Sending each word to the screen as it is produced rather than waiting for the whole answer. It does not make anything faster; it makes the wait visible and tolerable, which for a long response is the difference between a product that feels responsive and one that feels broken. The trade-off is that anything you wanted to check before the user saw it now has to be checked after they have seen it, or not at all.
Technical view
Implemented over server-sent events or websockets, with tokens emitted as generated. It reduces perceived latency and makes time to first token the number users actually feel. Complications include output-side guardrails, which either need the full response, defeating the purpose, or must operate on a delay over partial output; structured output, which is not valid until complete; and error handling once a response has partially arrived. Clients must also handle partial multi-byte characters and cancellation.
Related
Part of AI Terminology Index, section: Controlling and Extending Behaviour at Run Time.