Science and health

14.1AI, the new instrument of science

14.2Biology and drug discovery

Diagram14.1. The compression of discovery time. By digitally simulating the interaction between a molecule and its target, AI makes it possible to rule out candidates doomed to fail very early on, shortening the slowest and most expensive phase.

Under the hood

AlphaFold 3 (2024) no longer predicts only the shape of proteins, but also their interactions with small molecules (the drugs), DNA and RNA; its database lists more than 200 million structures, open to research. DeepMind's spin-off, Isomorphic Labs, has turned it into a drug-design engine (sometimes nicknamed "AlphaFold 4" by experts), raised 2.1 billion dollars in 2026, struck partnerships with major laboratories (Eli Lilly, Novartis, Johnson & Johnson), and is aiming for its first AI-designed drug (against cancer) to enter clinical trials in late 2026. Others, such as Insilico Medicine, already have AI-derived candidates in phase 2. Alongside molecule design, AI is also accelerating genomics (interpreting DNA, predicting the effect of mutations) and the analysis of cell biology. Beyond prediction, it now makes it possible to design from scratch proteins and enzymes with desired properties (a field likewise recognized by the 2024 Nobel Prize in Chemistry), paving the way for new vaccines, treatments and biomaterials.

In context

Biohacking and the democratization of biology

By lowering the skill barrier, AI puts cutting-edge biology tools into the hands of enlightened amateurs, a movement sometimes called biohacking. The most telling case: in 2025-2026, an Australian consultant with no training in biology, Paul Conyngham, used a conversational assistant and AlphaFold to design a personalized messenger RNA vaccine against the cancer (a mast cell tumor) of his dog Rosie. The approach mirrors the logic of precision medicine: comparing the healthy genome with the tumor genome to spot mutations, modeling the corresponding proteins with AlphaFold, then targeting these "neoantigens." He did not make the vaccine himself: a specialized university institute (at UNSW, in Sydney) synthesized it, for a few thousand dollars. The reported result: roughly a halving of the tumor, without being a cure. This episode illustrates the promise (AI does not replace researchers, but widens the circle of those who can contribute, toward a more participatory medicine) as much as its limits: as specialists point out, these are working hypotheses, not validated treatments; only the route of clinical trials can establish a real benefit, and manufacturing still requires a laboratory. The darker flip side of this democratization, the biosecurity risk, is addressed in chapter 24.

In context

AI and longevity

A related field is attracting considerable funding: research on aging and longevity. The idea is no longer merely to treat diseases one by one, but to understand, and even slow, the mechanisms of aging itself. AI plays several roles here: analyzing immense biological datasets to build aging clocks (estimating the "biological" age of an organism from molecular markers), spotting candidate molecules to act on senescent cells or the pathways of aging, and accelerating the discovery of treatments (the same engines as in the previous section). Specialized companies (such as Insilico Medicine, already mentioned, or Retro Biosciences and Altos Labs on the longevity biotech side) are betting on this convergence. Caution remains warranted: extending healthy life is a serious scientific goal, but the field also attracts very optimistic marketing, and no validated "anti-aging pill" exists to date. The line between rigorous research and oversold promises is particularly thin here.

14.3Materials, climate and other fundamental sciences

In plain terms

The same approach is being deployed across many disciplines, where AI acts as an accelerator and an explorer in the service of the researcher:

Materials science: AI systems have proposed hundreds of thousands of new crystal structures, candidates for better batteries, superconductors or solar panels (a link to the energy transition of chapter 10).
Meteorology and climate: AI models now predict the weather faster, and sometimes better, than the heavy classical physical simulations, and help model the climate.
Physics: AI has made it possible to better control the burning plasma of fusion reactors, and to analyze the mountains of data from large instruments (accelerators, telescopes).
Neuroscience: it helps decode brain activity, all the way to brain-computer interfaces.
Mathematics: it is beginning to propose conjectures and to verify (or even find) proofs.

