Organisations have decades of practice identifying and countering groupthink. Devil's advocates. Red teams. Pre-mortems. Structured dissent protocols. The premise behind all of them is the same: when a group's thinking converges too quickly, challenge it before the decision hardens.
These mechanisms work by detecting the social signals of premature convergence — a dominant voice going unchallenged, a room that stops producing genuine objections, a process that moves from divergence to conclusion too fast. They are designed to disrupt human social dynamics that suppress dissent.
They are entirely blind to what is now the most pervasive source of convergence in most organisations.
What invisible groupthink is
When teams simultaneously consult the same AI systems — querying the same foundation models, using the same enterprise AI tools, summarising the same documents through the same algorithms — they produce convergent thinking without any of the social friction that conventional groupthink detection relies on. Nobody is being dominant. No one is being silenced. The process feels open and analytical. And the outputs are systematically more similar than the team's independent thinking would have been.
A May 2026 paper from SSRN named this mechanism explicitly: Invisible Groupthink. AI-mediated conformity that operates without directive leadership and without social pressure. The safeguards designed to catch groupthink look for a person causing the convergence. Invisible groupthink has no such person.
When teams simultaneously consult homogeneous AI systems, they unknowingly undergo a process of AI-mediated conformity that is structurally invisible to the organisational safeguards designed to prevent it.
Why AI homogenises thinking
The mechanism is not subtle. Large language models are trained on overlapping corpora. When a team members queries a model with a strategic question, the model's response is shaped by the distribution of positions in its training data — not by the full range of perspectives relevant to this organisation in this context. Ask the same question to the same model through different interfaces, different phrasings, different team members — and the outputs will converge on what was statistically central in the training distribution.
This is not a failure of AI. It is what AI is designed to do: produce the most plausible continuation of the inputs given what it knows. The problem is that 'most plausible given training data' and 'most useful for your organisation's specific strategic challenge' are not the same thing. Often they are inversely related. The thinking that matters most in complex situations is frequently the thinking that diverges from the consensus — the peripheral perspective, the dissenting frame, the challenge that has not yet been expressed anywhere in any training corpus.
The epistemic diversity that disappears
Epistemic diversity — the range of ways of knowing that a group brings to a question — is not the same as demographic diversity. It is the diversity of epistemic grounds from which contributions come: expert knowledge, lived experience, speculative inference, institutional memory, challenge from outside the dominant frame. This diversity is the source of the unexpected connections, the productive anomalies, and the stress-tested conclusions that genuine deliberation produces.
When teams reach for AI at the point where they would previously have needed to consult each other, they are replacing the most epistemically diverse part of their process — the part where different ways of knowing collide — with the part that is most homogeneous. The thing that looked like a productivity gain is simultaneously a diversity loss.
What organisations are not measuring
Most organisations that have adopted AI tools have measured time saved, output volume, and perhaps quality of expressed conclusions. Very few have measured what happened to the diversity of their thinking in the same period. The metrics that would reveal invisible groupthink — epistemic range, quality of challenge, productive disagreement rate — are not in most organisations' reporting. What is invisible to the safeguards is also invisible to the dashboards.
The Forrester research cited in our earlier analysis of collective thinking found that only 6% of organisations are genuine transformation leaders. One of the consistent markers of that 6% is not that they use more AI — it is that they maintain structurally plural decision-making processes where AI is one input, not the frame.
How collective sensemaking operates differently
The protection against invisible groupthink is not avoiding AI. It is ensuring that the part of the process where epistemic diversity is built — where the full range of ways of knowing is surfaced, characterised, and connected — happens before and alongside AI analysis, not after it.
Hunome does not summarise perspectives into a single output. It builds a living structure in which every contribution carries its epistemic ground — expert fact, lived experience, belief, values, gut feel, speculative inference — and where the connections between perspectives with different grounds are as visible as the perspectives themselves. The Lens can then surface what is happening across the map: where clusters of similar ways of knowing are forming, where productive anomaly sits, where the epistemic diversity of the room is strong and where it is thin.
This is not a rejection of AI's analytical power. It is the layer that ensures AI operates on genuine collective intelligence rather than its simulation.