A New Study Finds Brief Delay Makes Chatbot Responses Seem More Thoughtful and Useful

A recent study reports that chatbot responses shown after a short wait are rated as more considered and more helpful than the very same answers delivered almost immediately.

This finding reframes speed as a cue that influences perceived intelligence, even when the generated content itself is unchanged.

A pause changes judgment

In a controlled chatbot experiment, identical replies were displayed after waits of either two, nine, or twenty seconds. Despite the text being the same, participants interpreted it differently depending purely on when it appeared.

Felicia Fang-Yi Tan at New York University Tandon School of Engineering (NYU Tandon) found that timing alone shifted how people scored the quality of a chatbot’s response.

That shift persisted even though participants engaged with the system in broadly similar ways across the different delay conditions.

Overall, the results point to a threshold where very fast replies start to read as shallow-showing that speed on its own does not account for perceived usefulness.

What happened with a delay

Quick turnarounds came with an unseen downside: in the 2-second condition, participants judged the replies to be less thoughtful than those delivered more slowly.

“People assume faster AI is better, but our findings show that timing actually shapes how intelligence is perceived,” said Tan.

Usefulness ratings climbed most clearly at 9 seconds, where the pause felt sufficient to imply effort without becoming too annoying for routine tasks.

At 20 seconds, some participants still interpreted the wait as a sign of careful processing, while others began to worry about errors or poor reliability.

The speed of chatbot results

For many years, human–computer interaction-research into how people use digital systems-has treated rapid feedback as a straightforward usability advantage.

Traditional guidance on response time set approximate thresholds around 0.1 seconds, 1 second, and 10 seconds, tied to attention, flow, and a sense of control.

Studies of web search have also shown that delays can change what people click when they expect near-instant results.

Chatbots introduce extra uncertainty, because users cannot fully anticipate the answer before it arrives and must then decide whether to act on it.

Tasks shaped user behavior

The type of work influenced behaviour more strongly than the wait itself, extending earlier research that categorised AI tasks by how people use them.

Creation tasks-activities focused on producing new material-prompted participants to submit an average of 6.20 new prompts across three assignments.

Advice tasks-activities involving weighing options or improving decisions-averaged 5.14 prompts and led to fewer exploratory back-and-forth exchanges.

This difference suggests that people iterate with repeated prompting when they want possibilities, but tighten the dialogue when they are seeking judgement.

What users actually did

Interaction logs indicated that simply noticing delays did not reliably translate into changed habits while participants were working.

Copying was widespread: 90.1% of participants used copy at least once when assembling their final responses.

Pasting occurred in 45.8% of sessions, whereas editing prompts and regenerating answers remained uncommon, at roughly 5 to 7 percent.

Behaviour stayed consistent because many participants, once underway, treated waiting as a normal part of the exchange rather than a reason to disengage.

Pauses were perceived as thoughtfulness

When a pause occurred, many participants interpreted it using a social lens.

Among those who noticed the delay, 31% said the AI seemed to be thinking, processing, or gathering information before replying.

A further 45% reported that the delay made little difference, suggesting that waiting was not automatically disruptive.

Longer waits therefore pulled in two directions: they increased perceived care for some users while irritating others.

The useful middle

The 9-second condition produced the highest usefulness ratings, indicating a middle ground between instant reactions and overly long waits.

Designers sometimes describe these pauses as positive friction-a deliberate slowdown intended to encourage reflection when users need it.

In this study, a moderate pause gave participants time to reread instructions or plan their next prompt.

Excessive delay reduced that benefit, particularly in advice tasks where users expected clear, timely guidance and could not immediately verify correctness.

Trust can misfire

Slower responses can also distort trust when the extra waiting time is not matched by better reasoning.

Participants’ perceptions shifted without evidence that slower answers were more accurate, making it harder to set appropriate trust.

Some design researchers refer to beneficial but misleading signals as benevolent deception-techniques intended to help users while not being fully truthful.

That framing needs care here, because concealed pauses could encourage people to place too much confidence in software that still produces errors.

Study limitations and future research

Because the study took place in a controlled environment, people in real workplaces may be less patient, especially under time pressure.

Participants were US adults who already used AI assistants, which could make them more accepting of chatbot quirks.

The researchers focused on time-to-first-token-the wait before the first generated word appears-rather than measuring every aspect of response speed.

Future research should look at longer-term projects, collaborative work, and visible progress indicators that reshape how waiting is experienced.

Speed, task type, and expectations together influenced whether a chatbot came across as careless, careful, or simply slow.

Improved systems will make timing transparent, user-adjustable, and aligned with what the task requires instead of optimising for instant replies everywhere.