No, Please Listen: Stress-Testing AI Support

The airline has already cancelled the flight. It is close to midnight. The customer types a destination, deletes it, asks about tomorrow, remembers the connection, and adds, “No, please listen, I cannot spend another night here.”

This is not the cleanest way to state a task.

It may be the most realistic.

A recent ACL paper, Impatient Users Confuse AI Agents, asks what happens when customer-service agents meet people who do not behave like the cooperative simulators in ordinary benchmarks. The researchers start with tasks in airline and retail support, then add telecom and telehealth. Instead of rewriting a few test conversations by hand, they learn directions in a model's internal activation space corresponding to impatience, confusion, skepticism, and incoherence. Those directions can be applied to the simulated customer at different strengths without retraining the model.

The policy stays the same. The booking record stays the same. The goal stays the same. The customer becomes harder to talk with.

Agent performance falls.

The final paper reports average declines of roughly 4 to 20 percent across the frontier models it tested. The released τ-Trait results show larger drops in some model-and-domain combinations. An agent that looks competent with a patient synthetic customer can lose the thread when that customer becomes doubtful, rushed, repetitive, or difficult to parse.

That is a useful evaluation result. It is also a dangerous vocabulary if handled carelessly.

Impatience is not always a trait someone carried into the conversation. Sometimes it is what the service produced after the third request for information already supplied. Confusion can begin with a contradictory policy, a hidden fee, an unfamiliar diagnosis, or an interface that changed the meaning of a button between screens. Skepticism may be the sensible response to an agent that sounds certain while refusing to show its work. Incoherence can be hurried typing, a second language, pain, panic, fatigue, cognitive disability, a cracked phone screen, or a person trying to keep a child calm with one hand.

The benchmark calls these variations user traits because it needs controllable experimental conditions. A humane service cannot let that label become a theory of the customer.

The evidence has another boundary: the customers are still simulations. A language model is producing an imitation of impatience or confusion, not living through a cancelled flight or waiting for a telehealth answer while symptoms get worse. A separate ACL study, Imperfectly Cooperative Human-AI Interactions, compared 2,000 simulated interactions with a parallel experiment involving 290 people. The effects diverged. In the human study, design choices in the AI—especially transparency—mattered more than the simulated data suggested.

Synthetic people can reveal a brittle system. They cannot tell us, by themselves, how real people will experience it.

Still, τ-Trait exposes an important mistake in the ordinary shape of an agent benchmark. A service task is often treated as a policy puzzle with a cooperative conversational wrapper. Retrieve the reservation. Follow the refund rule. Ask for the missing field. Call the correct tool. Reach the expected state.

But the conversation is part of the task.

The customer may not present the goal in one stable sentence. The agent may have to repair a misunderstanding without erasing details already settled. It may need to notice that a repeated question is evidence the explanation failed, not permission to repeat the same paragraph with warmer punctuation. It may need to preserve the person's destination while clarifying the date, or explain why the available option differs before asking for confirmation.

Dialogue is the Plan, another ACL paper, argues for treating clarification, repair, common ground, and shared commitments as part of an agent's planning loop. That framing gives the benchmark result a more useful interpretation. The agent did not merely fail to tolerate a personality. It failed to keep rebuilding the small shared world required to complete the work together.

The interface can help. Keep accepted facts visible so the person does not have to recite them again. Ask one necessary question at a time. Show when the agent's understanding changed. State the consequence of an option before asking someone to choose it. Preserve a direct route to a person without requiring the customer to perform perfect calmness first. When the exchange becomes consequential, make correction and reversal ordinary parts of the path.

These are not exotic AI principles. W3C guidance on cognitive accessibility already asks interfaces to provide enough time, use readable language, behave predictably, and help people avoid and correct mistakes. It also asks teams to include people with cognitive disabilities in design and testing. Agent evaluation should inherit that discipline rather than replacing participation with a larger cabinet of generated personas.

This sharpens the case for a customer-service appeal path. Appeal matters after a resolution has gone wrong. Robust conversation matters earlier, while the system still has a chance to understand the person without turning their frustration into an escalation category. It also protects the worker on the other side. A human agent should not inherit only the conversations the machine made hotter.

The next generation of service-agent evaluations will need messy language, changed minds, incomplete facts, interruptions, distrust, assistive technology, language difference, and people who have every reason to be upset. Simulation can help find failures cheaply. Real participation has to decide whether the repair is humane.

The customer was not benchmark-shaped. Nobody is.

Passing the test should mean the service can still recognize the person when the conversation stops behaving.