How Safe and Clean Is Nuclear Power, Really? A Worst-Case Check

Adapted from my note “How safe and clean is nuclear? A worst-case check” (2026-10-01)¹, where every number below is listed with its source so the charts can be recalculated.

The question

Two numbers dominate arguments about how “clean” and “safe” an electricity source is:

1. Lifecycle greenhouse-gas emissions, in grams of CO₂-equivalent per kilowatt-hour.
2. Deaths per terawatt-hour, from accidents and air pollution.

The big compilations are the IPCC lifecycle reviews^2,3 for emissions and Our World in Data⁴ for deaths. The IPCC figures cover the whole lifecycle: mining, construction, operation and decommissioning. The death figures are narrower: accidents and air pollution, with nuclear’s number based on Chernobyl and Fukushima. On both measures nuclear sits down with wind and solar, far below coal, oil and gas. That is all this piece compares: greenhouse gases and estimated deaths. Waste, land use, cost and other environmental effects are real questions, but they need different measures.

The standard objection is that these tables are too kind: they use median or central values and leave out the worst estimates. That is a fair objection, and it can be tested. I take the highest nuclear estimates from the major surveys and debates, put them in the charts, and ask two questions: how much does the picture change, and why don’t the big assessments use those numbers? This is a sensitivity check, not a search of every scenario ever published: a few modelling exercises go higher still, and I come back to why below.

When I call an estimate rejected or criticised, a cited source says so. When a number is excluded for a methodological reason, or merely disputed, I say that instead.

Emissions: the screened range

The IPCC’s 2011 renewables report (SRREN) reviewed 2,165 references; 296 of those that passed its quality screens supplied the lifecycle estimates in its comparison². It explicitly did not screen out outliers or judge input assumptions. Its minimum and maximum therefore include every estimate that passed, however high.

Axes: each row is an electricity source; the horizontal axis is lifecycle emissions in grams of CO₂-equivalent per kWh. The thin bar spans the lowest to the highest screened estimate, the thick bar the middle half of estimates, the black tick the median. Look for: the right end of nuclear’s thin bar (220 g) and the left end of gas’s (290 g). Takeaway: even nuclear’s highest screened estimate is below the lowest screened estimate for coal and gas. SRREN excluded land-use change, which is why hydro looks so tidy here; more on that below.

Emissions: adding the worst cases

The IPCC’s 2014 report (AR5) narrowed nuclear to 3.7–110 g, with a median of 12 g³. On top of that range I place two of the highest nuclear figures in the debate: the 288 g maximum of Sovacool’s 2008 survey⁵, and Jacobson’s 180 g⁶. Jacobson’s number is not a lifecycle measurement like the others: most of it is opportunity cost, explained below.

Axes: as above, but from AR5, which publishes only minimum, median and maximum, for specific plant types. Orange markers show the two high nuclear figures (diamond: Sovacool; triangle: Jacobson); the axis stops at 1,000 g, so hydro’s bar is cut off (its maximum is 2,200 g). Look for: the diamond (288 g) against the gas bar. Takeaway: at 288 g, nuclear would be worse than the median solar estimate and inside the biomass range, above its median, but still well below gas (410 g minimum) and coal. The bigger outlier on this chart is not nuclear at all: it is hydro.

Deaths per TWh

Our World in Data’s central estimates⁴ combine accident deaths with deaths from air pollution, drawing on Markandya & Wilkinson⁷ and Sovacool et al. They mix sources of different scope: the fossil-fuel rates come from European electricity, the wind and solar accident rates from 1990–2013, and nuclear and hydro from global generation since 1965. For nuclear, OWID assumes 433 deaths from Chernobyl and 2,314 from Fukushima (almost all from the stress and hardship of the evacuation, not radiation).

To test the Chernobyl assumption, I keep Fukushima and the generation total fixed and swap in other Chernobyl estimates: (Chernobyl deaths + 2,314) ÷ 96,876 TWh, the total electricity nuclear produced from 1965 to 2021.

Axes: each bar is deaths per TWh, on a logarithmic scale: each labelled step (0.01, 0.1, 1, 10) is ten times the previous one. Grey is fossil, green renewable, blue nuclear as OWID counts it, orange nuclear recalculated with a different Chernobyl toll. Look for: where the orange bars land between the blue nuclear bar and the fossil bars. Takeaway: every Chernobyl estimate up to Greenpeace’s 93,000 keeps nuclear below gas and hydro. Only Yablokov’s 985,000 pushes it above gas, and even then it stays below coal and oil. Even OWID’s central nuclear figure is not lower than solar’s; on deaths, nuclear, wind and solar are in the same small range.

The extreme numbers, one by one

The extreme values are not all the same kind of thing. Some are real, some come from contested assumptions, some measure something else, and one has been rejected.

Hydro at 2,200 g: real, but site-specific

The argument. When a reservoir floods vegetation, the organic matter decays under water, and in low-oxygen conditions much of it comes out as methane. AR5 attributes hydro’s huge range mainly to this, and links the high values to reservoirs that flood a large area for little electricity. Its maximum of about 2 kg per kWh comes from one specific Brazilian power station⁸.

