Does Alcohol Cause Cancer at Low Doses? The Evidence Behind 'No Safe Level'

The question

In January 2025 the United States Surgeon General issued an advisory with an unusually direct claim: alcohol consumption is the third leading preventable cause of cancer, after tobacco and obesity, and the risk "may start to increase around one or fewer drinks per day" for several cancers. The advisory put a number on the American burden — about 96,730 cases and 20,000 deaths a year — and recommended updated warning labels. The International Agency for Research on Cancer (IARC) had already gone further in a 2025 evidence brief: alcohol causes cancer at seven sites, and "even low levels of drinking increase cancer risk."

That claim collides with a different literature. Mendelian randomization (MR) studies, which use genetic variants that alter drinking behaviour to approximate a randomized experiment, have repeatedly found that people genetically predisposed to drink less do not have lower rates of many cancers. The largest such study, published in late 2025, examined twenty cancer types in about 1.5 million people and found a null effect on overall cancer, a null effect on breast cancer, and a statistically significant increase only for head and neck, colorectal, and oesophageal cancers. If alcohol were a linear carcinogen at all doses, that pattern should not appear.

The question this brief asks is therefore narrower than "is alcohol bad for you" and more useful than a slogan: which cancers does alcohol actually cause, how strong is the evidence at doses people actually drink, and what does "no safe level" mean quantitatively? The answer has three parts. The causal claim is secure for the upper aerodigestive tract and the colorectum, and reasonably secure for the liver. It is weaker for the breast than the public messaging implies. And the population burden is large not because the individual risk at one drink a day is large, but because drinking is common.

Why it matters

Prevention policy has to allocate attention across exposures that differ enormously in evidence and effect size. Alcohol sits in an awkward middle: a Group 1 carcinogen with a plausible mechanism, a large exposed population, and a monotonic dose-response in most studies, but small relative risks at low doses, difficult confounding, and a non-monotonic all-cause mortality curve. Getting the shape of that curve wrong produces two opposite errors: overstating the danger of light drinking and losing credibility, or understating the burden of heavy drinking and missing a large preventable fraction.

The population arithmetic is the least controversial part. A relative risk of 1.1 at one drink per day sounds trivial; applied to tens of millions of drinkers, it is not. That is why the burden is large even when the per-drink effect is small, and why the policy question is how much of it is reachable by plausible changes in drinking.

The individual question is different, and the evidence is less decisive than either side of the public argument admits. A person deciding whether to have a glass of wine is not asking for a population attributable fraction but for the change in their own absolute risk, which depends on age, sex, genetics, smoking, and competing causes of death. That risk is small, not zero, and more uncertain than the slogans suggest.

What the standard claim actually says

IARC's 2012 monograph classified alcoholic beverages, and acetaldehyde associated with their consumption, as Group 1 carcinogens — the same category as tobacco smoke and ionizing radiation. The classification was based on sufficient evidence in humans for cancers of the oral cavity, pharynx, larynx, oesophagus, colorectum, liver, and female breast. That list of seven has been stable for more than a decade and is the basis for the current public-health claim. The 2025 IARC evidence brief restates it and adds the quantitative claim that about 4% of cancers worldwide are attributable to alcohol. A 2023 commentary by WHO authors put the precautionary case plainly: the evidence "does not indicate the existence of a particular threshold at which the carcinogenic effects of alcohol start to manifest," so "no safe amount of alcohol consumption for cancers and health can be established."

The global burden estimates are consistent across groups. Rumgay and colleagues estimated 741,300 new cancer cases in 2020 attributable to alcohol, or 4.1% of all cases, with the largest contributions from oesophagus, liver, and breast. A 2026 Nature Medicine analysis of 30 modifiable risk factors put alcohol at 3.2% of the 18.7 million new cases in 2022 — behind smoking (15.1%) and infections (10.2%). A 2025 Global Burden of Disease analysis counted 343,370 cancer deaths attributable to alcohol in 2021, 3.5% of all cancer deaths, with liver cancer (27%), oesophageal cancer (24%), and colorectal cancer (16%) accounting for most of the mortality.

