Insights
How AI Quietly Limits Organizational Ambition
Picture a researcher with two questions on her whiteboard, and a small decision she makes without noticing. The first is the one she has been circling for three years — awkward, badly defined, the kind of question where you cannot say in advance what evidence would even count. The second is adjacent, narrower, and clearly answerable with the data her team already holds. She chooses the second. Not because it matters more. Because when she describes both to the model, one of them came back with a plan, and the other came back with a paragraph of encouragement. She would describe this as efficiency. I think it is something else.
Joshua Gans has a working paper (NBER 33566) modelling exactly this. When decision-makers gain AI that interpolates well inside known territory, scientists respond strategically. When the tool is weak, they ignore it. When it is very strong, they push outward into genuinely novel ground, because that is where the tool adds most. But in the middle range — where most organisations sit today — scientists “work to the AI.” They constrain the novelty of what they pursue to match the range of what the machine can reliably handle.
Work to the AI
Moderate AI, the model suggests, can reduce research ambition. I think that finding travels far beyond science. Every organisation is now full of people making that whiteboard choice several times a week. Which analysis to attempt. Which customer problem to open. Which strategic question to put on the agenda. And a new, quiet criterion has entered the decision: how well does this go when I describe it to the machine?
Nobody instructs anyone to apply that criterion. It arrives as relief. The tractable question produces a draft by Thursday. The important question produces a week of uncertainty and nothing to show at the review. One of those is much easier to be a person inside an organisation with.
This is not a story about AI being limited. It is a story about ambition being quietly re-scoped by whatever happens to be legible to our tools — and the re-scoping leaving no trace. A narrowed question does not look narrowed. It looks like focus. It arrives on time, well-structured, defensible. The abandoned question never appears in any document, so no one ever weighs the trade.
Ambition being quietly re-scoped
Research on collective creativity is arriving at a similar place from another direction: AI use tends to raise individual output while lowering the diversity of what a population of people produces. Each person is better off. The group is less varied. Nature Reviews Psychology published a piece this summer arguing that the main protection against this homogenisation is metacognitive — intellectual humility, the habit of noticing your own reasoning while it happens.
Which is an odd conclusion, if you sit with it. The safeguard against a technology of enormous scale turns out to be a small private act of attention.
The safeguard against a technology of enormous scale
I do not think individuals can carry that alone. If a mechanism operates below awareness and is rewarded by the calendar, willpower is the wrong instrument. It has to be built into how the organisation asks.
What that looks like, tentatively: keep a visible record of the questions that were considered and set aside, so ambition leaves a trace. Ask, in review, not only what was found but what was not attempted and why. Fund a portion of work explicitly on the criterion that nobody currently knows how to approach it. Treat “the model had nothing useful to say about this” as a signal of frontier rather than a verdict of infeasibility.
Execution has become cheap. That was supposed to free us to ask larger questions. It will only do so if we notice how strongly we are being pulled toward the ones that are easy to answer.
The most expensive thing an organisation can lose is not time. It is the question nobody remembers deciding not to ask.