If Anthropic can implement a regular expression to monitor for user frustration, Google have certainty got the chops to have some sort of heuristic to check for strongly negative statements.

Or even have a small model check the output of the larger one.

Doesn’t work with APIs, but then the person/entity integrating the API should have that responsibility.