Faster answers are not always better answers. In several studies of assisted learning and decision support, we found that response latency shaped trust more than marginal gains in accuracy — within a surprisingly wide band.
When systems answered instantly, participants accepted outputs they later could not defend. When systems took a visible beat — not a spinner theater, but a paced reveal tied to the structure of the answer — people checked more carefully and trusted the results more after they had checked.
Slowing down on purpose
This is not an argument for artificial delay as decoration. Delay without structure feels broken. What helped was latency that mapped to the work the person still needed to do: reading a claim, seeing its support, deciding whether to adopt it.
Accuracy still mattered. Below a competence threshold, no amount of pacing repaired trust. Above it, shaving another hundred milliseconds off the first token often reduced careful use.
Design implication
If your product’s job is human growth, optimize for the pace of understanding, not the pace of generation. That may mean refusing to stream a full answer before the person has a place to put it. It will look slower in a benchmark. It holds up better in a life.