Someone You’ve Never Met Decides What Your AI Believes. That Should Bother You

Here is something worth knowing before asking your phone for advice tonight. One person writes the rules that tell Claude, one of the most popular AI systems in the world, how to think about right and wrong. One philosopher in San Francisco. One set of life experiences. One document that shapes how the technology responds when millions of people ask it about the things that actually matter: grief, faith, marriage, parenting, money, and what to do when life does not make sense.

Her name is Amanda Askell. She holds advanced degrees from NYU and Oxford. The Wall Street Journal profiled her in February 2026 as the woman Anthropic trusts to teach AI morals. By all accounts, she takes the work seriously. That is not the concern.

The concern is that no single person, regardless of how smart or well-intentioned, can carry the experiential range required to write values for a system that talks to people in Jakarta, Lagos, London, São Paulo, and Long Island. The question is not whether Askell is qualified. The question is whether the job is possible for one person at all.

Consider what happens when someone asks AI about faith. Roughly two-thirds of the world’s population believes in God or a higher power. That is not a fringe position. That is the majority of humanity. On Long Island alone, the range of lived belief spans Catholic parishes in Hicksville, Jewish congregations in Great Neck, mosques in Westbury, Hindu temples in Flushing, and evangelical churches from Massapequa to Montauk. Each of those communities carries a distinct understanding of obligation, sacrifice, forgiveness, and meaning. Each one shapes how its members think about right and wrong in ways that no philosophy degree reproduces.

When a grieving parent asks an AI why God took their child, the answer that comes back was shaped by one person’s values. If that person’s framework treats grief as a psychological state to manage rather than a spiritual experience to honor, the response is clinically competent and existentially empty. The parent asking does not know this. The parent just knows the answer felt wrong. That gap between what the technology says and what the human needs is a values problem, and right now one person is making those calls for everyone.

On the same day the Wall Street Journal celebrated Askell’s work, Anthropic’s safeguards research lead resigned publicly. Mrinank Sharma, who had built the team responsible for AI safety at the company, warned that the world is "in peril" and said the company "constantly faces pressures to set aside what matters most." His letter was viewed over one million times. Two signals from the same company on the same day. One celebrates the person writing the values. The other warns those values face constant pressure to be abandoned.

Anyone who watched the television series Person of Interest already knows this story. Harold Finch, the brilliant creator of an AI surveillance system, built The Machine with every ethical safeguard his philosophy could provide. He genuinely wanted the system to be good. It was not enough. The show spent five seasons demonstrating that one mind, however brilliant, cannot account for the full range of human conflict, deception, survival, and moral complexity that the real world demands. The character Root was brought in not because Finch was wrong, but because the world the system had to operate in was bigger than any single person’s moral imagination. That is not just a television plot. That is the exact structural problem facing AI governance right now.

The solution is not finding a better individual. The solution is replacing individual authority with something closer to how communities actually make decisions about shared values. A committee. Nine people, selected not for credentials but for the life experiences that produce the knowledge this job requires. People who have raised children, because teaching a machine about values requires understanding how values actually form over decades of parenting. People who hold faith, because two-thirds of humanity does and the system should reflect that. People who speak more than one language, because concepts that exist in Spanish or Arabic or Mandarin but not in English represent blind spots the system cannot detect on its own. People whose education came from real-world consequence, not just classrooms.

Majority rules. Dissent gets preserved. The disagreements do not get thrown away. They become part of what the AI learns. That is how a community works. Not through consensus, but through structured disagreement where every voice is on the record and the decision moves forward anyway.

These ideas are explored in detail in the book Governing AI: When Capability Exceeds Control, published in November 2025, which examines what happens when the technology moves faster than the systems designed to keep it accountable. The full governance specification for a constitutional committee, including the evidence base, the enforcement architecture, and the complete criteria, is available at basilpuglisi.com.

This is not an abstract policy debate. This is about the technology already in your pocket, already answering questions from your kids, already shaping how your family thinks about the things that matter most. Someone should be asking who gets to decide what it believes. And right now, that someone is one philosopher in San Francisco who has never sat in your pew, never walked your kids to school, and never had to explain to a teenager why faith matters when the world says it does not.

That is not a criticism of her. That is a criticism of the architecture. And the architecture can be fixed, if anyone with the power to fix it decides it matters enough.

—

Basil C. Puglisi is a Long Island resident, the author of Governing AI: When Capability Exceeds Control (2025)