summary: In the world of technology, latency is usually the enemy. But a provocative new study suggests that the “need for speed” in AI may be backfiring on user perception.
The researchers tested 240 participants with varying response latencies (from 2 seconds to 20 seconds) and found that users consistently rated faster AI as less thoughtful and less useful. This reveals that humans apply social “thinking” cues to probabilistic AI, viewing short pauses not as technical lags, but as evidence of care and deliberation.
important facts
- Perception gap: Participants who waited 9 or 20 seconds for an answer rated the AI higher than those who received an immediate (2 second) answer, even if the actual content was the same.
- Anthropomorphic latency: Users interpreted the AI’s pauses as the machine “thinking” and projecting the norms of human conversation (quick responses appear impulsive, measured delays suggest deliberation) onto the software.
- Behavior and perception: Surprisingly, the speed did not change. how People used AI. Interaction frequency and prompts were consistent regardless of delay. The only thing that changed was the user’s subjective opinion of the AI’s intelligence.
- Task-driven interactions: of type Action was more important than speed of work. creation task (Brainstorming) led to more repeated exchanges, advice task (Evaluation) The result was less exchange and more focus.
sauce: new york university
In the race to not only improve the inference of AI models but also increase their response speed, latency (the delay before an answer appears) is often treated as a purely technical constraint that should be minimized and avoided. But how does this relentless pursuit of speed actually impact the people who use these systems every day?
There are many efforts in human-computer interaction to reduce response time and improve usability. However, AI models are fundamentally different from the deterministic systems that have been the basis of previous research. When waiting for a file to download or a page to load, the results are fixed and predictable.
AI models are probabilistic and cannot predict exact responses. Conversational interfaces allow users to naturally incorporate human social cues into their interactions. For example, a pause could be read as a “think” for the AI. Users are increasingly being asked to choose between faster models and slower, deeper inference models without any guidance on what that choice actually means for their experience.
A recent study presented at CHI’26 investigated how the timing of responses shapes how people use and evaluate AI systems. Felicia Fang-Yi Tan and Oded Nov, professor of technology management and innovation, recruited 240 participants and asked them to complete common knowledge work tasks using a chatbot. Some tasks focused on creation, such as brainstorming ideas and drafting text.
Others focused on advice, such as evaluating decisions and providing recommendations. Importantly, the system is designed to respond at different speeds. Some participants received their answers after just 2 seconds, while others waited 9 or even 20 seconds.
This result challenges the long-held assumption that faster is always better in human-computer interaction.
“People think the faster the AI, the better, but our findings show that timing actually drives how intelligence is perceived,” Tan said. “Even if nothing has changed in the underlying AI model, a brief pause signals attention and consideration, making the same response feel more thoughtful and helpful.”
Surprisingly, the speed of the AI’s responses did not significantly change people’s behavior (prompt frequency, copy-and-paste, etc.). Participants interacted with the system in much the same way, prompting and interacting with the system whether they waited 2 or 20 seconds. Rather, behavior depends more on the type of task.
As participants attempted creation tasks (including creating new content such as writing), users were further encouraged to refine and iterate on their ideas. Advising tasks (including providing guidance, critique, and evaluation) were less interactive and more focused.
It was a recognition that timing was important. Participants who received answers within two seconds consistently rated the AI’s answers as less thoughtful and less helpful. In contrast, those who experienced longer delays tended to view the same type of response more favorably. Many interpreted the stoppage as a sign that the system was “thinking” and assumed that great care and thought had been put into its output.
This effect highlights a subtle but powerful feature of human psychology. In everyday conversation, pauses have meaning. A quick reply may feel impulsive, but a long delay suggests remorse. People seem to apply similar social expectations to machines, even when they know they are interacting with software.
Its impact extends beyond the user experience. Given that delay is an inherent feature of today’s AI models, perhaps a more productive question is not how to eliminate it, but how it can be designed into.
Positive friction refers to intentional slowing down designed to promote cognitive benefits such as self-reflection. Instead of treating milliseconds of waiting as waste, designers might ask, “What can I do with this pause?”
This study also highlighted important ethical considerations. People who believe that longer response times mean better quality may place too much trust in a slow system, regardless of whether the output is actually better.
This raises ethical questions about whether AI systems should be designed to manage timing in a way that shapes users’ perceptions. And, if so, whether users should be notified of this.
Answers to key questions:
answer: We are introducing the concept of “Positive friction.” Speed is great for increasing efficiency, but intentionally slowing down promotes reflection and trust. However, research warns that this can be used deceptively and create substandard models. seems to be Just add a “thinking” delay to make it smarter.
answer: Because AI is conversational and probabilistic, it naturally uses the same social “mental models” we use for humans. In human conversation, a half-second response to a complex question is perceived as negative, but a five-second silence suggests that the person is seriously considering the answer.
answer: Not necessarily. According to research, feeling Decreased usefulness is the psychological effect of waiting. If you need real speed for repetitive tasks, faster models are better. If you’re looking for a “partner” for a complex advice task, your brain may naturally prefer the rhythms of a slower, reasoning-centered system.
Editorial note:
- This article was edited by the editors of Neuroscience News.
- Journal articles were reviewed in full text.
- Additional context added by staff.
About this AI research news
author: Leah Schmel
sauce: new york university
contact: Lia Schmerle – New York University
image: Image credited to Neuroscience News
Original research: Survey results were presented at CHI’26