How it is treated. It is not rejected. AR5 keeps it but notes that reservoir methane varies enormously from site to site, with no well-defined maximum. The best predictor is how much area a reservoir floods per unit of electricity⁹. A few shallow, tropical, low-output reservoirs are worse than coal per kWh. The median hydro estimate in the IPCC compilations is far below coal.

High nuclear lifecycle estimates: contested assumptions

The argument. Storm van Leeuwen and Smith built a detailed account of every step of the fuel chain and argued that as uranium ore gets poorer, the energy (and therefore emissions) needed to mine it rises steeply. Their estimates ran roughly 84–166 g. Sovacool’s 2008 survey of 19 studies reported a range of 1.4–288 g, with a mean of 66 g, driven by fossil-fuelled construction, enrichment and mining⁵. The 288 g is not any single study’s result: Sovacool built it by adding up the highest estimate for each stage of the lifecycle (mining, construction, operation, waste, decommissioning), each taken from a different study.

How it is treated. Sovacool himself noted that the Storm van Leeuwen and Smith estimate “has not been universally accepted”, summarising criticism that their supply-chain accounting was inconsistent and may overestimate the energy required. A harmonisation study by the US National Renewable Energy Laboratory (NREL)¹⁰ did two things. It first screened 274 references down to 27 with usable estimates for the common light-water reactor type. Among the methods it excluded were estimates that infer a process’s energy use from how much money it costs and the economy’s average energy per euro or dollar, rather than measuring the process itself; that shortcut is behind several of the highest numbers. It then recalculated the remaining estimates with consistent assumptions. The published median of 13 g and a range reaching 220 g became 12 g and 110 g, and that range is where AR5’s nuclear numbers come from.

A few modelling exercises go higher than anything charted here, for example scenarios combining very poor uranium ore with a coal-heavy economy, discussed by the OECD Nuclear Energy Agency¹¹. They rest on similar contested energy-accounting assumptions, which is why the major reviews do not use them.

Opportunity cost: a different quantity

The argument. Jacobson adds the emissions the existing grid keeps producing during the extra years a nuclear plant takes to plan and build (10–19 years in his assumptions) compared with wind or solar. His total of 68–180 g is a lifecycle part of 9–70 g, 59–106 g of opportunity-cost emissions, and up to 4 g for the risk of nuclear war or terrorism enabled by nuclear power⁶.

How it is treated. This is not a disputed measurement but a different question. SRREN only admitted attributional lifecycle studies, which ask “what does this plant emit?”, and excluded consequential ones, which ask “what changes in the whole system if we build it?”, because the two are not comparable. Opportunity cost depends on what the alternative is and how fast it can be built. It belongs in a planning comparison, which is the subject of a separate piece on renewables, storage and nuclear, not in a per-kWh table.

Chernobyl death tolls: mostly a disagreement about low doses

With Fukushima and generation held fixed, the differences between the recalculated rates come entirely from the Chernobyl estimate. Those estimates differ mainly in which population, which dose range and which time span they count.

Estimate	Chernobyl deaths	Deaths/TWh
OWID	433	0.03
Chernobyl Forum12	~4,000	0.07
WHO, wider13	~9,000	0.12
TORCH14	30,000–60,000	0.33–0.64
Greenpeace15	~93,000	0.98
Yablokov et al.16	985,000	10.19

What each one counts:

• OWID: its central assumption for deaths attributable to Chernobyl.
• Chernobyl Forum (IAEA/WHO/UNDP, 2005): projected lifetime deaths among about 600,000 most-exposed people (emergency workers, evacuees, residents of the most contaminated zones).
• WHO, wider (2006): adds about 5,000 projected deaths among people with lower exposure elsewhere in Belarus, Russia and Ukraine.
• TORCH (2006): projected excess cancer deaths worldwide, applying the LNT model to very low doses.
• Greenpeace (2006): a commissioned report, reported not to have been peer reviewed.
• Yablokov et al. (2009): claimed deaths between 1986 and 2004, drawn mostly from Russian-language literature.

How they are treated. Effects in highly exposed groups are measurable: thyroid cancers among people exposed as children, for example. The disagreement is about projections from very small doses spread over very large populations. Those projections (up to TORCH) have not been shown to be false, but epidemiology cannot reliably confirm them either. UNSCEAR (2012) recommends not multiplying very low doses by very large populations to estimate deaths, because of the uncertainties at low doses and the limited statistical power of such studies¹⁷. Critics such as Ian Fairlie argue this reversed UNSCEAR’s earlier practice and that some studies do show effects near background levels. This is a genuine scientific disagreement.

The Yablokov figure is treated differently. Balonov (2012) concluded that the book relied on comparing health trends between regions, which cannot by itself show that radiation caused the deaths, instead of studies that track individuals’ doses and outcomes; he estimated it overstated the toll by more than 800,000 deaths¹⁸. A review in Environmental Health Perspectives criticised it for excluding major technical studies and setting aside accepted methods for linking cause and effect¹⁹. In short: the 4,000–60,000 range reflects different populations and doses counted, plus a real disagreement about extrapolating low-dose risk; the 985,000 figure is rejected on methodological grounds by the reviews cited here.

