September 22, 2026
AI is making research cheaper to produce at precisely the moment it may be making research more valuable.
There is no real contradiction in that. We have simply spent a very long time confusing the value of research with the value of the document that contains it.
For most of its history, research has had one kind of reader: a human one. A portfolio manager reads a report and connects the concept to a previous one. An analyst recognizes an argument she has seen before, or notices that a colleague has begun to think differently about a company. A research director, having watched an industry for twenty years, develops an instinct for which changes matter and which ones do not.
This is how knowledge has accumulated in markets: imperfectly, selectively, and very humanly.
There is a limit to how much any one person can carry. Nobody remembers every forecast an analyst made over a career, every assumption that changed along the way, every abandoned thesis, every disagreement with a peer, and what eventually happened. Nobody can read twenty years of research from hundreds of analysts and hold all of it in mind at once.
So we built our idea of research around what its human reader could see.
A telescope does not make the stars more interesting. It reveals that what appeared to be a point of light has structure. A telescope may make stars more interesting to humans, and the availability of telescopes makes stars more interesting to larger numbers of people, but it does not change the nature of stars themselves. A microscope does something similar to a drop of water. A geological survey can turn an unremarkable stretch of ground into something quite different, not because the ground changed, but because our ability to see it did.
I wonder whether AI is doing something similar to research.
Take the most ordinary earnings note. To someone who has already read thirty reports on the company that morning, it may contain very little that feels new. Revenue was ahead, margins were disappointing, estimates moved, the price target changed, the thesis survived. Our industry has produced millions of documents like it, and their sheer number has made them feel unremarkable.
But another reader may see something else entirely.
It may notice that the analyst changed an assumption she had held for six quarters, recognize that her explanation for the margin decline differs subtly from everyone else's, connect that explanation to something she first argued eighteen months ago, and then compare the evolution of that argument with hundreds of others made under similar circumstances.
What looked like another earnings note begins to look like a moment in a much longer chain of thought.
Nothing about the research changed. What changed was the capacity of the reader.
This, I think, is where a peculiar dissonance is beginning to appear. Some in our industry talk about AI as though it must inevitably devalue research because it can summarize it, reproduce much of it, and increasingly write it. Yet the same technology can see relationships within the accumulated body of research that no human being has ever been able to see.
The document may indeed become cheaper, while the thinking inside it becomes more valuable.
For a very long time, the document and the idea have been almost inseparable. An analyst had a thought, tested it, developed it, and published it in a report, and the report carried the argument, the evidence, the author's identity, and the institutional reputation behind it. It was both the thought and the vessel in which the thought travelled.
AI pulls them apart.
Once that happens, an archive of twenty years of research no longer looks quite like an archive of documents. Seen by a machine, it becomes a history of people changing their minds: forecasts made and revised, assumptions abandoned, theses strengthened and then weakened. One analyst saw something before everyone else. Another resisted the consensus and turned out to be wrong. An idea appeared in one place, reappeared somewhere else, evolved, combined with another idea, and eventually influenced a decision.
There is extraordinary value in that history, but there is also danger in separating it from the documents that once kept it intact.
A machine can carry a conclusion forward and leave its reasoning behind, preserve an insight while losing the identity of the person who had it, or combine the thinking of several firms into an answer that appears to have no author at all. It can absorb years of intellectual work and return no signal to the analyst that her thinking was used, useful, challenged or influential.
The more powerful the new reader becomes, the easier it is to lose precisely what made the research worth reading in the first place.
This is why I no longer think machine-readable research is an adequate goal. The problem is not whether machines can read our research. They plainly can. The problem is whether humans and machines can recognize the same ideas inside it.
If an analyst makes an assertion, changes a thesis, identifies a catalyst, or challenges an assumption, the machine needs to recognize what kind of act has occurred. More importantly, when that thought travels, changes form or becomes part of an answer somewhere else, we need to be able to recognize it too.
Nothing more exotic, perhaps, than a way of marking that an assumption changed, that a catalyst was named, that a thesis gave way to another - small enough to attach to a single sentence, durable enough to survive that sentence being copied, summarized and rewritten a hundred times.
Not because humans and machines should think alike. That would defeat the purpose. Markets depend on disagreement, on different interpretations of the same facts, on the stubbornness of someone who sees something everybody else has missed. The language is needed so that, as ideas move between people and machines, we do not lose the path by which they came into being.
An answer should still lead us back to the reasoning behind it, the reasoning to the ideas and evidence from which it was built, and those ideas to their source and, ultimately, to the people who had them.
The path should work in the other direction too.
An analyst should be able to know where her thinking travelled. A research department should be able to see which ideas endured, which ones changed other people's thinking, and which disappeared. An institution should be able to understand not merely how much research it produced, but what became of the thought contained in it.
That would close a loop that AI is otherwise in danger of breaking, and it might also change what we believe a research archive to be.
There are decades of thought sitting inside these systems: not just conclusions, but arguments and mistakes, convictions and reversals, disagreements that were resolved and disagreements that never were. There are analysts with recognizable ways of reasoning, institutions with intellectual traditions, and ideas whose histories may be far longer than the documents in which we happen to find them.
But people thinking together, arguing with one another, inheriting ideas from those who came before them, and leaving ideas for those who come after them create something that persists. Until now, much of it has been buried in documents because documents were the only practical way we had to preserve it.
AI gives us a new way of seeing what is there, and that makes the language we use to describe it matter enormously. Perhaps the mistake we are making about AI and research is a very simple one.
We are watching the container become cheaper and assuming that what is inside it must be worth less.
The new reader may show us the opposite.