Marie Curie

In the [YOU] on AI Field Guide

The cycle asks what it means to take the orange pill: to see the machine clearly, without narcotic hype or paralysing fear. Curie is the cycle’s model of the third posture the discourse almost never occupies. Where most commentary swings between rapture and dread, she offers something harder: measure what can be measured, name carefully, refuse conclusions that outrun the evidence, and never pretend that the absence of felt harm is evidence of safety. She held a tube of radium by her bed because she found the light lovely. Her notebooks remain so radioactive they are stored in lead-lined boxes to this day. The pleasant warmth is not reassurance. It is the symptom.

Her relevance intensifies in the specific. The hidden labour beneath AI’s gleaming outputs—the data workers who tag images, rate outputs, and filter toxic content, often in the global South, often at psychological cost proportional to Curie’s burns—is the shed beneath the Nobel Prize. The dual-use nature of AI capability that heals and harms with equal indifference is radium: it burns tumours and healthy flesh with the same mechanism. The war between open and proprietary AI replays exactly the choice the Curies faced in 1898 and declined to resolve in favour of enclosure. And the harms of AI that arrive before they are felt—the slow drift of recommendation systems on attention, the deferred consequences of labour displacement, the erosion of the information commons—share radioactivity’s defining dangerous property: they cannot be felt as they accumulate.

The deepest point of contact is the discipline she modelled in the presence of a genuine mystery. The capabilities of large language models resist explanation in a way that resembles the early days of radioactivity: a real phenomenon, precisely measurable in its effects, opaque in its mechanism, surrounded by the temptation to claim more understanding than the evidence supports. Curie’s discipline—to label the measured phenomenon without asserting an explanation of it—is what the field most needs and least practises when it confidently attributes “understanding” or “reasoning” to systems whose inner workings remain genuinely unknown.

Her refusal to patent was also a refusal of a specific argument for enclosure: that only those who understand the dangerous thing can handle it responsibly. She and Pierre were the people most entitled to make that claim, and they declined to convert their expertise into a monopoly. The AI companies that keep their most powerful models closed appeal to safety; sometimes they are right, and sometimes safety is the respectable face of a strategy to corner a trillion-dollar market. Curie’s example tells us to ask, every time, whether the safety argument would survive if there were no money in the enclosure.

Origin

Born Maria Skłodowska in Warsaw in 1867 under Russian occupation, she emigrated to Paris to study at the Sorbonne, where women had only recently been permitted to enrol. She joined Pierre Curie in investigating the anomalous radiation emitted by uranium ores, and the work that followed was defined above all by its physical demands. From 1898 onward the Curies worked in a converted shed on the rue Lhomond—a leaking, unventilated structure described by a visiting chemist as a cross between a stable and a potato cellar—hauling pitchblende by the tonne and reducing it through fractional crystallisation repeated thousands of times to chase a substance present in a few parts per million. The radium that eventually emerged from four years of this labour existed as a tenth of a gram of nearly pure radium chloride. The insight and the labour were not separable: radium was found not by a flash of theory but by the willingness to do the unglamorous physical work of reduction until the answer appeared.

Pierre was killed in a road accident in 1906, and Curie succeeded him as Professor at the Sorbonne—the first woman to hold a professorial chair there—continuing research and directing the Radium Institute she had helped establish as a centre for research and treatment. During the First World War she developed mobile X-ray units, the petites Curies, to treat wounded soldiers near the front. Her decades of unprotected exposure to radioactive materials—a danger that was not understood because the science did not yet exist to understand it—destroyed her bone marrow. She died on 4 July 1934. She is buried in the Pauthéon, honoured on her own merits. Her papers remain too radioactive to be handled without protection.

Key Ideas

Measure first. Curie’s method was the patient accumulation of precise measurement before any conclusion. Her discovery of radioactivity did not begin with a theory of what was happening inside the atom; it began with the observation that the ionisation of air around uranium ores was proportional to the quantity of ore and independent of its chemical state—a reproducible, measurable fact that demanded explanation. The electrometer that made this possible extended human perception into a domain where the senses were blind. The AI analogue is the insistence on measuring what the systems actually do rather than inferring what they must be doing from the fluency of their outputs.

The discipline of not knowing. Curie held phenomena as real and unexplained without collapsing the gap prematurely. Radioactivity seemed to violate the conservation of energy; she measured it with relentless precision, named it carefully, and resisted the urge to claim she understood its source before the evidence permitted. The same discipline applied to AI means refusing both the confident claim that these systems understand and the confident claim that they cannot, because neither verdict is supported by the current evidence. The uncomfortable middle—we have built something real whose inner nature we do not understand—is the honest place to stand.

Knowledge belongs to everyone. The decision not to patent the radium-extraction process was a specific institutional act, not a vague aspiration. The Curies published the method in full so that any researcher could use it. The principle they invoked—that a discovery rather than an invention belongs to the world—maps imperfectly onto AI, where the models are artefacts built with enormous private investment. But the principle survives its imperfect application: the burden of justification falls on enclosure, and the default for world-altering knowledge should be openness. Open-source AI inherits this argument.

The externality is real. Curie’s death was not incidental to her science; it was the externality of radium made flesh. The benefits of her discovery diffused widely—to cancer patients, physicists, and nations. The cost concentrated in a few bodies, hers chief among them. This is the characteristic shape of the cost of a new power, and it is the shape AI’s costs are already taking: the data workers who bear psychological damage, the communities whose water and power feed data centres, the workers displaced by automation. None of these costs appears on the balance sheet of the firms that capture the benefit. Curie could not have known what her notebooks were doing to her marrow. We can see our technology’s costs forming and are choosing, for now, to handle the material without gloves anyway.