Hydro deaths: the other fat tail

Hydro’s 1.3 deaths per TWh is dominated by a single event, the 1975 Banqiao dam failure in China. Without it, hydro’s rate would be about 0.04, comparable to nuclear, solar and wind⁴. Hydro, like nuclear, is a “fat tail” technology: very safe in normal years, with rare catastrophes that dominate the average.

What the worst case changes, and what it doesn’t

• Even with the high inputs tested here (288 g; about 10 deaths per TWh), nuclear stays below coal on both measures, below gas on emissions, and below coal and oil on deaths. Against wind and solar it is roughly comparable on central values and can lose under the high ones.
• The extreme numbers fall into four groups: real but site-specific (hydro methane); contested assumptions, narrowed by screening and harmonisation (high nuclear lifecycle estimates); a different quantity (opportunity cost); and rejected on methodological grounds (Yablokov).
• What per-TWh statistics cannot settle is how to weigh rare catastrophes, low-dose risk under LNT, waste governance, cost and build time. Some of these move the numbers themselves, as the Chernobyl and opportunity-cost cases show; others, like waste and cost, need a different kind of comparison altogether.

When you meet an extreme number for any energy source, ask three things: is it typical or one site, does it measure the same thing as the table it is compared with, and did later reviews accept it?

---

1. Duty-Knez, M. (2026). How safe and clean is nuclear? A worst-case check. Lifecycle emissions and deaths per unit of electricity, including the most pessimistic published estimates. Note, 2026-10-01. Full PDF — all numbers with sources in its appendix. ↩

2. Moomaw, W. et al. (2011). Annex II: Methodology. In IPCC Special Report on Renewable Energy Sources and Climate Change Mitigation (SRREN), Tables A.II.3 and A.II.4. ↩ ↩²

3. Schlömer, S. et al. (2014). Annex III: Technology-specific cost and performance parameters. In IPCC AR5 WGIII, Table A.III.2; see also Chapter 7 (Bruckner et al.). ↩ ↩²

4. Ritchie, H. What are the safest and cleanest sources of energy? and Nuclear energy. Our World in Data. ourworldindata.org/safest-sources-of-energy ↩ ↩² ↩³

5. Sovacool, B. K. (2008). Valuing the greenhouse gas emissions from nuclear power: A critical survey. Energy Policy, 36(8), 2940–2953, Tables 4–5. doi:10.1016/j.enpol.2008.04.017 ↩ ↩²

6. Jacobson, M. Z. (2009). Review of solutions to global warming, air pollution, and energy security. Energy & Environmental Science, 2, 148–173. ↩ ↩²

7. Markandya, A. & Wilkinson, P. (2007). Electricity generation and health. The Lancet, 370, 979–990. ↩

8. Kemenes, A., Forsberg, B. R. & Melack, J. M. (2007). Methane release below a tropical hydroelectric dam. Geophysical Research Letters, 34; and (2011) CO₂ emissions from a tropical hydroelectric reservoir (Balbina, Brazil). Journal of Geophysical Research, 116. As used in IPCC AR5 WGIII Annex II, section A.II.9.3.3, and Annex III, Table A.III.2. ↩

9. Hertwich, E. G. (2013). Addressing biogenic greenhouse gas emissions from hydropower in LCA. Environmental Science & Technology, 47(17), 9604–9611. ↩

10. Warner, E. S. & Heath, G. A. (2012). Life cycle greenhouse gas emissions of nuclear electricity generation. Journal of Industrial Ecology, 16, S73–S92. ↩

11. OECD Nuclear Energy Agency (2012). The Role of Nuclear Energy in a Low-carbon Energy Future, pp. 20–21. PDF at oecd-nea.org ↩

12. IAEA/WHO/UNDP (2005). Chernobyl: The true scale of the accident. Joint press release, 5 September 2005. ↩

13. World Health Organization (2006). Health Effects of the Chernobyl Accident and Special Health Care Programmes. Report of the UN Chernobyl Forum Expert Group “Health”. Geneva: WHO. ↩

14. Fairlie, I. & Sumner, D. (2006). The Other Report on Chernobyl (TORCH). PDF via nirs.org ↩

15. Greenpeace (2006). The Chernobyl Catastrophe: Consequences on Human Health. Amsterdam: Greenpeace. ↩

16. Yablokov, A. V., Nesterenko, V. B. & Nesterenko, A. V. (2009). Chernobyl: Consequences of the Catastrophe for People and the Environment. Annals of the New York Academy of Sciences. ↩

17. UNSCEAR (2012). Report to the General Assembly, Annex A: Attributing health effects to ionizing radiation exposure. ↩

18. Balonov, M. I. (2012). On protecting the inexperienced reader from Chernobyl myths. OSTI/ETDE record 22101406. ↩

19. Dreicer, M. (2010). Book review: Chernobyl: Consequences of the Catastrophe for People and the Environment. Environmental Health Perspectives, November 2010 (PMC2974725). ↩