If everybody is thinking alike, somebody isn’t thinking!*
Human-AI collective intelligence: why most organizations are getting it wrong.
* Quote: Gen. George S. Patton
We recently reviewed our Substack as part of the redesign of the PolymathMind.ai website (coming soon), and in doing so identified four core themes across our posts. We wanted to share these themes as useful positions. The first two, published in May, covered ‘Judgement and Cognition under AI’, and AI’s potential for labor market disruption. This week, we have our third theme, examining collective intelligence
Your organization will become a hybrid cognitive system whether you have explicitly decided this or not. The quality of that system depends on capabilities most organizations have never deliberately developed.
Imagine that every AI deployment creates a small piece of organizational cognition, adds a small piece to the emerging picture. Machine output becomes part of how the company decides, frames, and acts. Most senior leaders have been forced to think hard about the technology side, but few (we believe) have thought as hard about what this implies for the way their organization now thinks. The result: rapid AI adoption, slow human capability development, and a growing gap that determines whether the AI investment pays off.
McKinsey announced in early 2026 that it has ‘a workforce of 65,000 — 40,000 humans and 25,000 AI agents.’ The number sounds extraordinary. But strip away the rhetoric and most of those agents do fairly prosaic things: retrieve documents, summarize reports, draft slides, scan datasets. The work that used to occupy junior consultants. The competitive advantage is not the AI itself, which is increasingly commoditized. It is what the AI is connected to — decades of accumulated institutional knowledge, suddenly searchable and recombineable. Every consulting firm has the same opportunity. The ones that win will be those that work out how the human-AI system actually thinks together.
Research on collective intelligence — from Anita Woolley’s work on group cognitive performance to Thomas Malone’s MIT Center for Collective Intelligence and his Supermind framework — has shown for years that group performance cannot be explained simply by the intelligence of the individuals involved. Some teams consistently outperform others because of social sensitivity, balanced participation, and the ability to integrate diverse perspectives. Intelligence exists at the level of the group. Historically that group intelligence emerged through human interaction. AI changes the architecture, but perhaps not in the ways that we might expect.
In a recent paper, published in the journal Nature Human Behavior, Thomas Malone (and co-authors Abdullah Almaatouq and Michelle Vaccaro)[i] found that on average, AI-human combinations do not outperform the best human-only or AI-only system.
“This was our most surprising finding. Some of the most important and interesting use cases for AI involve a combination of humans and computers. Many people would have assumed the combination would be quite a bit better, but it was statistically significantly worse.” Thomas W. Malone.
This contradicts the comfortable assumption that adding AI automatically improves outcomes. Our PolymathMind position is that the quality of the human capability layer is what determines whether the human-AI combination produces value rather than degrades it. Work by Lebovitz, Lifshitz-Assaf and Levina (2022)[ii] distinguish between two paths of human–AI interaction: engaged augmentation and unengaged augmentation. In engaged augmentation, professionals invest time in AI interrogation practices … intentional acts of relating the AI’s knowledge claim to one’s own. Professionals may overrule the AI, reflectively agree with it, or synthesize both views into a new insight. This reduces uncertainty and can produce a final judgement that integrates both sources.
By contrast, professionals may ignore divergent AI results or accept them without reflection. This looks like human-in-the-loop augmentation, but it is actually closer to functional automation. The human is present in the workflow, but not necessarily as an active epistemic agent. This passivity matters because organizations cannot assume that human involvement automatically guarantees good outcomes.
In April 2026, a multi-institutional team led by Cleotilde Gonzalez at Carnegie Mellon[iii] — with collaborators at MIT, Illinois and Microsoft — published Toward a Science of Human–AI Teaming for Decision Making: A Complementarity Framework. Their argument extends the collective intelligence tradition into the AI era directly. They treat reasoning, memory and attention as core cognitive processes that can be deliberately distributed across humans and AI systems — not by default, but by design. Realizing this potential requires deliberate design, rigorous evaluation, and principled governance.” The complementarity does not arrive on its own. It has to be built.
