In March 2015, the International Agency for Research on Cancer classified glyphosate, the world's most widely used herbicide, as probably carcinogenic to humans. Eight months later, the European Food Safety Authority concluded it was unlikely to pose a carcinogenic hazard. Both bodies drew on substantially overlapping literatures. Both employed credentialed toxicologists. Both claimed to be following the evidence.

The usual explanation is that one side got it wrong, through incompetence or capture by interests. The more revealing story lies elsewhere. IARC asked whether glyphosate could cause cancer under some conditions. EFSA asked whether it would at realistic exposure levels. The two agencies also weighted different kinds of studies differently, including unpublished industry data. These were not errors. They were choices, and the evidence alone could not settle them.

Risk science is where Thomas Kuhn's insight becomes most practically urgent. Facts about danger are never simply discovered. They are assembled through networks of instruments, protocols, institutions and judgments about what matters. Understanding that assembly does not dissolve the facts. It shows us where to look when experts collide.

Every risk assessment begins before any data is collected, with a decision about what the question is. Is the concern mortality or morbidity? Individual exposure or population burden? Acute harm or chronic accumulation over decades? Each framing constructs a different object of inquiry, and each makes different evidence relevant. A study perfectly suited to one frame may be irrelevant in another. The study is not flawed. It simply answers a question no one asked.

Consider radiation protection's Reference Man, a standardised body defined by the International Commission on Radiological Protection: a Caucasian adult male, twenty to thirty years old, weighing seventy kilograms. For decades, dose calculations were calibrated to this figure. Women, children and fetuses, who are more radiosensitive, were pushed to the margins. No finding put them there. The reference point itself did. A technical convenience quietly encoded a judgment about whose body counted as the default.

Philosopher Richard Rudner made the deeper point in 1953. No hypothesis is ever confirmed with certainty, so scientists must decide how much evidence is enough, and that depends on the consequences of being wrong. Heather Douglas later developed this into the argument from inductive risk. Choosing a significance threshold, a statistical model or an extrapolation from animal to human doses involves weighing false alarms against missed harms. That weighing is irreducibly evaluative.

In Bruno Latour's terms, a risk is not found lying in nature, waiting to be measured. It is stabilised through a chain of translations: from laboratory rats to dose-response curves to regulatory thresholds. Each link reflects decisions about what to include. When experts disagree, they are frequently not disputing the same fact. They are holding different chains, built on different frames, and both chains can be internally rigorous.

Takeaway

Before asking which expert is right, ask which question each one is answering. Many scientific disputes about risk are disagreements about framing that look like disagreements about data.

Uncertainty is often treated as a neutral property of evidence, a measure of what we do not yet know. In risk controversies, it behaves more like a resource. It can be amplified, minimised, relocated or reframed, and actors with stakes in the outcome learn to deploy it with considerable skill. Uncertainty always exists. The real question is who gets to define how much of it matters.

The canonical case is tobacco. A 1969 Brown & Williamson memo stated plainly that 'doubt is our product'. Historians Naomi Oreskes and Erik Conway documented how the same strategy migrated to acid rain, ozone depletion and climate change. By insisting the science was unsettled, interested parties turned a normal feature of empirical research into a reason for regulatory paralysis. Demanding certainty before action is itself a policy position, one that presents itself as epistemic caution.

Yet the strategic use of uncertainty does not belong to one side. Advocates for precaution sometimes present preliminary or contested findings as settled, collapsing error bars to sharpen a message. Regulators may project confidence they do not possess to preserve institutional authority. Sociologist Brian Wynne observed that official risk communication often performs certainty as a way of managing publics. That performance tends to erode trust when events contradict it.

Recognising these dynamics does not mean every claim about uncertainty is equally suspect. It means asking sharper questions. Who bears the cost if the uncertainty resolves badly? Who benefits from delay? Whose standards of proof are being applied, and were they chosen before or after the stakes became clear? Uncertainty is a scientific fact, but how it is distributed across a controversy is a social achievement.

Takeaway

Uncertainty never speaks for itself; someone always speaks on its behalf. Asking who benefits from emphasising or downplaying it is a basic act of scientific literacy, not cynicism.

In 1983, the US National Research Council's influential 'Red Book' proposed a clean division. Risk assessment would belong to science, estimating probabilities and magnitudes. Risk management would belong to policy, deciding what to do about them. The separation promised to keep values out of the laboratory. Yet values had already entered upstream, through framing, evidential thresholds and the handling of uncertainty.

Sociologist Thomas Gieryn called the drawing of such lines boundary-work: the ongoing rhetorical labour by which communities decide what counts as science and who counts as an expert. These boundaries are negotiated rather than discovered, and they distribute authority. Wynne's study of Cumbrian sheep farmers after Chernobyl shows the cost of drawing them badly. Scientists predicted radioactive caesium would soon become immobilised, based on its behaviour in clay soils. On the acidic upland peat the farmers knew intimately, it stayed mobile, and restrictions on some farms lasted until 2012. Local knowledge, dismissed as anecdote, had been relevant all along.

Philosophers Silvio Funtowicz and Jerome Ravetz argued that a different kind of science applies where facts are uncertain, values in dispute, stakes high and decisions urgent. They called it post-normal science. In that domain, quality control requires an extended peer community that includes those affected by the decision. Sociologists Harry Collins and Robert Evans offered a caution: views are not all equally informed, and genuine expertise must still be told apart from mere opinion.

The productive answer lies between these positions. Experts are indispensable for describing what is known, what is not, and what different choices would likely produce. But the decisions embedded in framing, acceptable error and the tolerability of harm belong to democratic deliberation. The aim is not to strip science of authority but to locate that authority precisely.

Takeaway

Expertise can tell us what is likely to happen under different choices; it cannot tell us which risks a society should accept. The boundary between the two is not a wall but a seam that must be stitched with care.

Scientists disagree about risk not because science is broken, but because risk is a hybrid object, part measurement and part judgment. The glyphosate dispute was never only about tumours in rodents. It was about which question to ask, how much doubt to tolerate, and who should decide.

Seeing the social architecture of risk science does not weaken it. It makes visible the decisions that were always there, so they can be debated openly rather than presented as technical necessity.

This is the constructivist argument at its most useful. Knowledge built carefully, with its frames and values exposed, is more trustworthy than knowledge pretending to have no builders. Objectivity in risk science is not the absence of values. It is their accountable presence.