An AI’s Pause Can Change How Its Answer Feels

Imagine receiving the same short answer in two different conversations. In one, it appears almost immediately. In the other, there is a pause before the words arrive. Would you describe the second answer as more considered, or wonder whether the system had stalled?

This is a hypothetical comparison, but the question has a research basis. Response time is part of how a conversation feels. When people work with AI, the experience of waiting can acquire meaning beyond the clock: attention, effort, uncertainty or inconvenience. Those interpretations deserve examination alongside the answer itself.

A pause can become a cue for thoughtfulness

A CHI 2026 paper by Felicia Fang-Yi Tan and colleagues examined this question using an analytic sample of 240 US adult AI users. Participants completed creation or advice tasks with delays of 2, 9 or 20 seconds before the first output appeared. The system buffered GPT-4o responses to impose the assigned timing; it did not create the longer waits by instructing the model to reason more deeply.

Two-second responses were rated less thoughtful than nine- and twenty-second responses. Usefulness was significantly higher at nine seconds than at two; the other pairwise usefulness differences were not significant. Some participants interpreted waiting as deliberation, while long waits could also prompt frustration or reliability concerns. Interaction logs showed no significant latency-related differences in the measured behaviours.

These were controlled tasks and selected timing conditions. The findings concern people’s perceptions and interactions, not proof that a slower answer contained better reasoning.

Experience changes what a pause can mean

Research published on 30 May 2022 by Ulrich Gnewuch and colleagues helps explain why one timing rule may suit people differently. In a laboratory experiment with 202 participants, a chatbot responded instantly or with a delay during a fictional mobile-phone-plan recommendation task.

Delayed replies increased perceived social presence among novice chatbot users but reduced it among experienced users. Social presence describes the sense of a socially present conversational partner; the study also examined intentions to use the chatbot. This was a particular service interaction, conducted before today’s LLM interfaces. It does not establish how every experienced or first-time user will respond.

The useful implication is to investigate expectations directly. A designer should avoid deciding that someone wants a slower, more human-like rhythm simply because another group preferred it.

Explanations have limits too

A CUI 2024 paper by Zhengquan Zhang, Konstantinos Tsiakas and Christina Schneegass tested textual explanations for chatbot timing. Following an initial design evaluation, their online experiment involved 194 participants and compared immediate replies with delays that varied according to answer length.

The selected explanations improved perceived transparency and trust in the immediate-response condition, but showed no effect in the dynamic-delay condition. That result gives no universal instruction to add reassuring text whenever an answer is slow. The wording, timing and situation need to be considered together.

Separate the wait, the feeling and the work

For everyday use, try a three-line review. This is a practical suggestion, not a validated psychological assessment. Its purpose is to stop one impression from standing in for several different questions.

• The wait: What did you observe? For example, a blank interval, a visible progress message or words appearing gradually.
• The feeling: How did that timing affect your experience? You might have felt reassured, interrupted, impatient or largely unaffected.
• The work: Did the response meet the task’s requirements? Identify the relevant facts, reasoning, constraints or creative decisions you can actually examine.

Keep those lines separate even when all three are positive. You can enjoy a conversational rhythm and still ask for a missing explanation. You can dislike a delay and still recognise that the completed answer is useful. Neither reaction requires pretending the other did not happen.

Notice, too, when you have inferred an invisible process. “The answer took longer” is an observation. “The system must have checked every claim” is a further conclusion that needs evidence about what occurred. The visible delay alone cannot supply that evidence.

A fictional cover-design discussion

Suppose an independent author asks an assistant to compare two proposed book covers. Her stated priority is whether the title remains readable at a small display size. After a pause, the response offers an elegant discussion of mood and colour, but never addresses that requirement.

She may find the exchange attentive and still need a better comparison. A useful follow-up would identify the omission: “Please assess the title’s small-size readability in each supplied design, explain what you can observe and flag anything you cannot determine from these images.” Any observation the assistant offers still needs checking against the designs.

This imagined example demonstrates a review method, not the frequency of an AI failure. The author can retain a preferred conversational style while making approval depend on the question she actually asked.

Design the waiting experience honestly

For product builders, the practical recommendation is to make the system’s state understandable. A progress message should describe work the application can substantiate. “Checking sources” needs a real source-checking operation behind it. A decorative delay should not be presented as evidence that verification occurred.

Where an application supports them, make cancellation, retrying and returning to unfinished work understandable. Explain any consequences before the user acts. These are proposed design commitments to test with intended users, rather than a claim that every existing AI product provides those controls.

The same concern for understandable choices appears in MaryChuks’ guide to AI voice preference and continuity. That article proposes previews, accessible changes and honest explanations when an available voice changes. Timing deserves similar attention to what the person experiences and what the interface promises.

When reviewing a prototype, ask people what they believe happened during the wait. Then compare their interpretation with the actual process. If they think a loading animation means a claim was independently verified, the design has created an expectation the team needs to examine.

Choose a rhythm without outsourcing judgement

No universal timer follows from these studies. A useful conversation can feel brisk or measured, and a person’s preference can depend on what they are doing. Ask about that preference rather than assigning a personality or level of expertise from a reaction to one delay.

If you want a moment to reflect, you can also create your own review pause after the answer arrives: restate the task, identify the strongest reason offered and note what remains unresolved. This places the reflective step under your control.

The research makes timing an interesting part of human–AI interaction. Treat it as part of the experience you can discuss and design. Give the completed answer its own assessment, grounded in the work you needed it to do.

Sources

• Tan and colleagues, The Impact of Response Latency and Task Type on Human-LLM Interaction and Perception, CHI 2026, April 2026. Author manuscript.
• Gnewuch and colleagues, Opposing Effects of Response Time in Human–Chatbot Interaction: The Moderating Role of Prior Experience, Business & Information Systems Engineering, published 30 May 2022.
• Zhang, Tsiakas and Schneegass, Explaining the Wait: How Justifying Chatbot Response Delays Impact User Trust, CUI 2024, July 2024.