The Chair We Offered the Machine

The question circulating ahead of a gathering I’m attending this week is a good one: what happens when AI becomes another voice in the room? The range of possible answers is sensible: Let the AI speak only when invited. Disclose your system prompts. Use an enterprise license, so your words don’t become training data. Have the AI play devil’s advocate, but not too often. Give it a name and face, or don’t.

Every one of these is an answer to the question of how. Not one of them is an answer to what for.

When considering meeting design, I think in terms of a sequence: Purpose → People → Process. What is the meeting for? Given that purpose, who can contribute to its achievement? Given that purpose and those people, what supporting process is appropriate?1

The order is important. Skip the first question, and the third becomes unanswerable because there are no criteria by which any process could be judged better or worse. I wonder whether that’s the state of the debate about AI in meetings. It is often a Process argument in the absence of Purpose.

I can think of at least three general purposes for a meeting, and the AI question resolves differently for each.

Convergence. Here, our goal is to coordinate, inform, and ratify something that is already substantively settled. In this instance, friction is costly. AI summarization, transcription, and action-item extraction serve this purpose. The AI-as-efficiency-tool is correctly specified for these meetings. That’s why it sells.

Generative. In these meetings, we find out what we actually think, especially when we disagree and haven’t said so. Here, friction can be productive. Summarization is majority-weighted by construction. The modal view survives, and the lone objection is, by definition, what gets cut for length. A recording is not a notetaker. Notes are lossy, paraphrased, and deniable; a verbatim transcript is attributable, searchable, permanent, and available to people who weren’t in the room. Candor has always depended on some measure of forgetting. The technology whose business model is remembering is a strange player here.

Constitutive. By constitutive, I mean that the conversation alters the participants in some way, and that alteration is the purpose. David Gurteen’s Knowledge Café is an example: the conversation is unrecorded, and there is no note-taking or reports. The outcome is what each person carries away in their own head.2 No artifact is produced, which (I think) means there is nothing for AI to help produce.

I suspect much of our anxiety comes from tools built for convergence being carried into meetings to be generative or constitutive. We’re not necessarily wrong about the tool; we’re wrong about its application to a particular meeting.

The debate can be framed as a question of how this new participant should be admitted to the conversation, which places the AI in the People category before anyone has tested whether it should be offered a seat at the table. For generative purposes, a contributor must be able to hold a view, bear some cost for holding it, and be moved off it by a better one.

A facilitator or advisor also lacks a personal stake but is accountable. They can be questioned, can revise their position under pressure, and carry reputational risk. Those accountability loops are what the criteria are really pointing at.

An AI meets none of these conditions and closes none of those loops. It has no stake, no standing, and nothing at risk for being wrong. Furthermore, it can’t transmit conviction.

The research is at least consistent with this, though the authors read their own results more optimistically than I do. Soohwan Lee and colleagues at UNIST put an LLM devil’s advocate into twelve four-person groups deliberately stacked three “seniors” to one “junior.”3 Satisfaction with the conversation rose when the AI was present. But the content of the counterarguments moved little. What changed was that the junior was no longer the only person in the room holding a different view. The AI didn’t voice her dissent. It gave her cover to voice it herself.

The senior–junior gap in felt psychological safety, meanwhile, stayed large in both conditions. The AI made people happier about a conversation whose power structure it left intact.

Lee’s group went on to test the obvious next move: an AI that takes the minority member’s actual view, paraphrases it, and posts it as its own. Anonymity by proxy. It increased participation but reduced psychological safety and satisfaction.4 Speaking through the machine got the words into the room and left the person outside it. Anonymity and authenticity turn out not to be the same thing.

Charlan Nemeth’s work would predict something like this. Dissent that no one actually holds doesn’t carry — whether the AI invents it or relays it. In the worst case, a devil’s advocate can provoke a group into bolstering the very view dissent was meant to unsettle.5

So, the AI is not People; it’s Process. More precisely, it’s infrastructure. Once acknowledged, discussions of etiquette — addressing the obsequious flattery and concerns over domineering interjections — fall away. We don’t need to socialize a technology into a role that it’s structurally unqualified to fill…for now.

I think when we rephrase the question, the design challenge becomes tractable. What supporting process, in service of this purpose, with these people, is useful? For conversations with a generative purpose, the honest answer regarding AI is not in the room, and mostly not during. Here’s where I think it can help:

• Surface dissent and tensions before the meeting. Private, individual reflection gathered in isolation away from the group’s gravitational pull toward consensus. Weak signals are preserved rather than averaged away.

• Give cover, not content. An AI that names a tension during a conversation makes it cheaper for the human who holds it to claim it aloud.

• Prepare the facilitator. This pre-meeting intelligence can include tension maps, fault lines, and minority views that can help a facilitator shape the conversation.

• Force a considered discussion. David Creelman recently described wanting an LLM that won’t budge until you’ve addressed its criticisms one by one. It’s a forcing function that makes you do the thinking rather than doing it for you.6

In my essay, No Villain in the Room, I wrote how we blame silence on power differentials, because that gives us someone to confront. More often than not, we self-censor for reasons that have nothing to do with malice or power.

Mo Houtti and colleagues found something similar occurring in their research. They designed a “virtual co-host” tasked with improving engagement across a group by, for example, sending gentle, private reminders to over-participators. The recipients of messages telling them they were dominating rationalized away the messages. One over-participator, told he was taking too much air, concluded he had been perfectly fair — while an under-participator in the same meeting sat quietly frustrated with him.

The co-host was gentle and private because that’s what the users asked for in the design sessions. They wanted an intervention that wouldn’t make anyone squirm — and squirming was the active ingredient. The machine didn’t bring a preference for comfort into the room. It found one there and honored it.

The authors’ own explanation is the one that matters: the AI had too little social pressure. A machine can tell you that you’re crowding the room. It cannot make you feel uncomfortable about doing so.

The meetings didn’t improve, but participants rated them higher anyway.7 We risk allowing AI interventions to change our perceptions without changing the underlying substance.

Real meetings are mixed. A single ninety-minute meeting can be convergent for one participant, generative for another, and constitutive for the meeting’s sponsor. My tripartite classification system is too neat and tidy. The really interesting failure mode is a meeting that silently disagrees about what it’s for.

The classification is convenient for me. I’ve been building a before-the-meeting process. So, of course, I conclude that key generative work happens before the conversation. Discount my analysis accordingly.

Cover is not conviction. If the AI’s real contribution is lowering the cost of claiming a dissenting view, that’s permission. In Conviction by Proxy, I speculated that dissent, without a person standing behind it, goes inert. Permission is necessary but insufficient.

One other thing: moving AI out of the People category doesn’t make it neutral. It makes it infrastructure, and infrastructure is where unexamined choices go to hide.

Should the AI speak? feels like a question about manners. For me, the more pointed questions are: What is this meeting for? Would we notice if AI were serving a purpose different from ours? I suspect that much (most?) of the time we wouldn’t. That’s less of a statement about the technology and more about how rarely we make the purpose of our conversations explicit.