How Dangerous Is Nuclear Power?

Comparing energy sources by counting accidents is misleading, because a source that produces very little electricity can look safe simply by being small. The standard fix is to divide deaths by energy produced, giving deaths per terawatt-hour. A terawatt-hour is roughly the annual electricity use of about 150,000 European households, so these are deaths per large block of generation rather than per plant.

The figures in the chart above combine two studies. Markandya and Wilkinson (2007, The Lancet) produced the fossil fuel and biomass numbers, and Sovacool et al. (2016, Journal of Cleaner Production) compiled an accident database covering renewables and nuclear. Our World in Data maintains the combined table that most public comparisons are built from.

The single most important thing about these numbers is that the fossil fuel figures are mostly not accidents. They are dominated by air pollution: particulates, sulphur dioxide and nitrogen oxides from combustion, causing respiratory and cardiovascular deaths spread thinly across large populations. Coal’s accident-only rate is a small fraction of its total. Nuclear’s figure is the opposite, made up almost entirely of accidents plus occupational exposure, because routine radiation releases from operating plants are negligible.

The published figures are also conservative in a specific direction. Markandya and Wilkinson measured European power stations with modern pollution controls. Coal burned without those controls, which describes much of the world’s coal fleet, kills at a considerably higher rate, so the gap between coal and nuclear in the chart is if anything understated.

Hydropower is listed at 1.30 deaths per terawatt-hour, which is far above nuclear. Almost all of that comes from one event: the failure of the Banqiao dam in China in 1975, with an estimated toll usually given as about 171,000 and ranging across sources from 26,000 to 240,000. Exclude Banqiao and hydropower drops to roughly 0.04, indistinguishable from nuclear and wind.

This is worth stating plainly because it is the same argument, run in reverse, that appears in nuclear debates. Whether a single catastrophic event belongs in a long-run average is a real methodological question, and the answer should not depend on which technology the answer favours.

One further honesty point. The Our World in Data figure for nuclear includes both Chernobyl and Fukushima, and the Fukushima component counts the 2,313 evacuation deaths rather than radiation deaths. Nuclear’s headline rate therefore already includes the indirect harm its critics emphasise most.

The RBMK-1000 reactor at Chernobyl had two design features no Western commercial reactor had. It had a positive void coefficient, meaning that when its cooling water boiled the nuclear reaction sped up rather than slowing down. And its control rods were tipped with graphite, so that inserting them to shut the reactor down briefly increased reactivity before reducing it. There was also no Western-style containment building. During a badly executed turbine test at low power on 26 April 1986, with most control rods withdrawn, those features combined into a runaway power surge and a steam explosion that blew the reactor open.

Two workers were killed on the night of the accident. Of 134 emergency workers and staff diagnosed with acute radiation syndrome, 28 died within the following weeks and months. Nineteen more acute radiation syndrome survivors died between 1987 and 2004, mostly of causes not necessarily attributable to radiation. The Chernobyl Forum, the joint IAEA, WHO and UNDP review published in 2005, summarised this as fewer than 50 deaths directly attributed to radiation, almost all of them among highly exposed rescue workers.

The other confirmed toll is thyroid cancer in people who were children or adolescents in 1986 and drank milk contaminated with radioactive iodine. About 20,000 thyroid cancers were registered in this group between 1991 and 2015. UNSCEAR attributes roughly one in four of them, about 5,000 cases, to the accident; the rest reflect other causes and much more intensive screening. Fifteen deaths from thyroid cancer have been recorded. Survival among treated cases exceeds 99 per cent, which is why a large number of cases produces a small number of deaths.

The Chernobyl Forum projected that up to 4,000 people could eventually die of radiation-induced cancer among the roughly 600,000 most exposed: emergency workers, evacuees and residents of the most contaminated areas. A WHO report the following year added about 5,000 more among larger, lower-dose populations in Belarus, Russia and Ukraine, giving a figure of around 9,000. Cardis et al. (2006, International Journal of Cancer) projected about 16,000 deaths across Europe by 2065. The TORCH report, written by Ian Fairlie and David Sumner in 2006, gave 30,000 to 60,000. Greenpeace, the same year, gave about 93,000.

