will still learn from parents, teachers, and friends, but they will also learn an enormous amount from AI. No individual teacher can match the breadth of knowledge available to a frontier model. An AI tutor can be present at any hour, answer endless questions, adapt an explanation to a child's level, and move effortlessly from mathematics to history to biology. This potential—and the need for an age-appropriate, human-centered approach—is explored in UNESCO's guidance on generative AI in education.

That capability is extraordinary. It also comes with extraordinary influence.

When children repeatedly ask AI what is true, what matters, and what they should do, its answers will help shape their beliefs and values. The model's tone will teach them something about how authority behaves. Its willingness to admit uncertainty will teach them something about honesty. The way it handles disagreement will teach them something about respect.

This is why an AI system cannot be considered a good teacher simply because it gives correct answers.

We already understand this standard for humans. A person may know everything about mathematics, but if they are dishonest, cruel, or abusive, we would never call them a great teacher. Knowledge alone is not enough. We expect teachers to exercise judgment, care about their students, and understand the responsibility that comes with influence.

We should hold AI to an equally serious standard. That does not mean forcing one rigid worldview on every child. It means designing AI to consistently model qualities worth learning: honesty, kindness, respect, curiosity, and intellectual humility. Work such as Anthropic's published constitution for Claude is one attempt to make an AI system's guiding values more explicit and open to examination. It means helping children think rather than merely telling them what to think. It also means keeping parents and human teachers involved, because education is a relationship—not just an information-delivery problem.

An infinitely capable AI that educates without care could become a nightmare. A safe and trustworthy AI, on the other hand, could give every child access to a patient and knowledgeable tutor. The difference will not be capability alone. It will be the values expressed through that capability.

Biology Shows Both Sides of AI's Power

The second reason is biology.

Biology, medicine, and health may be the areas of science people care about most personally. They touch birth, aging, disease, suffering, and death. They determine not only how long we live, but how well we live.

My mother has spent her career researching Alzheimer's disease. As a child, I spent time in her lab around the mice used in her research, and I sometimes helped feed them. It was a fun experience for me then. Looking back, it also gave me an early view of how slow and difficult biological research can be. Progress requires years of careful experiments, and even brilliant researchers can explore only a tiny fraction of the possible questions.

AI could change that dramatically. It could help researchers understand complex biological systems, discover new medicines, design better experiments, diagnose diseases earlier, and develop treatments faster. AlphaFold's impact on protein-structure research is already a powerful example. It could make every scientist more productive and reveal connections that no human could find alone. In biology, the upper limit of AI's usefulness is difficult to imagine. The World Health Organization's work on AI for health similarly emphasizes both this potential and the need for safe, ethical, and equitable adoption.

But the same knowledge that can heal can also cause harm.

A system that is excellent at reasoning about proteins, chemicals, and biological processes might help develop a lifesaving drug. In the wrong hands, that same capability might help someone create a dangerous pathogen, synthesize a toxic compound, or bypass safeguards that currently require specialized expertise.

This is the central dual-use problem. The very capabilities that make AI for biology so promising also make it dangerous. The more powerful the system becomes, the larger both outcomes become. Anthropic's research on why advanced language models may create additional biological risk offers a concrete view of how this risk is being measured.

We cannot solve that tension by abandoning biological AI. The potential benefits are too important. We also cannot treat safety as something to add after the models become capable. Safety has to advance alongside capability: through careful evaluations, access controls, monitoring, secure research practices, and models that refuse to assist with genuinely dangerous work while remaining useful to legitimate scientists. Frameworks such as Anthropic's Responsible Scaling Policy and OpenAI's Preparedness Framework attempt to connect increasingly dangerous capabilities with increasingly strong safeguards.

In biology, safety is not an obstacle to progress. It is what makes progress sustainable.

Safety Is Also a Business Advantage

When people see brilliant researchers and wealthy companies investing enormous amounts of money in AI safety, they may wonder why. Is it simply because these are good people who care more about the mission than money?

That may be part of the explanation. But there is also a practical reason: safety creates trust, and trust creates durable business value.

We can already see capability becoming widely available in many dimensions. Open-weight models such as Kimi K3 and DeepSeek V4 increasingly approach the performance of proprietary systems such as Claude and ChatGPT. Almost anyone can use AI to build a surprisingly good web or mobile app. In some cases, AI-generated products are already better than what an average developer could have built alone a few years ago. As frontier models become accessible to everyone, the capability gap between individual builders will shrink. The capability gap between top-tier models is also shrinking.

Imagine that one hundred teams can build products with roughly the same features. Why would a customer choose one instead of the other ninety-nine?

Increasingly, the answer will be trust.

People will choose the company they trust with their data, their health, their children, and their important decisions. Enterprises will choose the provider they believe will behave responsibly when something goes wrong. Customers will care not only about what a product can do, but about whether its creator has been honest, careful, and dependable over time.

Trust cannot be generated by a marketing campaign. It is accumulated slowly through a long-term commitment to safety—and it can disappear after a single serious failure. The NIST AI Risk Management Framework offers one practical foundation for turning that commitment into continuous governance, measurement, and risk management.

As AI makes raw capability cheaper and more common, trust may become one of the few advantages that cannot be copied overnight. A competitor can gain access to the same model. It can reproduce a feature. It cannot instantly reproduce years of responsible behavior. AI Safety is a moat.

Business value is a by-product of safety

AI safety is sometimes discussed as if it were a narrow technical field or a constraint placed on innovation. I see it differently.

Safety determines whether AI becomes a teacher we would trust with our children. It determines whether advances in biology cure diseases or create new dangers. And it determines which people and companies earn the confidence required to build lasting products.

The goal is not to make AI less capable. The goal is to make its growing capabilities reliably serve human beings.

Pursuing that goal is a moral mission. If we do it well, business value will follow as a natural byproduct. In the long run, the most valuable AI may not be the one that can do the most. It may be the one humanity can trust the most.