In context

AI for nuclear power and particle physics

Two fields deserve a spotlight, because they reverse the usual relationship: AI is no longer merely energy-hungry there, it becomes a tool in the service of science and energy. In civil nuclear power, it is already optimizing plant operation: predictive maintenance (anticipating an equipment failure), anomaly detection in the streams of sensor data, outage planning and fuel management, and accelerating reactor design (including small modular reactors, SMRs). The International Atomic Energy Agency (IAEA) has coordinated dedicated programs since 2021 and has designated reference centers (at Purdue for nuclear power, at MIT for fusion). In fusion, the flagship challenge is real-time control of the plasma: a gas at several million degrees, intrinsically unstable, which reinforcement learning systems have learned to confine by steering the magnetic fields, where classical methods struggle; AI also helps explore the parameters of future reactors. Finally, in particle physics, laboratories such as CERN have long relied on machine learning to sift through the mountains of collision data from the Large Hadron Collider (LHC) and pick out the rare signals of interest. AI and energy thus feed on each other (chapter 10).

In context

AI and space

The space sector has become a major consumer of AI. In Earth observation, models continuously analyze satellite imagery to track deforestation, crops, disasters, ice melt or emissions, turning a flood of images into actionable information. In astronomy, machine learning sorts through the massive data of large telescopes: detecting exoplanets in light curves (a method that revealed new worlds in the Kepler data), classifying galaxies, spotting transient events. In exploration, AI assists the autonomous navigation of rovers and probes (where the communication delay rules out live piloting), the planning of observations and the predictive maintenance of the craft. Finally, space surveillance (tracking debris in orbit, preventing collisions) increasingly relies on these techniques as low orbit fills up with satellites. Here again, AI acts as a force multiplier for scientific and operational work.

In context

AI that forecasts the weather

One of AI's most tangible successes for science is weather forecasting. Models trained on decades of atmospheric data (reanalyses, such as ERA5) have caught up with and then surpassed, on many metrics, the traditional physical systems that run on supercomputers. The pioneer, GraphCast (Google DeepMind, 2023), produces a ten-day forecast in under a minute on a single graphics card, where the reference model of the European Centre (the ECMWF) takes hours, and outperformed it on about 90% of the metrics tested; its probabilistic successor, GenCast (2024), generates ensemble scenarios and beat the European ensemble system on nearly 97% of the targets. Other players are in the race: Pangu-Weather (Huawei), particularly good for cyclones, Aurora (Microsoft), a "foundation model" of the atmosphere, and the ECMWF itself, which developed its own AI model (AIFS) and made it operational. In practice, these models do not (yet) replace the physical models: weather services and hurricane centers use them as an additional source of forecasts, and the divergence between the AI and the classical model becomes itself a piece of information about uncertainty. Two limits temper the enthusiasm: trained on the past, these models may poorly anticipate unprecedented extreme events (a question all the more pressing in a changing climate, chapter 10), and their resolution remains limited. The trajectory is no less spectacular: a field where accuracy gained about one day of forecast per decade has seen AI compress that progress into a few years.

In context

The integrity of science under threat

Beyond reproducibility, generative AI poses a more direct threat to the reliability of the scientific literature. Articles are now partly, or even entirely, written by models, sometimes without this being declared; revealing turns of phrase left in by mistake have been spotted in journals, along with fabricated images or data. The phenomenon feeds the "paper mills," which mass-produce fake publications, and puts strain on peer review, already overwhelmed, which some are trying (cautiously) to assist with AI. The risk is twofold: polluting the scientific corpus with dubious results, and eroding trust in publication. In response, publishers and institutions are developing detection tools, rules for disclosing the use of AI, and heightened vigilance. Here, science must guard against a tool it otherwise uses to make progress.

14.4Everyday health

In plain terms

At the patient's bedside, AI is advancing on several fronts.