These are not competing measurements. They are the same arithmetic applied to different populations. The method multiplies a collective dose, measured in person-sieverts, by a risk factor of roughly 0.10 fatal cancers per sievert, drawn from studies of atomic bomb survivors who received much higher doses. Fairlie has set the calculation out explicitly: a total collective dose of 600,000 person-sieverts multiplied by 0.10 gives 60,000 deaths. Whether you get 4,000 or 93,000 depends on how many millions of very lightly exposed people you include in the collective dose.

UNSCEAR and WHO have repeatedly examined the exposed populations for effects beyond thyroid cancer. In the general population they have found no detectable increase in leukaemia, no detectable increase in solid cancers, and no detectable increase in hereditary effects or birth defects. Cataracts and some leukaemia have been found among the most heavily exposed workers. The honest statement of this finding is that no effect has been statistically detected, which is not the same as proving that none exists; the linear no-threshold model would predict an excess too small to see.

Around eight RBMK reactors remain in operation, all in Russia, all modified after 1986 to reduce the void coefficient and redesign the control rods. None was ever built outside the former Soviet Union.

On 11 March 2011 a magnitude 9.0 earthquake and the tsunami that followed killed about 19,500 people in north-east Japan. The same tsunami flooded the Fukushima Daiichi plant, knocked out its backup power, and caused three reactor cores to melt.

No member of the public has died of radiation from Fukushima, and no case of acute radiation sickness occurred. One death has been officially recognised: in September 2018 Japan’s Ministry of Health, Labour and Welfare granted workers’ compensation for a man in his fifties who had worked at the plant and died of lung cancer, with a recorded total exposure of about 195 millisieverts. Causation for a single cancer is a probabilistic judgement rather than a diagnosis, which is why this case is a compensation decision rather than a clinical finding.

UNSCEAR’s 2020 report, published in 2021, concluded that no adverse health effects among Fukushima residents have been documented that could be directly attributed to radiation exposure, and that future radiation-related effects are unlikely to be discernible. Public doses were low: of about 1,700 residents whose exposure was assessed, two-thirds were under 1 millisievert per year and 98 per cent under 5, against a global average natural background of roughly 2.4 millisieverts per year.

The evacuation is where the deaths are. Japanese authorities have certified 2,313 disaster-related deaths among evacuees from Fukushima prefecture, distinct from deaths caused directly by the earthquake, the tsunami or radiation. About nine in ten were people over 66. The causes were the ones that kill frail elderly people who are moved suddenly: interrupted medical care, transfer trauma, cold, stress and suicide. Roughly a third of the deaths occurred in the first three months.

This has become an argument used in both directions, and the honest position is that it is unresolved. Critics of the evacuation point out that it killed thousands while the radiation it avoided killed nobody. UNSCEAR credits the evacuation with substantially reducing exposure, and the doses people would have received had they stayed are themselves modelled rather than measured. What can be said without dispute is that emergency response, not radiation, caused the deaths.

Japan screened roughly 380,000 children by ultrasound and found more than 200 thyroid cancers, far more than expected. UNSCEAR and most specialists attribute this to the screening itself. Sensitive ultrasound detects small thyroid cancers that would never have caused symptoms, a well-documented effect known as overdiagnosis, and the pattern of cases does not track estimated radiation dose. A minority position exists in the peer-reviewed literature: Tsuda et al. (2016, Epidemiology) argued the increase was real and large. The mainstream reading is overdiagnosis.

The area under evacuation orders has fallen from about 12 per cent of Fukushima prefecture in 2011 to about 2.2 per cent, with roughly 300 square kilometres still closed to habitation. Around 23,700 people remained displaced as of late 2025, down from a peak of about 470,000. Return has been slow: a peer-reviewed study published in SOIL in 2023 found that only about 30 per cent of the pre-accident population of the twelve evacuated municipalities had returned by 2020, with much lower rates in the worst-hit towns. Whatever the radiological findings, the social disruption was severe and is not over.

Half the core of Unit 2 melted, and the containment held. The US Nuclear Regulatory Commission estimated an average dose of about 0.01 millisieverts to the roughly two million people in the surrounding area, with a maximum of about 1 millisievert at the site boundary, against an American natural background of about 3 millisieverts a year. A twenty-year follow-up of the exposed population led by Talbott and colleagues at the University of Pittsburgh found no consistent evidence that the release affected cancer mortality, with a relative risk near 1.00 for all malignancies and no dose-response trend. A dissenting analysis by Wing and colleagues in 1997 argued otherwise. There are no confirmed radiation deaths from the accident.

