Among large Italian and European companies, 90% now use artificial intelligence (AI) mainly for one task: translating emails. Professor Telmo Pievani, philosopher of biology and evolutionary biologist, who joins the University of Milan on 1 October 2026, cited the figure in his lecture at the Molecular Medicine PhDay, and used it to open a discussion of AI in scientific research. We have an immensely powerful tool at our disposal, he argues, and we are mostly underusing it.
In a talk centred on serendipity, Pievani set out two very different ways of using AI, in the laboratory and across scientific research more broadly: one “muscular”, the other heuristic.
If most companies use AI mainly to rewrite emails, Pievani explains, we are using only a tiny fraction of what it can do. «As human beings, we are brilliant at writing emails in the right tone, with kindness, irony or sarcasm. And yet we ask artificial intelligence to do it for us», Pievani observes, noting that the technology also carries a significant cost in energy and water. «We need to push for a smarter, more creative use of artificial intelligence, for instance for people’s wellbeing, for research, for discovery», he adds.
Pievani draws a simple distinction. On one side is the “muscular” use of AI, essentially brute-force computing: researchers keep doing what they did before, only much faster and more efficiently. One example he gives is comparing two complete genomes, such as those of Homo sapiens and a Neanderthal, which now takes a fraction of the time it once did.
On the other side is the heuristic use: asking new questions that could not even be posed before. AI’s computing power rapidly evaluates thousands of possibilities that no person could check one by one, at least not in any reasonable timeframe. It is still the researcher, however, who decides which question to ask.
A report by Google DeepMind and MIT FutureTech, published in September 2026 and based on a survey of 637 researchers in the United States and the United Kingdom, puts figures on this divide. Of the scientists surveyed, 67% say AI now allows them to tackle more ambitious research questions. At the same time, 49% say their own work increasingly involves safer, incremental questions, compared with 28% who report taking on more high-risk work. The heuristic potential is there; putting it into practice is harder.
In his lecture, Pievani cited three examples of AI used heuristically: enzymes designed by Frances Arnold with AI assistance, materials discovered by DeepMind, and Project CETI on cetacean communication, covered in full in our article on serendipity. Three more cases, which Pievani did not mention, illustrate the heuristic approach just as well.
The first comes from James Collins’s laboratory at MIT. In 2020 his group trained a neural network to screen hundreds of thousands of molecules for their ability to inhibit bacterial growth, and identified halicin, a compound active even against strains resistant to common antibiotics (the study appeared in Cell). In 2023 the same group repeated the feat with abaucin, which is effective against Acinetobacter baumannii, one of the hardest bacteria to treat in hospitals (Nature Chemical Biology). In both cases, AI narrowed a vast pool of candidate molecules down to a handful, few enough to test in the lab.
The second example takes us into pure mathematics. In 2021 a team at the Technion published the “Ramanujan Machine” in Nature: an algorithm that automatically generates conjectures about fundamental mathematical constants, such as pi or Euler’s number, expressed as continued fractions. The algorithm proposes connections no one has yet noticed; proving them is up to mathematicians. It is perhaps the most literal example of heuristic use: the conjectures point to where to look, and humans supply the proofs.
Heuristic AI is also at work closer to home, on the San Raffaele campus. In a kidney cancer study, the S-RACE platform, developed by Vita-Salute San Raffaele University (UniSR) and IRCCS San Raffaele Hospital, identified two blood markers that clinicians had not been considering: disease signals that no one had asked the system to look for.
Asked what threatens serendipity in science today, and with it the chance of unexpected discovery, Pievani identifies one enemy: the algorithmic logic of social networks and the platforms we use every day. These systems are built to profile users from their past choices and keep serving up variations of what they already know. The result is an environment in which it is hard to stumble on something new, something you were not looking for.
«We would need serendipitous algorithms, capable of offering us something truly unexpected, but that is not how they are designed», says Pievani. An informed, critical use of AI in scientific research, by contrast, widens the field of inquiry.
Today’s algorithms have one clear limitation: they tend to be sycophantic, telling users what they expect to hear, and above all they rarely admit ignorance.
We asked Pievani whether AI might one day generate serendipitous research questions on its own. «I’m rather sceptical about that. Knowing that we don’t know, and understanding why, is something that sets us apart as human beings, and algorithms cannot yet replicate it. Harder still to reproduce is the moment when a discovery reveals something we didn’t even know we didn’t know». The philosopher of science calls this awareness generative ignorance, and paradoxically it can become a driver of new knowledge.
«As scientists, we should begin with a simple question: I have this new technology, highly effective and powerful in one specific area, so how can I use it to find more than before, and to find something new? The same principle applies as much to the telescope as to AI», Pievani concludes.
M. Codreanu, A. Imas et al., “AI in Science: Early Insights”, Google DeepMind / MIT FutureTech (September 2026). Survey of 637 US and UK researchers.
J.M. Stokes et al., “A Deep Learning Approach to Antibiotic Discovery”, Cell 180, 688–702 (2020). Discovery of halicin, MIT.
G. Liu et al., “Deep learning-guided discovery of an antibiotic targeting Acinetobacter baumannii”, Nature Chemical Biology (2023). Discovery of abaucin, MIT.
G. Raayoni et al., “Generating conjectures on fundamental constants with the Ramanujan Machine”, Nature 590, 67–73 (2021). Technion.
“Artificial Intelligence in Medicine: S-RACE and the Challenge of Real Clinical Data”, UniSR blog (5 June 2026). AI platform for predictive analysis in urological oncology, UniSR / IRCCS San Raffaele Hospital.