Designing AI for people in groups

Almost everything difficult that humans accomplish, they accomplish in groups — and almost every AI product ships as if its user were alone. The mismatch is not cosmetic. A tool that makes each individual faster can still make the group dumber: private assistants fragment shared context, fluent summaries hide the disagreements a team needed to have, and the social fabric of checking one another's reasoning quietly goes slack.

Five principles currently guide our group-facing work. They are working hypotheses, written down so they can be wrong in public.

• The group's shared picture is the primary asset. Any system that improves individual output while degrading shared context is a net loss; design for the room, not the seat.
• Provenance over fluency. In group settings, where a claim came from matters more than how well it reads. Every contribution should carry its sources on its face.
• Surface disagreement; never average it. Divergence between members is the group's most valuable signal. A system that smooths it away is destroying information.
• Strengthen the practices that make groups smart. Articulating reasons, checking one another, building shared vocabulary — assistance should exercise these muscles, not replace them.
• The group owns its memory. Shared context built by a team is the team's, portable and inspectable — not a side effect locked inside someone else's product.

The larger conviction behind the principles: collective intelligence is the original augmentation. Language, writing, and institutions were technologies for thinking together long before anyone thought alone with a machine. AI that joins that lineage — that makes the room smarter — is the kind worth building.