Why Faster AI Isn’t Always Better

In the race to make AI models not just reason better but respond faster, latency — the delay before an answer appears — is often treated as a purely technical constraint, something to minimize and move past. But how is this relentless push for speed actually impacting the people using these systems every day?

There is a rich body of work in human-computer interaction linking faster response times to better usability. But AI models are fundamentally different from the deterministic systems that previous research was built on. When you wait for a file to download or a page to load, the outcome is fixed and predictable. AI models are probabilistic — you cannot anticipate the precise response. Their conversational interface means users naturally read human social cues into the interaction. A pause might be read as the AI "thinking," for instance. Users are increasingly asked to choose between faster models and slower, deeper-reasoning ones, without guidance on what that choice actually means for their experience.

A recent study presented at CHI’26 explored how response timing shapes the way people use and evaluate AI systems. Felicia Fang-Yi Tan and Technology Management and Innovation Professor Oded Nov recruited 240 participants and asked them to complete common knowledge work tasks using a chatbot. Some tasks focused on creation, such as brainstorming ideas or drafting text. Others centered on advice, like evaluating decisions or offering recommendations. Crucially, the system was engineered to respond at different speeds. Some participants received answers after just two seconds, while others waited nine or even twenty seconds.

The results challenge a long standing assumption in human-computer interaction that faster is always better.

“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.”

Surprisingly, how quickly the AI responded did not significantly change how people behaved (e.g., frequency of prompting, copy-pasting). Participants prompted just as much and interacted with the system in broadly similar ways regardless of whether they waited two seconds or twenty. Instead, behavior depended more on the type of task. Participants attempting creation tasks (which involve producing new content such as writing) prompted more back and forth, with users refining and iterating on ideas. Advice tasks (which involve providing guidance, critique, or evaluation) led to fewer, more focused exchanges.

Where timing did matter was in perception.

Participants who received two-second responses consistently rated the AI’s answers as less thoughtful and less useful. In contrast, those who experienced longer delays tended to view the same kinds of responses more favorably. Many interpreted the pause as a sign that the system was “thinking,” attributing greater care and deliberation to its output.

This effect highlights a subtle but powerful feature of human psychology. In everyday conversation, pauses carry meaning. A quick reply can feel impulsive, while a measured delay suggests reflection. People appear to apply these same social expectations to machines, even when they know they are interacting with software.

The implications extend beyond user experience. Given that latency is an inherent feature of today's AI models, perhaps the more productive question is not how to eliminate it, but what it can be designed to do. Positive friction refers to intentional slowdowns designed to promote cognitive benefits such as reflection. Rather than treating every millisecond of waiting as waste, designers might ask: what can this pause do?

The study also surfaces important ethical considerations. If people equate longer response times with higher quality, they may place undue trust in slower systems, regardless of whether the output is actually better. This raises ethical questions about whether AI systems should be designed to manage timing in ways that shape user perception. And if so, whether users should be informed when they are.