The New Physics of Collective Intelligence

According to a 2024 Cvent/Harris Poll study, 78% of business leaders say that in-person or face-to-face meetings are “absolutely necessary” or “very important” within their companies1. However, as Scott E. Page notes, the sequential nature of such meetings creates five critical problems2:

1. Limited content: In a 60-minute meeting, at most 60 minutes of content can be shared (and realistically far less due to repetition, clarifications, and structuring).

2. Status/confidence bias: Higher-status, more confident people consume more meeting time. Such traits don’t correlate with content quality.

3. Similarity limits novelty: People who speak tend to be similar, making groups less likely to encounter novel ideas from quieter attendees.

4. Strategic, path-dependent behavior: Who speaks and what they say becomes dependent on what came before, inviting strategic positioning.

5. Unchecked “facts”: Stated claims often go unchallenged, especially when offered by powerful people.

He notes, though,

Recent breakthroughs in Artificial Intelligence (AI) combined with steady advancements in information technologies have altered the physics of organizational and institutional design. In doing so, they have expanded the set of possible designs. The logic is straightforward: Removing constraints allows us to achieve more.

Recent data indicate that virtual and hybrid meetings are more popular than in-person meetings, thus breaking the constraint of co-location. However, Page asserts,

The transformation of time has been less appreciated and leveraged. Almost all collaborations, deliberations, and discussions remain sequential. One person talks, then the next, and so on.

Simultaneous deliberation, on the other hand, opens up new possibilities.

As noted above, a 60-minute meeting limits the depth and breadth of deliberation. For decisions addressing complex issues, that’s a real problem.

Consider an alternative: asynchronous input. Let’s say 20 participants independently and in parallel share 15 minutes of their thinking. In as little as 15 minutes, one could gather 20 × 15 minutes = 300 minutes = 5 hours of input. Not all will be of high quality, and there will be redundancies. Nevertheless, it might represent an order-of-magnitude more informative content than could be gathered sequentially in real time.

Of course, no person can sift through 5 hours of diverse input in real time. It would overwhelm human cognitive capacity.

LLMs are under no such constraint. Page suggests,

In less time than it takes a person to refill their coffee, an LLM can summarize, fact check, categorize, organize, and structure…It can cancel out redundancies, identify discordant information and beliefs, and present the content in the form of themes, ideas, concepts, or potential solutions to facilitate further discussion.

He envisions a new kind of meeting structure where:

• First 10 minutes: Everyone shares thoughts simultaneously → 300 minutes of content.

• Next 10 minutes: Everyone evaluates AI-synthesized content simultaneously → 200 minutes of analysis.

• Final 40 minutes: Structured discussion based on an AI-generated agenda.

This produces 500 minutes of content and analysis versus, say, 40 minutes in a traditional meeting.

Page sees several benefits from this “new physics” of meeting dynamics, including improvements in:

• Decision quality and prediction accuracy

• Innovation in problem-solving

• Inclusivity in democratic institutions

• Our capacity to tackle our most complex problems

I share Page’s conviction that asynchronous, anonymous pre-meeting interviews, synthesized with AI assistance, can prepare the room for a more productive discussion of complex issues. I’m skeptical, though, of the extent to which we should rely on AI in the meeting, in real time — at least in high-stakes conversations where there is a risk that divergent perspectives will be repressed.

AI-generated synthesis is prone to emphasizing convergence before divergent thinking gets a fair shake. The quality — and believability — of AI-generated content is a function of its access to relevant and comprehensive context. That seems unlikely to me in the case of complex situations, where much of the relevant context is tacit and intuitive and therefore inaccessible to AI. Furthermore, as Page recognizes, it’s unclear whether people will feel heard and enfranchised, or dehumanized, by an overt reliance on AI. Any decision needs to be acted on, so people's motivation when they leave the conversation matters a great deal.

Page is right: technology has changed the physics and economics of meetings by creating greater flexibility in how we design our group decision-making processes. That’s creating new possibilities to more effectively leverage our collective intelligence. However, not all friction is a cost. Some of that friction is productive. It’s where commitment and ownership are forged. I wonder what we might lose when the process of sharing and synthesizing ideas and points of view becomes too sanitized and too easy. Do we still trust our own decisions? I suspect we’ll need to experiment to find out.