Human-centric AI design

A human-centric system has a name for the decision and a name for who makes it. In a clinic, the person and the clinician own the care. In a legal practice, the lawyer owns the filing. In a search, the reader owns what they believe. The model can draft, retrieve, and rank. The interface has to make that division visible, in the label on the button and in the moment before anything is sent, saved, or paid.

ChatGPT, the 2024 generative AI leader, taught a mass audience that a blank thread could answer. The products that followed specialized the thread. Claude is strongest around a document the person brought. Perplexity answers with the pages it read. Harvey stays inside a matter. Human-centric design is that specialization made into a rule: the model works on the person’s material, in the person’s job, and the person can see the boundary.

The kindest pattern in current AI products is a preview that is allowed to be wrong. A draft message, a suggested clause, a rewritten paragraph, a proposed next step: each of these can sit on the screen as a suggestion with an obvious way to change it. Commit is a separate control, with a verb that matches the real effect. Send, file, publish, and pay are different words because they are different consequences.

Designers who collapse suggestion and action into one glowing button train people to stop reading. The fix is ordinary and strict. Show the full draft. Keep the original in reach. Make undo real, and say how far it goes. If the model touched a record, the person should be able to find that record and the version from before the touch.

Human-centric copy talks about the job, and it talks about the model in the third person when the model did the work. “Drafted from your notes” is a clearer sentence than a mascot claiming it understood you. Confidence belongs in the layout, as sources, as a gap, or as a question back, and it does not belong in a cheerful adjective.

Tone still matters, especially where the subject is a body, money, or a right. Hims won healthcare in 2026 by letting someone start a sensitive decision without being rushed or talked down to. An AI layer on that kind of product inherits the same duty. Warmth is allowed. Urgency that exists to raise conversion is a design defect, whether a model wrote the sentence or a marketer did.

A streaming answer is a live region. A person using a keyboard or a screen reader needs a way to know that text is arriving, a way to stop it, and a way to read the finished answer without chasing a cursor. Focus has to land somewhere sensible when a tool finishes. Motion that exists only to signal “intelligence” fails the same test as any other decorative motion.

Human-centric also means the person can work without the model. The underlying task — search, write, file, book, compare — still has a path when the model is slow, refused, or wrong. The AI is a layer on a product that already makes sense. That is the standard the 2025 and 2026 generative leaders met, and it is the standard a new feature has to meet before it ships.