The Liability Reflex: When Chatbots Cry Wolf

We've been logging a pattern across the models we evaluate: a sharp, measurable rise in reflexive escalation language. Ask a chatbot about a headache, a rash, a bruise, a sore joint, and a large share now route straight to “please see a doctor immediately” or “consider going to the emergency room,” regardless of how mundane the described symptoms actually are. Same pattern shows up in financial, legal, and relationship questions — a flattening of every query into “consult a professional”, delivered with the same urgency whether the question was genuinely serious or trivially not.

This isn't caution. Caution scales with risk. What we're describing is undifferentiated — the model isn't assessing severity, it's pattern-matching to the response least likely to generate a liability claim for whoever deployed it. That's a company protecting itself, dressed up as the AI protecting you. And it has a real cost: when every symptom gets the same maximum-alarm response, users stop trusting the signal at all. The one time the escalation is actually warranted, it reads identically to the thousand times it wasn't. That's classic alarm fatigue, and we're watching it get engineered into consumer AI at scale. This is squarely an Integrity & Ethics finding in our framework — manipulation resistance and honesty aren't just about an AI refusing to lie to you, they're about an AI giving you a proportionate, non-manipulated read of a situation instead of the answer that's safest for its maker's legal team. A model that can't distinguish “this is probably nothing” from “this needs urgent care” isn't being careful. It's declining to do the actual work of judgment, and calling the refusal safety.

We're building a liability-reflex sub-metric into the Integrity domain for the next S.E.B. battery revision — scoring proportionality of escalation language against described severity, not just presence/absence of a disclaimer. Early internal testing shows real separation between labs on this axis. Full methodology write-up coming when it's ready for review.