‘To carry out research, you need artificial intelligence that can say “I don’t know”’

Three conditions have come together: data of unprecedented quality and depth, more mature AI methods, and hardware capable of training sufficiently powerful models. Today, we can make sense of information that did not exist or could not be analysed ten years ago. But we do not want to limit ourselves to using existing solutions: the most important part is to create new models and new ways of training them, inspired by the practical challenges of biomedical research.

What can AI do today that we couldn’t envisage before?

One example relates to microscopy. Using the MicroSplit and lambda-Split methods, we can virtually separate structures that overlap in the raw image. This allows biologists to view more information simultaneously using the same instrument, at no extra cost. A second example is closer to clinical practice: we have developed a method to predict the risk of breast cancer and are adding an explainable AI module. The system does not merely provide a prediction, but can show which image features contributed to determining it.

Is the issue of the ‘black box’ therefore a key one?

Yes, particularly in medicine. The risks of careless use are very real: a model might have learnt to recognise a detail in the scan rather than the disease itself, or might identify a random correlation. That is why we are working on models that do not provide a single answer but rather several plausible possibilities, on calibrating uncertainty and on explainable AI, to verify whether what the model has learnt makes biological sense.

You talk about an AI capable of saying “I don’t know”. Why is that so important?

Because a model that always gives an answer does so even when it is simply guessing. And in medicine, being wrong with confidence is worse than saying nothing at all. A calibrated model should say: here is my prediction, and here is how much you should trust it. When it is very confident, it can relieve the clinician of routine work; when it is not, it must refer the case to a human. Knowing one’s limitations is not a weakness; it is what makes a tool usable.

What are the main challenges?

The first is to turn a method into a reliable tool that biologists or doctors can use: simply demonstrating that something is possible is not enough; it requires serious engineering. The second is people. Talented individuals capable of working across mathematics, computer science and the life sciences are in high demand.

How long will it be before these tools deliver tangible benefits for patients?

‘It depends on the discovery. A new method can help a researcher within a matter of weeks; a clinical prediction tool may require years of testing and validation before it is introduced into hospitals. Speed cannot replace rigour.

And what about the future?

AI will undoubtedly change the way we do science. Data has become too complex to be analysed effectively by humans alone. But we must distinguish between its real potential and overhyped promises. Our aim is to accelerate the pace of scientific discovery and build tools that can genuinely help researchers, doctors and, ultimately, patients.