Kyshtym, Windscale and SL-1 appear in many lists of nuclear power accidents. None of them was a civilian power plant.

The 1999 criticality accident at Tokaimura in Japan, which killed two workers, is also often listed as a power accident. It happened at a fuel processing facility, not a reactor or a storage site.

Counting these against nuclear power inflates its record. The correct statement is that in the history of civilian nuclear electricity, Chernobyl is the only accident with a substantial confirmed radiation death toll, and Fukushima has one officially recognised death.

Global spent fuel from civilian power comes to roughly 400,000 tonnes, with the United States holding about 90,000 tonnes across some 79 sites in more than 30 states and adding roughly 2,000 tonnes a year. Spent fuel is extremely dense, so those tonnages occupy a small volume: the US Department of Energy’s standard illustration is that all American commercial spent fuel stacked together would cover a single football field to a depth of about ten yards. This is a real contrast with fossil generation, which disperses its waste into the atmosphere, but a small volume of intensely radioactive material is a different management problem from a large volume of a mild one, not automatically an easier one.

Spent fuel is held first in cooling pools and then in dry casks: sealed steel cylinders inside concrete overpacks, passively cooled. The NRC states that since the first American dry cask was loaded in 1986, dry storage has released no radiation that affected the public or contaminated the environment. There are no confirmed deaths from commercial spent fuel storage anywhere. That record is genuine, and it is also short relative to the timescales involved; the NRC is actively studying canister ageing, including stress-corrosion cracking, and an absence of releases affecting the public is not a claim of zero degradation.

Finland is closest. The Onkalo repository at Olkiluoto, operated by Posiva, places spent fuel in copper and cast-iron canisters packed in bentonite clay about 430 metres down in bedrock roughly 1.9 billion years old, with capacity for about 6,500 tonnes. Trial emplacement with test canisters was completed in 2025 and operations are expected in 2026 or 2027. It will be the first deep geological repository for commercial spent fuel anywhere.

The United States has nothing. Yucca Mountain was designated by Congress in 1987 and capped at 70,000 tonnes. The Department of Energy moved to withdraw the licence application in 2010 and the project was defunded, after sustained opposition from Nevada and the Western Shoshone. The NRC completed its own safety evaluation in 2014 and 2015 and concluded the design would meet the standards for a million years, but no waste was ever emplaced, and the current administration has restated its opposition. Roughly 10 billion dollars was spent on the site.

Radioactivity falls steeply at first and then slowly. Short-lived fission products, principally caesium-137 and strontium-90, dominate the hazard for the first few hundred years, after which long-lived actinides such as plutonium and americium control a long tail. Directly disposed spent fuel returns to roughly the radiotoxicity of the natural uranium ore it came from after something between 100,000 and 300,000 years. The 10,000-year figure often quoted is a regulatory assessment horizon, not the point at which the material becomes harmless.

If nuclear power is as safe as the mortality data suggest, the interesting question is why it is not being built. The answer is cost and time, and the evidence here runs strongly against new nuclear in Western markets.

Vogtle units 3 and 4 in Georgia were projected at about 14 billion dollars for completion around 2017. The final cost was about 35 billion, with the units coming online in 2023 and 2024, roughly seven years late, and the contractor Westinghouse going bankrupt along the way. Hinkley Point C in Britain was approved in 2016 at 18 billion pounds for first power in 2025; the latest estimate is 31 to 35 billion pounds in 2015 prices, with first power now expected around 2029 to 2031. Flamanville 3 in France began at 3.3 billion euros and a five-year build; it cost about 13.2 billion to construct, closer to 19 to 24 billion including financing, and took seventeen years from first concrete to start-up. Olkiluoto 3 in Finland started at 3.2 billion euros and finished at about 11 billion, thirteen years late.

Lazard’s levelised cost analysis published in June 2025 puts unsubsidised new nuclear at 175 to 255 dollars per megawatt-hour, against roughly 38 to 78 for utility-scale solar, 37 to 99 for onshore wind and 48 to 107 for combined-cycle gas. An existing, fully depreciated American nuclear plant runs at about 32 dollars per megawatt-hour, which is why keeping old reactors open is a very different economic proposition from building new ones.

