If you’re asked over and over, “do you love me?”, the problem isn’t necessarily the answer, but that the question ends up draining it of its meaning.

In Being and Nothingness, Sartre put it in his own terms: we all long to be loved freely. But a proof obtained under pressure stops, precisely, being a proof, and the only “I love you” that counts is the one you never had to extract.

With data, it’s often the same story.

In many companies, the numbers aren’t there to decide but to make a decision acceptable after the fact. So we drown in indicators, analyses and dashboards. Then we pick the ones that let us say: “I’m not the one who chose, the data did.”

This isn’t necessarily bad faith, because deciding remains hard. Launching a product, changing a strategy, winding down an activity: none of these decisions come with a guarantee. So we look for something that reassures us. A validation, even a permission to act.

The trouble is that data which can only confirm what we already believed carries almost no information. Its value doesn’t depend on its precision, or its sophistication. It depends on a much simpler question: if the result had been different, would I have changed my mind?

If the answer is no, then the number isn’t there to teach us anything, but to comfort us…

This is an idea Karl Popper placed at the heart of his account of knowledge. A theory isn’t interesting because we’ve found examples that confirm it. It becomes interesting when we expose it to a test that could refute it.

Take a bridge. We don’t know it’s sound because it’s still standing. We know because it has borne a load it might not have borne. Well, with data it’s exactly the same.

An indicator only becomes credible the moment we accept the possibility that it proves us wrong. An A/B test has value because it can invalidate our intuition. A control group has value because it can reveal that our idea doesn’t work. A forecast has value because we’ll be able to compare it against what actually happened.

Conversely, many organizations produce numbers that never risk anything. Dashboards consulted after the decision, analyses whose every possible result would leave the intended action unchanged, or metrics that give the impression of controlling reality without ever exposing themselves to being contradicted by it.

So the useful question isn’t “what is the right data?”, but rather “what data could change my mind?”

And above all: am I genuinely ready to listen to it? This distinction matters even more with AI.

By making the production of plausible arguments almost free, analytics agents can now write a report, synthesize information, draw a curve or produce a convincing conclusion in seconds. But they don’t automatically create the test that would tell us whether that conclusion holds. Quite the opposite.

On problems where a check exists (a program that compiles, a calculation that’s right or wrong, a forecast confronted with the facts), the mechanism still works. But on more ambiguous, human or organizational matters, the risk is different. AI excels at manufacturing coherent explanations, not necessarily true ones. It may simply be the one we wanted to hear…

So the danger isn’t that it deceives us, but that it indulges us, sparing us the moment when we might have discovered we were wrong.

In the end, data is only worth something as long as it keeps the right to contradict us. Everything else is just a more sophisticated form of reassurance.