Summary: Latency is usually viewed as a performance problem in technology. A provocative new study, however, suggests that faster responses from AI can sometimes reduce users’ perceived thoughtfulness and usefulness of those systems.
Researchers tested 240 participants with chatbot responses intentionally delayed between 2 and 20 seconds. Even when the content was identical, participants consistently judged slower responses as more thoughtful and more useful. The findings indicate people apply social cues — interpreting pauses as signs of deliberation — to probabilistic AI systems.
Key Facts
- The Perception Gap: Participants who waited 9 or 20 seconds for answers rated the AI more favorably than those who received near-instant (2-second) replies, despite identical outputs.
- Anthropomorphizing Latency: Users read human conversational norms into AI interactions: a quick reply can seem hasty, while a pause suggests the system is “thinking” and caring about the answer.
- Behavior vs. Perception: Response speed did not significantly affect how people used the system — prompting frequency and interaction patterns stayed similar — but it did change subjective judgments about the AI’s intelligence and usefulness.
- Task-Driven Interaction: The nature of the task mattered more for behavior than timing. Creation tasks (brainstorming, drafting) encouraged iterative exchanges, while advice tasks (evaluation, recommendations) produced shorter, focused interactions.
Source: NYU
Background: Efforts to reduce latency in AI—making models respond faster—are common because earlier human-computer interaction research linked speed with better usability. But AI systems differ from deterministic applications like file downloads: AI outputs are probabilistic and conversational, which leads users to apply social expectations to machine behavior.
Because AI responses are not predetermined, users naturally treat conversational latency as meaningful. A delay may be interpreted not as a technical hiccup but as a sign the system is weighing options — the same cue people use when judging another person’s reasoning.
At CHI’26, Felicia Fang-Yi Tan and Professor Oded Nov from NYU explored how response timing shapes both the use of and attitudes toward AI assistants. They recruited 240 participants to perform everyday knowledge-work tasks with a chatbot. Tasks fell into two categories: creation (generating ideas, drafting text) and advice (evaluating choices, recommending actions). The chatbot was set to respond after 2, 9, or 20 seconds for different participants.
The results challenge the commonly held belief that faster is always better. While speed had little effect on measurable behavior — participants edited, copied, and re-prompted at similar rates across conditions — it strongly influenced subjective evaluations.
Participants who received the fastest replies judged the AI as less thoughtful and less useful. Those who encountered longer delays tended to rate the same responses more favorably, attributing greater deliberation and care to the system. In short, a brief pause often made identical content feel deeper and more valuable.
This reveals a subtle human tendency: conversational timing conveys intent and competence. In ordinary dialogue, an instant response to a complex question can feel dismissive, while a pause signals consideration. People unconsciously transfer these expectations to AI, even when they understand the interaction is with software rather than a person.
Designers and product teams should consider the broader implications. Latency is often treated as purely negative, but it can function as “positive friction”: an intentional, measured delay that encourages reflection and may boost perceived trust or thoughtfulness. The study suggests the pause itself can be designed to support user goals rather than treated solely as a performance cost.
At the same time, there are ethical concerns. If users equate longer response times with higher quality, designers could exploit this by slowing responses to create a false impression of depth. That raises questions about transparency and whether users should be informed when timing is manipulated for perception rather than reflecting actual processing differences.
Key Questions Answered:
A: The study introduces the idea of “positive friction.” Intentional slowdowns can promote reflection and a sense of trust, but they risk being deceptive if used to make a weaker model appear smarter. Any design that manipulates timing should be considered carefully and transparently.
A: Because AI interactions are conversational and unpredictable, people rely on familiar social cues. In human exchanges a rapid reply to a complex prompt can feel glib, while a pause implies thought; users transfer that interpretation to AI behavior.
A: Not necessarily. The perceived usefulness tied to slower responses is a psychological effect of waiting. For repetitive, time-sensitive tasks, faster models remain advantageous. For complex, collaborative advice work, users may prefer the rhythm of a slower, deliberative system.
Editorial Notes:
- This article was edited by a Neuroscience News editor.
- The journal paper was reviewed in full.
- Staff added contextual information to clarify implications for design and ethics.
About this AI research news
Author: Leah Schmerl
Source: NYU
Contact: Leah Schmerl – NYU
Image: The image is credited to Neuroscience News
Original Research: Findings presented at CHI’26