The Lost Thinking Room — and How AI Is Rebuilding It

For years, the debate about AI-assisted writing has been framed almost as a moral question. If AI helped write something, is the work really yours? Is it less authentic, less original, perhaps even less legitimate?

Sometimes that skepticism is justified. Someone who barely understands a subject can ask an AI system to produce something plausible, make a few edits, and suddenly appear remarkably well informed. We have, in effect, invented expertise at scale. The result can be polished, confident, and completely wrong.

But that is only one way to use AI. In the hands of someone who already understands the domain, AI can function less like a ghostwriter and more like a thinking partner. It can accelerate research, challenge assumptions, reorganize evidence, expose weak logic, generate alternatives, and make many more rounds of iteration practical than would normally be possible.

That does not diminish expertise. It gives expertise leverage. The useful distinction is not between human-written and AI-written. It is between outsourcing thought and accelerating thought.

We used to think together

Long before AI, organizations had another way to develop ideas quickly. We called it going to the office.

People stopped by one another’s desks. They argued at whiteboards. Someone floated an idea, and another person immediately explained why it would never work. A third remembered a customer who had already tried something similar. Someone else connected the discussion to a theory, a prior project, or a problem in another part of the business.

Thirty minutes later, the original idea might be almost unrecognizable—and considerably better.

Good teams did more than divide work. They developed thought collectively. Experience, intuition, data, theory, and dissent collided in real time. Much of this was accidental. Nobody scheduled a meeting called Cross-Functional Serendipitous Idea Refinement, 2:00–2:30. People were simply there.

Then work became distributed. Remote and hybrid work brought enormous advantages. Commutes disappeared. Geographic boundaries weakened. Talent pools expanded. Many forms of individual productivity improved. But something less visible became harder.

We became very good at coordinating work remotely and, in many cases, not quite as good at thinking together remotely.

We recreated the meetings

Video conferencing solved the obvious problem of how people in different places could keep meeting, and we embraced it enthusiastically. In fact, we may have recreated the meeting a little too successfully.

What we did not reproduce nearly as well was everything around the meeting: the five-minute conversation, the overheard problem, the quick challenge, the person who happened to know something you did not yet realize you needed.

Remote collaboration tends to be more intentional. You schedule the call, send the document, prepare the presentation, and arrive with a position. The old office often worked differently. People sometimes developed the position together.

The distinction sounds small, but it matters. A weak assumption survives longer when nobody challenges it. Two ideas remain disconnected when the people holding them rarely meet. A difficult problem can sit with one person because assembling three useful minds requires a calendar negotiation worthy of international diplomacy.

Remote work did not eliminate collaboration. It increased the transaction cost of spontaneous intellectual collaboration.

That reduced what I think of as thinking density: the amount of useful challenge, synthesis, pattern recognition, and refinement that can occur in a short period of time. And thinking density matters because it affects decision velocity. Ideas improve faster when they encounter other informed minds early and often.

AI changes the transaction cost of thinking

This may be one of AI’s more interesting organizational effects.

AI does not recreate the office, but it may recreate part of what made the office valuable as a thinking environment. It can provide an immediately available source of challenge, research, alternative explanations, and cross-domain connections without requiring three calendars to align.

A good AI system can help test an idea before it hardens. It can ask what evidence would disprove the hypothesis, compare an argument with established theory, search for contrary cases, reorganize a messy collection of observations into competing explanations, or take one industry example and ask whether the same underlying mechanism appears elsewhere.

Because the cost of another iteration is so low, an idea can go through many more cycles of challenge and refinement than would ordinarily be practical. The first version of an idea is rarely the best one, but organizational coordination costs often limit how many people can engage with it and how many times it can be reconsidered before a decision has to be made.

AI reduces some of that cost. It does not replace the people in the thinking room. It makes the room easier to enter.

It also has the advantage of being available at 11:30 at night, requiring no calendar invitation, and rarely complaining that the discussion really should have been an email.

Expertise still matters

AI is not, of course, the colleague who has lived through three failed product launches, knows which executive will quietly block the idea, or remembers why a remarkably similar strategy failed five years ago. It does not replace judgment, trust, organizational memory, political awareness, or the human willingness to say, “I think we are solving the wrong problem.”

And AI can still be confidently wrong.

That makes expertise more important, not less. A thinking partner is useful only if someone can recognize when the thinking is weak. The person using the system still has to know which questions matter, which assumptions deserve challenge, which evidence is credible, and which answer sounds elegant but violates something fundamental about the business.

AI can increase the number of ideas, challenges, connections, and iterations available to us. Expertise determines which ones deserve to survive.

That is why I find the argument over whether something is “AI-written” increasingly unhelpful. A more useful question is what role the human played in the thinking. Did the person understand the subject? Did they form the underlying thesis? Could they defend the reasoning? Did AI replace judgment, or did it increase the amount of useful work that judgment could perform?

Those are more meaningful distinctions than whether every sentence began life under a human fingertip.

The office was also a thinking system

Perhaps the office was never just a building. It was also a low-friction system for combining minds.

People exchanged weak signals before they became formal issues. Knowledge moved through informal networks. Ideas collided unexpectedly. Someone could ask a question without first finding an available half-hour next Thursday.

Distributed work weakened some of those mechanisms even as it improved many others. AI will not restore the old office, nor should that necessarily be the objective. But it may help restore some of the intellectual density we lost.

A persistent thinking partner can carry context across discussions, bring relevant evidence into the conversation, challenge assumptions on demand, and make experimentation cheaper. It cannot reproduce every valuable element of human collaboration, and it does not need to.

If AI can lower the transaction cost of thinking together—or even of thinking better alone—it can restore part of the organizational capability that became harder when work became more distributed.

Which leads to an appealing possibility.

We may be able to rebuild some of the best parts of the office without rebuilding the commute.