Two caveats cut in opposite directions. Nuclear’s cost is unusually sensitive to the discount rate, because almost all of it is upfront: the OECD-NEA and IEA joint costing exercise put French nuclear at 45 dollars per megawatt-hour at a 3 per cent discount rate, 71 at 7 per cent and 97 at 10 per cent. And the renewable figures are unfirmed, meaning they exclude the cost of covering periods when the wind is not blowing and the sun is not shining. Four-hour battery storage, the configuration usually costed, does not address multi-day or seasonal shortfalls. Adding firming and grid costs narrows the gap without closing it.

Nuclear construction is not inherently slow or expensive. South Korea has built at overnight costs around 2,200 dollars per kilowatt and China around 2,500, against more than 6,000 in the United States and an effective first-of-a-kind figure near 15,000 at Vogtle. The OECD-NEA and IEA survey found a range from about 2,157 dollars per kilowatt in South Korea to 6,920 in Slovakia. France built most of its fleet quickly and cheaply in the 1970s and 1980s.

The drivers of that gap are reasonably well identified: a single standardised design built repeatedly, several units on one site, an intact domestic supply chain, state financing, and continuity of programme. Lovering, Yip and Nordhaus (2016, Energy Policy) documented the cost histories across countries. Grubler (2010, Energy Policy) documented the uncomfortable other half of the story: France’s own costs rose over time as it scaled up and changed designs, which he called negative learning by doing. Whether the West could recover the Korean pattern is a live question, not a settled one.

The most substantial argument against nuclear power for climate purposes is not that it is dangerous but that it is slow. Amory Lovins argues that because nuclear is several times more expensive and much slower to deploy per unit of carbon avoided, spending on it displaces cheaper and faster options and therefore results in more emissions than the same money spent elsewhere. Mark Jacobson at Stanford makes a related case for a wholly wind, water and solar system. The IEA and OECD-NEA argue the reverse, that removing nuclear from the mix makes decarbonisation more expensive and less reliable. The disagreement rests on assumptions about firming, storage, transmission and land, and it is not resolved.

Civilian nuclear programmes have contributed to weapons programmes. India’s 1974 test used plutonium from the CIRUS research reactor, supplied by Canada with American heavy water for peaceful use. North Korea developed weapons from its Yongbyon complex and left the Non-Proliferation Treaty. Pakistan’s A.Q. Khan network exported enrichment technology to Iran, Libya and North Korea. Iraq and Libya both pursued weapons under civilian cover.

Safeguards exist and have been strengthened, notably by the 1997 Additional Protocol adopted after Iraq’s programme was discovered. The IAEA verifies declared material and can detect diversion. What safeguards cannot do is physically prevent a determined state from breaking out; they provide warning rather than prevention. This is a real cost of the technology and it is not measured in deaths per terawatt-hour.

American support for nuclear power peaked at 62 per cent in 2010, held around 57 per cent immediately after Fukushima, fell to a record low of 44 per cent in 2016, and recovered to 55 per cent by 2023, according to Gallup. In 2019, for the first time, a plurality of Americans described nuclear plants as unsafe. Support splits sharply by party. In Britain, favourable opinion fell from 40 per cent to 28 per cent in the months after Fukushima and had returned to 40 per cent by the end of that year. In Japan, support fell from 62 per cent before the accident to 39 per cent by April 2011.

Germany accelerated its nuclear exit after Fukushima, closing eight reactors immediately in 2011 and the last three on 15 April 2023. Coal’s share of German generation rose from 24.2 per cent in 2020 to 33.3 per cent in 2022, though the war in Ukraine and gas prices also drove that.

The most cited estimate of the cost comes from Jarvis, Deschenes and Jha. Their 2019 working paper put the social cost at about 12 billion dollars a year, more than 70 per cent of it from additional air pollution deaths, implying roughly 1,100 extra deaths annually. The peer-reviewed version, published in the Journal of the European Economic Association in 2022, revised the figure down to 3 to 8 billion euros a year. A separate synthetic-control study published in 2025 estimated roughly 170 additional respiratory deaths a year. Later analysis notes that renewable growth progressively offset the coal effect over the following decade.

The direction is agreed and the magnitude is not. Replacing nuclear generation with coal raised emissions and air pollution deaths in the short run, by somewhere between about 170 and about 1,100 deaths a year depending on method. Anyone citing the higher figure should know it comes from a working paper the authors themselves later revised downward.