Sean Rintel

Design Principles

1. Treat the medium as a participant in its own right. Communication technologies shape what counts as presence and contribution. The channel is never neutral, so study the social practices that grow up around it.
   In practice

   • AI narration is a social act. When AI produces a summary or an action-item list, it creates an authoritative account that shapes who is held accountable: some contributions get foregrounded, others quietly dropped. Design for visible narration choices, and let participants co-author the AI's account of what happened.
   • AI uptake reproduces who counts as a speaker. ASR and language models trained on majority speech systematically under-transcribe accented, overlapping, or quiet voices. When AI hears some people better than others, the medium decides who counts as a participant. Audit AI uptake with the rigour you already apply to interface accessibility.
2. Embrace asymmetry. People arrive at a meeting from different bodies, devices, situations, and positions of power. Equal participation comes from tailoring affordances to each person's situation.
   In practice

   • AI models the meeting from a single vantage point. Most meeting AI produces one account, as if no one in particular authored it. But a meeting is lived differently by the host, by remote participants, and by those with less power, and one account privileges a single perspective while appearing objective. Design AI that can take a stated perspective, or lay several accounts side by side.
   • Personalization is not individualization. AI that learns your preferences is tracking past behaviour, which reveals little about your structural position in the room. A remote participant differs from a collocated one by role and access rather than taste. Adapt AI to role and context, treating preference history as only a weak signal.
3. Put intent and process ahead of output. Help people think about why before optimising how; speed without purpose just produces faster noise. Goals should persist across sessions as named artifacts, legible to people and AI alike.
   In practice

   • Surface the intent layer before the generation layer. AI that jumps straight to output skips the most valuable step: making intent explicit. Ask users to say why they want something before it generates anything. That articulation is cognitive work worth supporting in its own right, and users who author their intent get more useful outputs and feel more ownership of them.
   • Distinguish AI-assisted articulation from AI-substituted articulation. Helping you find words for what you already intend is a different thing from proposing new goals on your behalf. The first is assistance. The second quietly trains people to accept AI-generated intent. Make assistance the primary mode, and treat any AI-proposed goal as scaffolding to own and revise.
   • Treat goals as persistent artifacts. AI that forgets a meeting's goals when the session ends treats intent as ephemeral, but goals are durable commitments. Keep them across sessions and visible to everyone, so AI can refer back to them over time.
4. Scaffold metacognition and calibrate confidence. Using GenAI well means monitoring your own thinking: knowing when to trust output and when to question it. That is hard, and a design opportunity. Build interfaces that help users calibrate trust and resist treating fluency as authority.
   In practice

   • Fluency gets mistaken for authority. GenAI produces polished output regardless of accuracy, and people read well-written as correct. Counteract that: show where a claim comes from, and mark uncertainty in the interface itself rather than burying it in a disclaimer.
   • Design for productive scepticism. Disclaimers and confidence scores inform consent, but they don't create active doubt. Build moments that make people think, such as asking them to predict an answer before AI shows one, or flagging when its output has changed since last time.
   • A heavy verification burden is itself a design failure. When checking falls entirely on users, it disadvantages those with less expertise or time. Make it tractable: link claims to their sources, and flag which outputs are high-stakes enough to be worth checking at all.
5. Enable thinking by doing. Insight rarely arrives before action. It emerges through making and revising. AI tools should treat externalisation as the work itself: lower the cost of producing rough, provisional things so people can think through artifacts, not only about them. Reversibility matters more than first-try fluency.
   In practice

   • Treat the first AI output as a provocation. "Generate, then edit" treats AI output as 80% done, but seeing a response anchors your thinking to its structure. Have AI open the possibility space first, with alternatives and counterproposals, before narrowing it. Its first job is discovery, and production only after that.
   • Make AI outputs structurally reversible. If AI makes experimentation cheap, make undo and branching cheap too. Design AI interactions as explorations with clear return paths, where committing takes more deliberate effort than exploring.
   • Design AI for dialogue about the task. AI is most useful when it helps you work out what to argue and what to leave out, before it writes anything. Support that conversation up front: ask what the group is trying to do, and what would count as success.
6. Design for meaningful shared attention. A firehose of notifications does not add up to collaboration support. As AI becomes a participant in shared work, interfaces should cultivate sustained attention instead of competing for it.
   In practice

   • Distinguish AI-generated urgency from human-generated urgency. AI notifications create synthetic urgency that can override priorities the group negotiated. A person raising a point exercises judgment. AI flagging drift is just executing code. Design AI that reads the room before intervening, weighing more than its own relevance.
   • The unit of attention is the whole meeting. AI optimised for one person's attention fragments the group's shared focus. Design at the group level: ask what everyone needs to see together.
   • Design for attentional recovery. Most AI features compete for attention. The harder and more useful problem is helping people return to shared focus after a drift or tangent.