Medical imaging: it detects tumors and anomalies on X-rays, CT scans or MRIs, sometimes with a sensitivity exceeding the human eye, and speeds up diagnosis.
Diagnostic support: conversational AI systems achieve, in some studies, better results than physicians on textbook cases (a result to be confirmed under real-world conditions).
Surgery: operating robots are being enhanced with AI copilots able to anticipate complications (chapter 13).
Personalized medicine: tailoring a treatment to the patient's genetic and clinical profile, and repurposing existing drugs.
Administrative relief: transcribing consultations, drafting reports, which can give medical time back. This is today one of the most widely deployed uses: so-called ambient assistants listen to the consultation (with the patient's consent) and automatically draft the report, integrated into the patient record. The most advanced example is the French company Nabla, used in mid-2026 by some hundred thousand practitioners across roughly one hundred and ninety healthcare organizations, which integrates with the main electronic health records and aims above all to reduce the administrative burden and the burnout of caregivers.
Mental health: conversational agents offer a first line of psychological support, accessible and non-judgmental. This is a promising avenue for access to care, but a high-risk one (chapter 19): such a tool does not replace a professional, can give bad advice, and must absolutely steer people toward human help in cases of distress.

In context

Brain-computer interfaces

At the frontier of neuroscience and the clinic, brain-computer interfaces (BCI) directly connect the brain to a computer, and AI is indispensable there to decode neural signals into intentions. The most publicized, Neuralink, implanted its first human patient in early 2024 (a quadriplegic man who was able to move a cursor, play and navigate by thought); by mid-2026, around twenty patients had received its implant (the size of a coin, fitted with thousands of electrodes placed by a surgical robot), as part of trials on motor control (PRIME), the restoration of speech (VOICE) and of vision (Blindsight). But Neuralink is not alone: Synchron inserts its sensor through the bloodstream, without opening the skull (less resolution, but far less risky); Precision Neuroscience, Paradromics or Blackrock Neurotech are exploring other approaches, while non-invasive headsets (EEG) detect mental states without reading thoughts. The promises are immense for paralyzed people, but so are the stakes: safety, the privacy of the most intimate data there is (brain activity), and the more distant prospect of an augmentation of capabilities, which raises the question of "neural rights." In the very long term, some players (Neuralink first among them) advertise a more radical ambition: a human-AI symbiosis, that is, a very-high-bandwidth interface designed so that humans "do not get left behind" by an ever more powerful AI. For now, this prospect belongs to the realm of professed vision, very far from today's real capabilities, and raises immense questions of both feasibility and ethics.

14.5Limits and risks specific to health

Debate

The health field concentrates specific risks that demand caution.

Liability. If an AI makes a wrong diagnosis, who is responsible: the physician, the software publisher, the hospital? The legal framework remains unclear (chapter 21).
Bias. A model trained mainly on Western populations may be less reliable for other groups, creating inequalities in care (chapter 21).
Validation and regulation. In health, the demand for evidence is at its highest. The European Union classifies AI-based medical devices as "high-risk," imposing audits and transparency (chapter 25).
Confidentiality. Training on health data, among the most sensitive there is, must comply with strict rules (the GDPR in Europe, chapter 21).
Hallucination. A language model can invent a reference or a dosage: in a medical context, this can be deadly. Hence a golden rule that holds for all of science and health: AI assists, but a qualified human decides.

Key takeaways (chapter 14)

AI is becoming a major scientific instrument: not to converse, but to discover, and even to formulate hypotheses ("co-scientist"). Symbol: AlphaFold (2024 Nobel Prize in Chemistry).
In drug discovery, it compresses timelines: AlphaFold 3 (200 million structures), Isomorphic Labs (2.1 billion raised, clinical trials targeted for late 2026), Insilico (already in phase 2).
The impact extends to materials, climate, fusion, neuroscience and mathematics; but poorly used AI also threatens scientific reproducibility.
In health, it is advancing in imaging, diagnosis, surgery, personalized medicine and mental-health support (the latter high-risk).
The risks are specific: liability, population bias, validation, confidentiality and hallucination. The principle: AI assists, the qualified human decides.

From scientific discovery to the transmission of knowledge: chapter 15 examines how AI is upending education.

14.1AI, the new instrument of science#

14.2Biology and drug discovery#

14.3Materials, climate and other fundamental sciences#

14.4Everyday health#

14.5Limits and risks specific to health#

Key takeaways (chapter 14)

14.1AI, the new instrument of science

14.2Biology and drug discovery

14.3Materials, climate and other fundamental sciences

14.4Everyday health

14.5Limits and risks specific to health