An illustration titled ‘What Radioactivity Is’, with six panels’ reading (top) ‘Water in pond is ‘stable,’ it expends no energy,’ water can be carried up-hill and its energy level raised,’ ‘as it runs downhill it gives off energy and reaches a stable state’, (bottom) ‘atom is stable, it does not expend nuclear (atomic) energy,’ ‘atom can be bombarded with ‘atomic’ particles and its energy level raised, atom is now radioactive,’ and ‘as it disintegrates it gives off energy in the form of radiations and reaches a stable state,’ United States, circa 1955. (Photo by FPG/Archive Photos/Getty Images)
Getty Images
A newly released study from Columbia University and Japan’s Radiation Effects Research Foundation (RERF) may reshape the regulatory map of United States nuclear policy. In a world-first, researchers applied machine learning to the storied dataset of Japanese atomic bomb survivors to tackle a long-standing question: How dangerous is low-dose radiation?
Atomic Survivors and the Uncertainty of Low Doses
For decades, Japanese atomic bomb survivors have served as a cornerstone of global radiation risk modeling. The challenge has always been what to do at the low end of the dose spectrum. The effects of radiation doses below 0.1 Gray (Gy), the equivalent of a few CT scans or years of background exposure, are difficult to assess with precision.
Conventional models like the Linear No-Threshold (LNT) model assume that even the smallest dose of radiation increases cancer risk in a directly proportional way. However, critics argue this oversimplifies biological complexity. One competing theory is hormesis, which proposes that low doses of a harmful agent may actually trigger beneficial biological responses. In radiation biology, this would mean that small exposures might activate repair mechanisms or adaptive responses that reduce the likelihood of disease, rather than increasing it. The LNT has been enshrined in the US nuclear regulatory process, while hormesis has been mostly ignored. Recently, renewed calls for a review of the long-standing convention have become louder. This study supports those who have been calling for a reform.
No Evidence of Harm Below 0.05 Gy
Above 0.05 Gy, radiation exposure causally increases all-cause mortality. But below that threshold, there was no statistically significant increase in risk.
The study employed cutting-edge Causal Machine Learning techniques, to reanalyze mortality data for over 86,000 survivors. Unlike standard statistical models that assume a fixed dose-response shape (such as linear or threshold-based models) CML lets the data define the relationship without imposing any specific dose-response structure. It lets the data speak for itself.
And the data indeed spoke. The 50 mGy level is especially significant because occupational limits in the United States currently allow up to 50 mGy per year for most adults. This has long been considered a safe threshold by regulators. The finding that mortality risk is not significantly elevated below this level supports the idea that existing occupational limits are not only conservative but may already include a buffer of safety.
For comparison, a single chest CT scan delivers around 7 mGy of radiation. Natural background radiation varies globally between 1 to 10 mGy per year depending on altitude, soil composition, and building materials. For example, people living in Ramsar, Iran, or certain regions of Kerala, India, receive more than 10 mGy per year from natural sources alone and yet show no consistent pattern of elevated disease incidence.
The finding challenges the ultra-cautious stance of regulators who apply the LNT model even at microdose levels. This approach has led to overregulation of nuclear power, radioactive waste handling, and medical imaging. It is driving costs upward for both clean electricity generation and life-saving diagnostic medicine. In environmental remediation, it translates into excessive cleanup mandates, such as those seen around Fukushima or United States Superfund sites. Perhaps most consequentially, the public fear of even the lowest doses of radiation has been fueled more by mathematical assumptions than by data-driven causal relationships.
Regulatory Implications: Challenge to the LNT Model
This new CML-based study does not disprove the LNT model, but it strongly questions its continued necessity as the default assumption in radiation protection. If regulators accept these findings, the likely outcome will be a recalibration of low-dose exposure limits based on empirically supported causal thresholds. This would also support a shift away from blanket LNT applications, particularly in environmental cleanup, where the stakes involve billions of dollars and decades of remediation.
More nuanced public health messaging could follow, helping to avoid unnecessary fear over trivial exposures. With radiation risk better understood through flexible, modern tools like CML, regulatory agencies would be equipped to distinguish between levels of concern and levels of convenience.
The need for scientifically grounded, adaptable regulation has never been greater. At a time when nuclear energy is being re-embraced as a cornerstone of low-carbon energy policy and when diagnostic imaging continues to revolutionize modern medicine, it might be time to update the nuclear rulebook with data that better reflects reality than precaution alone.