“The key challenge is no longer whether humans and AI will collaborate, but how to structure this collaboration to achieve true complementarity, conditions under which Human-AI teams outperform either humans or AI-only teams.” Gonzalez et al
In a traditional meeting, the cognitive resources available to a group are limited to the knowledge, memory and attention of the people in the room. When AI enters the system, each participant effectively gains a cognitive node. Person A + AI, Person B + AI, Person C + AI. The cognitive system expands, working memory grows, the space of ideas becomes larger. But this is not equivalent to adding another person. As Andy Clark and David Chalmers[iv] argued in their extended mind work, tools that participate in cognition reshape the system of thought itself.
Our own survey of senior decision-makers (March 2026) showed 90% agreement that future organizational performance will depend on collective intelligence — humans and AI thinking together. 84% agreed human skills need to evolve, not diminish. 82% agreed human judgement remains accountable for outcomes. The strategic intent is clear. But the open comments told a different story: concerns about ‘blind following’ of AI outputs, anxieties about speed of decision-making outrunning the capability to challenge it, the ‘black box’ problem of decisions whose reasoning cannot be interrogated. Leaders accept the destination but the human operating model needed to get there is not yet in place.
Field research reinforces the gap. The Harvard-BCG study (Dell’Acqua et al., 2023)[v] showed AI gave consultants a 12-25% productivity boost on tasks well within its capabilities — but produced 19% more incorrect solutions when used on tasks beyond its capabilities. They called it the jagged technological frontier. Performance depends not just on having AI but on knowing which problems to use it for. That is a human capability question, not a technology question.
The implications
1. Treat collective intelligence as a strategic capability. It is not a woolly metaphor … it should be a real capability. Your competitive position will increasingly depend on how well humans and AI think together inside your organization.
2. Audit where AI is making decisions you have not noticed. Are your teams abdicating decision-making to algorithmic outputs? Is your organization’s AI contributing to or making more consequential decisions than you realize.
3. Build challenge into the system. AI-generated outputs sound coherent regardless of whether they are right. Without deliberate human challenge, group decision-making degrades. The friction is really important to judgement and effective decision-making.
4. Recognize that the bottleneck is human, not technical. Better AI will not automatically produce better organizational thinking. It may produce faster mediocre thinking, faster groupthink, faster strategic drift. The competitive edge sits on the human side.
PolymathMind’s central commercial argument is that organizations are dramatically underinvesting in the human capabilities required to work intelligently with AI. We call those capabilities Power Skills — the framework we introduced in 2023. Three years of subsequent work has tested and refined them. They are the practical answer to the collective intelligence question senior leaders are now confronting.
This is the work PolymathMind does — helping senior leaders close the gap between AI adoption and the human operating model around it. If your organization is wrestling with what to keep meaningfully human, how to protect the routes through which expertise is built, or how to make accountability real rather than decorative, those are conversations we’d welcome.
Where to go deeper in our Substack articles
Your Colleague is an Algorithm (early 2026) — What McKinsey’s 25,000 agents actually represent
Cognitive Leaps and Bounds (Feb 2025) — The Supermind framework and the human side of the cyborg
AI Task Optimisation (11 Mar 2025) — When AI-human collaboration works — and when it doesn’t
Mind the Gap (10 Mar 2026) — What 90% of senior leaders agree about — and what they quietly worry about
Beyond the Noise (24 Sep 2025) — Collective human-AI intelligence as the real direction of travel
[i] Vaccaro, M., Almaatouq, A. & Malone, T. 2024. When combinations of humans and AI are useful: A systematic review and meta-analysis. Nat Hum Behav 8, 2293–2303
[ii] Lebovitz, S., Lifshitz-Assaf, H. and Levina, N., 2022. To engage or not to engage with AI for critical judgments: how professionals deal with opacity when using AI for medical diagnosis. Organization Science, 33(1), pp.126–148.
[iii] Gonzalez C, Donahue K, Goldstein DG, Heidari H, Jalali MS, Schelble B, Singh A, Woolley AW. 2026. Toward a science of human-AI teaming for decision making: A complementarity framework. PNAS Nexus.
[iv] Clark, A. and Chalmers, D., 1998. The extended mind. Analysis, 58(1), pp.7–19.
[v] Dell’Acqua, F., McFowland, E., Mollick, E.R., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F. and Lakhani, K.R., 2023. Navigating the jagged technological frontier: field experimental evidence of the effects of AI on knowledge worker productivity and quality.