Put the Values Where the Machine Can See Them

I've spent my career in innovation, crafting my own blend of human-centered design crossed with business strategy. This year, I moved my research practice from innovation to AI, though the thread is older than that. In 2014, I co-founded a machine learning startup, hand-coded the training sample, and wrote the labeling guidelines that taught a classifier what mattered. The model worked; the buyer thesis didn't. I've been designing for that gap ever since.

Human-centered design brings the human into the innovation process. Software development is research-design-test; so is HCD, but there's a person in each room. We ask people what they want, not what businesses want them to have. We co-design. We let people actually try the thing.

Human-centered AI design does two things at once: it brings AI into each of those rooms, and it aims the whole process at AI. We research it, design it, test it — and we do that with AI itself.

What changes is the asymmetry you design around. In HCD, the gap is meaning: what I intend isn't what you understand, and the fix is context: the right information at the right time, so each person can act well in the role they're in. AI has that gap too, but now the meaning has to be written where the system actually looks. A system adopts the values of whatever feedback is most legible to it.

AI has two more that are particularly sticky. State: context windows, memory, what the system knows now versus an hour ago. And authority: who decides, who verifies, what happens when the system is confident and wrong. But neither is new. Human teams have always had memory limits and authority confusion, and HCD is the discipline of designing around exactly that.

This year, I developed an AI-based report-drafting system for a clinician with all three asymmetries managed at the code and the product level. I elicited her terminology so the drafts came back in her language, not the model's. I replaced the context window with an attributed fact ledger, so the system reasons from settled evidence instead of whatever is still in the window. And it never asserts a fact it cannot source.

I'm making these learnings usable by anyone, whether they're on a research team or running a side hustle alone. You'll see posts from me on applying HCD to AI, with AI, and lessons from the field.