The essay argues that nuclear risk has been consistently inflated in policy discourse through a recognizable playbook: LNT extrapolation, collective-dose multiplication, relative-risk framing, subgroup mining, and citing upper bounds of confidence intervals as point estimates. None of the techniques fabricate data — they exploit legitimate statistical tools in ways that make widely-distributed, low-dose exposures sound apocalyptic even when no epidemiological signal exists above background noise
By submitting the piece under the title 'The Radiation Exposure Lie,' the submitter foregrounds the framing that mainstream radiation risk reporting is not just imprecise but actively deceptive. The 115-point score signals that the HN audience found the methodological critique compelling enough to elevate.
The editorial extends the essay's relevance beyond nuclear policy by noting that every technique catalogued — linear extrapolation from tiny effects, multiplying small per-unit risks by huge populations, leading with relative rather than absolute change, fishing for significant subgroups, and citing upper confidence bounds — has a direct analog in product analytics and experimentation. The argument is that engineers should recognize the playbook because they encounter (and sometimes deploy) it w
Works in Progress published a long essay titled *How to Lie About Radiation* that hit 115 on Hacker News. The piece, framed as a sardonic how-to, catalogs the statistical and rhetorical maneuvers that have made nuclear risk one of the most consistently overstated hazards in modern policy discourse. It is not a defense of any particular reactor, nor a polemic against any particular regulator. It is a methods paper dressed as satire.
The author walks through the standard playbook. Start with the Linear No-Threshold (LNT) model — the assumption that a millisievert is a millisievert, that risk scales linearly down to zero dose with no safe threshold — then multiply that per-person risk by a very large population, and you can generate any death toll you want from any radiation event you want. This is how a Chernobyl plume that delivered a few extra microsieverts to a hundred million Europeans becomes tens of thousands of statistical cancers, even when no epidemiological signal is detectable above background noise.
The second move: pick your denominator carefully. Report the relative risk increase ("40% higher!") and not the absolute baseline ("from 0.4% to 0.56%"). Report collective dose in person-sieverts, a unit that scales with population and so makes any widely-distributed exposure sound apocalyptic. Cite the upper bound of a confidence interval as if it were the point estimate. Subgroup-fish until you find a cohort with a p-value below 0.05 and lead with it. None of this is fabricated data. All of it is technically true. All of it is misleading.
The radiation literature is where these techniques were perfected, but they're not confined to radiation. Every move in the essay — LNT extrapolation, collective-dose multiplication, relative-risk framing, subgroup mining, upper-bound citation — has a direct analog in the dashboards and A/B test reports that engineers ship every week. The reason it's worth a developer's time to read a 6,000-word piece on becquerels is that the epistemic failure modes are the same.
Consider LNT. In radiation, it's the assumption that effects observed at high doses (atomic-bomb survivors at hundreds of millisieverts) can be linearly extrapolated down to ambient background levels. In analytics, it's the assumption that a treatment effect measured on power users generalizes to occasional visitors, or that a conversion lift measured in week one persists at month six. Both are extrapolations beyond the regime where you have data. Both are convenient because they let you say something quantitative where honesty would require saying "we don't know."
Collective dose is even more familiar. Multiply a tiny per-user effect by a hundred million users and you get a headline number. "Our redesign cost users 47,000 hours of cumulative scroll time per day." Is that bad? Compared to what? The trick works because the human brain treats large absolute numbers as meaningful even when the per-unit effect is statistically indistinguishable from zero. Anyone who has sat through a quarterly business review knows that the easiest way to make a metric move is to change what you're summing over, not what you're measuring.
Subgroup fishing is the one that should make every PM uncomfortable. The essay notes that radiation studies frequently slice cohorts by age, sex, exposure duration, geography, and cancer type until something pops. Run twenty subgroups at p<0.05 and one will be "significant" by chance alone. This is the entire methodological substance of half the A/B test postmortems in the industry: the overall effect was flat, but mobile-Safari-users-in-Tier-2-markets-on-Wednesdays showed a 12% lift, ship it. Multiple comparisons corrections exist; almost no one applies them.
The Hacker News thread surfaced the inevitable rebuttals — *but LNT is the conservative assumption*, *but you can't ethically run the controlled experiment*, *but the precautionary principle*. The essay anticipates these. The argument isn't that conservative assumptions are wrong, it's that conservative assumptions presented as point estimates are dishonest. There's a meaningful difference between "we use LNT as a regulatory floor because we cannot rule out low-dose harm" and "LNT predicts 47,000 deaths from this incident." The first is a policy choice. The second is a fabricated number with a citation.
Three concrete habits worth borrowing from the essay's critique. First: when you report a relative effect, always show the absolute baseline next to it. "Conversion up 22%" means nothing without "from 0.18% to 0.22%." If your dashboard library doesn't make this easy, fix the dashboard library. The single highest-leverage change you can make to internal data culture is mandating absolute-and-relative side by side in every chart.
Second: write down your subgroups before you look at the data, not after. This is preregistration, and it's the only defense against the garden of forking paths. If you discover a subgroup effect post-hoc, label it as exploratory and run a confirmation experiment on fresh data before acting. The radiation epidemiologists who get this right do it; the ones who get it wrong are the ones generating the studies the essay is mocking.
Third: be suspicious of any metric whose magnitude scales with your user base rather than with the underlying phenomenon. "Cumulative hours of productivity lost" is the collective-dose unit of the SaaS world. It tells you almost nothing about whether the per-user effect is real, large, or worth acting on. Per-user effect sizes with confidence intervals are the boring, honest alternative. They will get less attention in the all-hands. Ship them anyway.
The essay's deeper point is that statistical literacy is no longer a specialist skill — anyone consuming a dashboard, reading a model evaluation, or sitting through an experiment readout is exposed to these moves daily. The radiation literature is just the laboratory where they were refined over seventy years, against a backdrop of public terror and regulatory caution that made motivated reasoning easy and falsification hard. Your A/B testing platform has a faster feedback loop, but the incentives are structurally similar: someone wants the number to mean something specific, and the methods that get it there are sitting right there in the toolkit. The defense isn't more sophisticated statistics. It's the discipline to show the absolute baseline, preregister the subgroup, and refuse to multiply small effects by large populations to manufacture significance.
Paradoxically active ongoing research suggests that the same dose of (medically determined) radiation delivered over a very short period of time increases its efficacy against the cancer compared to damage to the surrounding tissue.Here's one: https://pmc.ncbi.nlm.nih.gov/article
> Fifteen people have already died of these cancersDon't forget that USSR and Belarus authorities did everything to attribute illnesses to anything but radiation. And it's really hard to prove that some illness is _due_ to radiation anyway. The reporting was way less transparent and non
Even before Linear No Threshold was a thing scientists were doing experiments showing that dosing fruit flies with radiation all at once would lead to a highly mutated second generation, but spreading that radiation out over the course of a month wouldn't.
> Drinking one beer a night for a year is a lot less harmful than drinking 365 beers in one go. The same applies to radiation exposure, but regulation doesn’t agree.I know I may be overreacting, but when the dentists try to convince me that the annual x-ray of my family members is the same as bei
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
Reminded of something that Geraldine Thomas, founder of the Chernobyl Tissue Bank, wrote when the Fukushima drama was going on, "Look at the science – smoking and obesity are more harmful than radiation":* https://www.theguardian.com/environment/2011/apr/26&#x