Interfaces for Collaborative Intelligence

The dominant AI interface is a private conversation: one person, one model, one scrolling transcript. It is a remarkable interface for solitary work and a strange fit for almost everything else. Teams reason together. They argue, delegate, build shared vocabulary, and maintain a common picture of the problem that no single member holds completely. When each member retreats to a private booth to consult a model, that common picture fragments — five people return with five confident, mutually inconsistent answers, and the team's real work becomes reconciling them.

For the past year we have been prototyping interfaces in which the model is a participant in a shared workspace rather than a private oracle. The setting is deliberately mundane: small teams doing document-heavy work — policy analysis, due diligence, curriculum design. Three patterns have survived contact with real groups.

Shared context, visibly owned

The first pattern is a context pool that the whole group can see and edit: the documents, decisions, and standing assumptions the model works from. Every model contribution cites which pieces of the pool it used. This sounds bureaucratic and turns out to be the opposite — arguments about a conclusion become arguments about a source, which groups already know how to resolve. The pool also gives the group something it rarely had before: an explicit, inspectable record of what it collectively believes.

Reasoning as a shared object

The second pattern is exposing intermediate reasoning in a form the group can point at. Not raw chains of thought — those are noise at group scale — but structured claims with stated support, laid out spatially so that two people can stand in front of the same step and disagree about it. In our trials, groups using pointable reasoning caught model errors at roughly twice the rate of groups reading fluent prose conclusions, for the simple reason that prose conceals joints and structure exposes them.

Disagreement as input

The third pattern treats disagreement between group members as signal rather than friction. When two members mark the same claim in opposite directions, the system surfaces the split instead of averaging it away, and routes the question back with the competing rationales attached. The model's job becomes sharpening the disagreement — what evidence would distinguish the positions? — rather than adjudicating it. Teams reported this as the single feature they would keep if they could keep only one.

A theme runs through all three patterns: the interface succeeds when it strengthens the group's own cognition — its memory, its vocabulary, its capacity to disagree productively — and fails when it substitutes for them. Collaborative intelligence is not a model property. It is a property of the room, and interfaces decide whether the room gets smarter.