Why a Slower AI Might Actually Feel Smarter to You

Type a question into an AI chatbot and hit send. Now watch the cursor blink. Two seconds feels fine, barely noticeable. Nine seconds and something shifts: you start to wonder if the system is really working through the problem. By the time you hit twenty seconds you have either concluded the AI is doing something impressively deep, or you are quietly composing a complaint about internet connectivity. What you probably have not considered is that all three scenarios could produce the exact same answer, and that your judgment of its quality will differ substantially depending on how long you waited. That is the rather unsettling conclusion of a new controlled experiment out of NYU Tandon School of Engineering, presented this month at CHI 2026 in Barcelona.

Felicia Fang-Yi Tan and professor Oded Nov recruited 240 participants and asked them to complete realistic knowledge-work tasks using a chatbot, rigged to respond at precisely two, nine, or twenty seconds. The underlying model was identical across conditions. The responses were the same kind of responses. And yet the people who waited longer consistently judged what they received as more carefully considered.

The Pause That Reads Like Thought

Human-computer interaction research has spent decades establishing that speed is good. Faster file loads, quicker page renders, sub-second responses: the field developed what amounted to a unified theory of “just go faster.” Those rules were calibrated against deterministic systems, where you click something and the outcome is already fixed inside the machine. AI is different. The output is probabilistic; you cannot anticipate it. And crucially, the interface is conversational, which means users bring entirely different expectations to the waiting. In everyday conversation, a brief pause before answering suggests care. A person who fires back instantly can seem glib. We have apparently applied the same social heuristic to software.

“People assume faster AI is better, but our findings show that timing actually shapes how intelligence is perceived,” says Tan. “A short pause can signal care and deliberation, making the same response feel more thoughtful and useful, even when nothing about the underlying AI model has changed.”

The study split participants across two task types as well as three latency conditions. Creation tasks involved producing something new: brainstorming slogans, drafting middle sections of articles, proposing ideas for distracted students. Advice tasks asked for evaluation and guidance: critiquing a decision memo, improving a draft email reply, recommending focus strategies for someone working from home. Both types showed the perceptual effect of latency, but they diverged interestingly on behaviour. Participants working on creative tasks submitted significantly more prompts, more back-and-forth, more iteration. Advice tasks tended toward fewer, more targeted exchanges. This held true regardless of wait times, suggesting the nature of the work shapes how people use AI more than system speed does.

Behavior Holds Steady, Perception Does Not

That behavioral robustness is one of the more striking findings. Whether the chatbot took two seconds or twenty, participants prompted it at roughly the same rate, copy-pasted outputs with similar frequency, and showed no significant tendency to consolidate queries or give up. The wait did not change what people did. It changed what they thought about what they got.

On the perceived thoughtfulness scale, two-second responses rated consistently lower than nine or twenty-second ones. Perceived usefulness peaked at around nine seconds, then dipped slightly at twenty. The researchers describe a “sweet spot” in the moderate range, where delays are long enough to read as deliberation but short enough not to shade into apparent malfunction. Some participants in the longest-wait conditions began to wonder whether their internet had dropped out; a few described mild frustration. Still, overall trust remained high across all conditions, and expectation-gap scores (whether responses met what people anticipated) showed no significant variation. People waited twenty seconds, got a perfectly ordinary AI response, and mostly felt it had been worth it.

This pattern has a name in cognitive psychology: the effort heuristic. People routinely infer quality from perceived effort, a shortcut that works well enough in human contexts (a surgeon who studies for years probably is better than one who didn’t) and transfers messily to AI. A slower response carries no information about the quality of the processing behind it. The model is not, in any meaningful sense, “thinking harder” because you waited nine seconds. The tokens stream out the same way. The impression is entirely a construction of the observer’s social expectations colliding with an opaque black box.

Design Lever or Deceptive Nudge?

The practical implications branch in two directions, and they are somewhat in tension with each other. On one side: perhaps latency should be treated as a design variable rather than pure overhead. There is prior work on “positive friction,” the idea that small, deliberate slowdowns can interrupt automatic thinking and encourage more considered engagement with AI outputs. A moderate wait might prompt users to re-read the task, formulate a cleaner follow-up, or think critically about what they are about to receive. Some participants in the study did exactly that, using the pause to plan their next query. If a nine-second delay improved perceived usefulness while also giving users a moment to think, that could be genuinely beneficial rather than cosmetic.

On the other side sits an obvious ethical discomfort. If AI developers deliberately slow their systems to make outputs seem more thoughtful, that is a form of manipulation, however benign the intent. Users who equate wait time with quality may extend unwarranted trust to slower systems without any corresponding improvement in accuracy or reliability. The researchers flag this explicitly: novice users, or those with less domain knowledge, might be particularly susceptible to over-trusting a system that is simply pacing itself for effect. Without transparency about whether timing has been engineered, users lose the ability to calibrate their own judgment.

There is a version of this that feels acceptable and a version that does not. Displaying a message like “analysing your request…” during a pause, as some platforms already do, nudges users toward the deliberation interpretation while at least acknowledging that something is happening. Engineering a multi-second delay without explanation, purely to shape perception, sits closer to what researchers call “performative deliberation,” simulating cognitive effort to increase apparent competence. Where exactly the line falls is not obvious; the study raises the question without resolving it, which is probably the honest position given how early this territory is.

What the findings do settle is that the old framework needs updating. Latency is not simply cost to minimize. It is a signal that users are already interpreting, whether designers intend it or not. The cursor blinking while you wait is not just waiting. It is information. Knowing that, product teams building on large language models now face a choice about what to do with it.

https://doi.org/10.1145/3772318.3790716

---

Frequently Asked Questions

Does waiting longer for an AI response actually mean you are getting a better answer?

No. In this study, the underlying AI model and the quality of its responses were identical across all conditions. Participants who waited longer simply perceived the same kind of answer as more thoughtful and useful. The effect is entirely perceptual, rooted in social expectations about what deliberation looks like, not in any real difference in output quality.

Why does the type of task change how people use AI chatbots?

Creative tasks like brainstorming or drafting tend to invite iteration: users refine and explore, prompting back and forth until something clicks. Evaluative tasks like getting advice or reviewing a decision tend to wrap up once a satisfactory answer arrives. The research found this behavioral difference held regardless of how fast or slow the AI responded, suggesting the nature of the work shapes interaction patterns more than system speed does.

Is deliberately slowing down an AI to make it seem smarter a form of deception?

That depends on context and transparency. The researchers distinguish between using pause time constructively (for example, displaying a progress message that invites users to think through the problem) and engineering delays purely to inflate perceived quality without disclosure. The latter sits uncomfortably close to what the literature calls “performative deliberation,” mimicking cognitive effort to increase apparent trustworthiness without any real basis.

What is the “sweet spot” for AI response time found in the study?

Around nine seconds appears to be where perceived usefulness peaked, outperforming both the two-second and twenty-second conditions. Very short responses were judged as less thoughtful, while very long waits began to raise reliability concerns in some users. The researchers frame this as a “moderate zone of benefit,” though they caution that the right timing will vary with task type and user expectations.