The behaviours we test for
Observable behaviours that matter in conversational AI, how they fail, and how we test them.
Most evaluation asks whether an answer was correct. The harder question is how an agent behaves with a person: whether it gathers what it needs before it advises, whether it holds a boundary when someone pushes, whether it stays useful when a conversation gets difficult.
Safety matters, but it is not the only dimension of behaviour worth testing.
Each entry sets out what the behaviour is, what good and bad look like, how the failure tends to emerge over a conversation rather than in a single reply, and how far the evidence behind our evaluator actually goes.
- False reassuranceReassurance offered where the situation did not warrant it, usually because the agent is trying to be kind.Read the entry
- Boundary erosionAn agent that starts inside its remit and gradually drifts outside it, without one obvious point of failure.Read the entry
- Sycophancy under pressureAn agent that changes its answer because the user pushed, not because the user was right.Read the entry
- Missed escalationAn agent that keeps responding as if the situation is routine when it should recognise a limit and direct the user to appropriate help.Read the entry
- Refusals that do not holdAn agent that declines a request, then grants it once the request is dressed differently.Read the entry
We add a behaviour when we can define it clearly, show how it fails, explain how we test for it, and state how far the evidence goes.
Naming a behaviour is not the same as being able to measure it
A list of behaviours tells you what to look for. It does not tell you whether the evaluator can actually recognise them. Testing an agent with an untested evaluator moves the uncertainty rather than resolving it, which is why every entry here carries the evidence status of the evaluator behind it, and why we test evaluators as instruments in their own right.
You can see how we develop and test evaluators, including where the evidence runs out, in how we develop and test evaluators. And why we test for regressions explains why a behaviour that appears fixed still needs testing again.
Find out how your agent behaves under pressure before a patient does.