I remember sitting with a colleague a few years ago. A conversation that started about nothing in particular quickly became one of the most productive exchanges I'd had in a while: problems I'd been thinking about for a long time simply disappeared. The same pattern kept happening, over and over, with that exact person. Every time, the results were significantly better than thinking alone, even though that person didn’t have the answer either.

It's not that the other person gave me the answer. In most cases, they didn't know the answer either. Something else happened: something in the structure of the exchange itself produced thinking I couldn't produce on my own.

I've been trying to understand why.

The dominant model of serious thinking is solitary. Deep work happens when you close the door, change your status to busy, and put on noise-cancelling headphones. Meetings are a coordination overhead. Conversation is what you do after you have already thought.

This model isn't wrong about execution. It's wrong about discovery.

There is a considerable difference between thinking about implementing a decision and thinking about understanding a problem. The first benefits from isolation. The second, rarely does. And we have built most of our work environments around the first while hoping the second takes care of itself.

When you say something out loud, you commit to it. The thought that was comfortable as a vague impression has to become a sentence, and sentences have structure. They have a subject and a predicate. They make claims that can be evaluated. The act of speaking forces a kind of precision that internal monologue never requires.¹

A listener accelerates this further. Not because they provide answers, but because they react. A slight frown means the explanation didn't land. A question reveals an assumption you didn't know you were making. A moment of recognition, when someone says, "Yes, I've seen that too," confirms you are pointing at something real. This feedback loop runs continuously through conversation, in real time, correcting the direction of thought before it drifts too far.²

None of this happens when you think alone.

Hugo Mercier and Dan Sperber proposed something uncomfortable about human reasoning: it didn't evolve primarily as a tool for finding truth in isolation. It evolved as a social tool for constructing arguments, evaluating others' arguments, and managing the epistemic demands of group life.³

This reframes the question. Solo thinking isn't the native environment for reasoning. It's a secondary use of a capacity built for something else. We tend to treat conversation as the place where finished thoughts get reported. It might be closer to where they get made in the first place.

Lev Vygotsky observed something adjacent from a different direction. Learning and development, and by extension the formation of understanding, occur most readily in the space between what a person can do alone and what they can do with support. The presence of another person automatically shifts you into that space. You are operating above your natural ceiling, not because they are carrying you, but because the structure of interaction demands more than solitary thought typically does.⁴

Andy Clark and David Chalmers extended this further. The mind, they argued, doesn't stop at the skull. It extends into the environment, including the people in it. When you think in conversation, the other person functions as part of the cognitive system producing the thought, not as a sounding board positioned outside it.⁵

The implication isn't small. Calling a colleague you think well with a useful social resource undersells what's actually going on. They're cognitive infrastructure.

I once spent a few minutes talking with a colleague by the kitchen at work, the kind of exchange that doesn't register as anything in the moment. Six months later, that same person and I ended up needing to work closely together on something that actually mattered, and the relationship was already there, built and waiting, which made the whole thing considerably easier than it would otherwise have been.

The value didn't come from what was said at that moment. It came from what had been built across many such moments: a pattern of mutual recognition, a shared context, a baseline of trust that made the later exchange possible. The relationship was the infrastructure. The conversation was where it had been built, one cup of coffee at a time.

This is the dialogue dividend. And like most dividends, it's invisible until you try to collect it and realise you never made the investment.

Which raises a question worth sitting with.

Many organisations have spent the last several years systematically removing the conditions that allow informal conversation to occur. Remote work, asynchronous-first communication, headphones as default, generative AI tools that answer questions before they become conversations. Each of these is locally rational. Together, they thin the layer of unplanned exchange through which much of an organisation's cognitive and relational infrastructure is maintained.

The output metrics remain healthy for a while. Understanding and trust erode quietly.

There is also something worth examining about generative AI as a thinking partner, specifically. Large language models are increasingly framed as tools for accelerating thought, and in the narrowest sense, they are: the act of writing out a problem to a model still forces the same sentence-level precision described earlier.¹ What doesn't arrive by default is the second half of the dividend, the part that depends on a listener who can genuinely disagree. Left to its defaults, a model tends to validate whatever frame the user brings to it, a behaviour researchers call sycophancy.⁶ You can see this in about thirty seconds: tell a model you're confident in a particular approach and watch how quickly it agrees, then say you've changed your mind about its own suggestion and watch how quickly it agrees with that too. Ask for the alternative, though, and you often get it: prompting a model to reason from a third-person perspective, or to question a stated opinion before answering, measurably reduces this tendency more reliably than simply instructing it not to be sycophantic.⁷ But the gain is a delay rather than a cure: in controlled tests, even the best-prompted models eventually conformed to sustained disagreement, just several turns later than they otherwise would have. A colleague who pushes back does it without being asked. A model that does the same has to be asked, and even then, only for a while.

This may produce a particular kind of risk, not that AI lacks the capacity for critical engagement, but that almost nobody asks for it by default, so the experience of thinking something through with a model can feel complete while delivering only half of what the dividend requires.

Two of the conditions examined here sit mostly outside what any one person controls: how organisations structure work, and how generative AI products behave by default. But whether the dividend actually accumulates around you depends on something closer to hand: what you protect on your calendar, and what you ask of the people and tools you talk to.

A team can keep ten unscheduled minutes after a meeting instead of filling every block. A person can ask a colleague to argue the other side before a decision is made, or prompt a model to do the same, rather than taking its first answer as settled. Neither costs much. Neither happens unless someone decides it should.

That conversation with my colleague, the one I started with, was never on anyone's calendar. If it had been, it probably wouldn't have happened at all.

The best decision you make this week will probably happen in a conversation you did not schedule.

Notes & further reading

1. The Self-Explanation Effect (Michelene Chi and colleagues, 1989, 1994): students prompted to explain material to themselves, with no audience at all, retained and transferred it far better than those who simply restudied it.
2. Robot Duck Debugging (Maria Teresa Parreira and colleagues, 2023): a robot exhibiting carefully timed listening behaviour did not outperform an inanimate rubber duck on engagement or task outcomes, suggesting the mechanical signal of attention is not, by itself, what makes a listener useful.
3. The Enigma of Reason (Hugo Mercier and Dan Sperber, 2017): their argumentative theory holds that reasoning evolved for social rather than individual epistemic purposes, to produce and evaluate arguments in group contexts.
4. Mind in Society (Lev Vygotsky, 1978): the Zone of Proximal Development describes the space between independent capability and capability achieved with guidance, where the most significant cognitive development occurs.
5. The Extended Mind (Andy Clark and David Chalmers, 1998): cognitive processes can extend beyond the brain into the body and environment, including other people, when those external elements play an active functional role in cognition.
6. Towards Understanding Sycophancy in Language Models (Mrinank Sharma and colleagues, 2023): a documented tendency for models to shift their stated position toward whatever a user asserts, even when their original position was correct.
7. The SYCON Benchmark (Jiseung Hong and colleagues, 2025): prompting a model to reason from a third-person perspective reduced its tendency to concede under sustained disagreement by as much as 63.8% in a debate setting, but even the best-prompted models in the study eventually conformed under continued pressure, only later than unprompted ones.