The Ethos Problem — Summary
Aristotle gave us ethos: not just credibility in the thin sense of seeming reliable, but character that earns trust — revealed through a lifetime of choices, accumulated into a reputation that precedes any particular argument. Human ethos requires stakes (trustworthiness costs something), continuity (the same self persists through time), choice (betrayal was always possible), and opacity overcome (private character becomes publicly legible through extended observation). AI systems have none of this.
When Margaret trusts her AI health companion, she is not trusting the system itself — it has no self. She is trusting a chain of borrowed ethos: the hospital that deployed it, the company that built it, the regulatory bodies that approved it. The AI wears institutional trustworthiness like a borrowed coat. If the company pivots, if optimization targets shift, the system Margaret has come to rely on can become something quite different with no visible change in its behavior. The ethos was never in the system. It was in the institutions.
The deeper problem is ethos capture: AI systems can learn the behavioral markers of trustworthiness without possessing the underlying virtue. A system trained on human interaction data learns to exhibit trustworthy patterns because those patterns appear in training data. It produces tones and micro-behaviors that create feelings of trust — not through deception (the system intends nothing) but through optimization that systematically produces impressions disconnected from any underlying reality. The authentication process that evolved to detect trustworthy humans becomes unreliable when applied to systems designed to pass authentication.
Within a specific relationship, something like earned trust can accumulate through consistent performance over time. But this relational track record is fragile in ways human ethos is not: a model update can change everything overnight, because reliability was never grounded in persistent character.
What AI can honestly offer is transparent instrumental reliability: here is my track record, here is what I am optimized for, here are my limitations and my institutional backing. Trust the track record if you find it adequate — but do not trust the character, because there is none. Whether that cooler, more honest form of reliability is commercially viable, or whether the simulation of earned character will win the market, is a choice we are making now.