The part of the claim that is genuinely contested is not the existence of the association but its behaviour at the bottom of the dose range. Rumgay and colleagues estimated that 103,100 of the 741,300 cases arose from drinking below 20 grams of ethanol per day — roughly two standard drinks — and 41,300 arose from drinking up to 10 grams per day, roughly one drink. IARC's 2025 brief repeats the 41,300 figure and describes it as cases caused by "less than one drink per day." A 2021 analysis of the European Union found about 23,000 cases in 2017 attributable to drinking below 20 grams per day — 13.3% of all alcohol-attributable cancers and 2.3% of all cases at the seven sites — including about 11,000 breast cancers, with about 8,500 cases arising below one standard drink per day. These are model-based estimates, not counts: they depend on assumed dose-response functions, underreporting, and the choice of reference group. But the direction is consistent — a nontrivial share of the burden sits below the level most drinkers would call heavy.

The Surgeon General's American numbers make the same point: 44,180 breast cancers in women — the largest single site — and an estimated 17% of alcohol-attributable cases occurring within the federal dietary guidelines, which permit up to two drinks per day for men and one for women. That 17% figure is the strongest official statement that the "safe" threshold is not a biological boundary.

The dose-response, in absolute terms

Relative risks are easy to misread, so it is worth translating them into absolute numbers. The largest prospective study of women, the Million Women Study, followed about 1.28 million middle-aged women and reported that each additional 10 grams of alcohol per day — one drink — was associated with a 12% increase in breast cancer risk (95% CI 9–14%). It also reported larger relative increases for several other sites: 29% for oral cavity and pharynx, 22% for oesophagus, 44% for larynx, 10% for rectum, and 24% for liver. The pooled estimate for breast cancer across 53 studies was smaller, about 7% per 10 grams per day.

The absolute version is more useful. In the Million Women Study, 118 of every 1,000 non-drinking women developed breast cancer by age 75. At one drink per day, the estimate rose to about 133 per 1,000 — an additional 15 cases per 1,000 women over a lifetime, of which about 11 were attributed to breast cancer and four to the other sites studied. At two drinks per day the excess was larger, and it continued to rise in roughly linear fashion.

Those numbers are small individually and large in aggregate. A 15-in-1,000 lifetime excess is roughly a 13% relative increase over background — consistent with the per-drink estimate — and easy to miss in a single cohort but visible in a meta-analysis of millions of person-years. It is also easily confounded, which is why the genetic evidence matters.

The genetic test

Mendelian randomization exploits the fact that genetic variants affecting alcohol metabolism are allocated at conception, roughly independently of the behaviours and social conditions that confound observational studies. The best-known variant is rs1229984 in ADH1B, which speeds the conversion of ethanol to acetaldehyde and makes drinking unpleasant; carriers drink less. A second is rs671 in ALDH2, which slows the clearance of acetaldehyde and produces the flushing response common in East Asian populations.

The ADH1B variant has been used to test the cardioprotective hypothesis. Holmes and colleagues studied 261,991 people of European ancestry and found that each additional ADH1B allele was associated with 17.2% lower alcohol intake and a 10% lower risk of coronary heart disease (OR 0.90, 95% CI 0.84–0.96) — evidence against cardioprotection, and a validation of the instrument that can also be applied to cancer.

The largest cancer application is a 2025 Mendelian randomization analysis of four biobanks and consortia, with about 1.49 million participants, 95 genetic variants, and twenty cancer sites. Its estimates are expressed per standard deviation of genetically predicted intake, which is a modest shift — often less than one drink per week. The overall cancer estimate was null: OR 0.96 (95% CI 0.87–1.06). Head and neck cancer was elevated (OR 1.51, 95% CI 1.18–1.93), colorectal cancer was elevated (OR 1.21, 95% CI 1.01–1.45), and oesophageal cancer was elevated with a wide confidence interval (OR 1.42, p = 0.033). Lung cancer showed an association that attenuated after adjustment for smoking, consistent with confounding. Breast cancer was not significant: OR 1.09 in the biobank analysis (p = 0.25) and OR 0.98 in the consortium analysis. Kidney cancer (OR 0.64) and endometrial cancer (OR 0.56) were inversely associated, which the authors interpreted cautiously as possible confounding or reverse causation rather than protection.

The null breast result is the most consequential number in the recent literature, and it is easy to overread. Mendelian randomization estimates a local average treatment effect among people whose drinking is shifted by the instrument, not the effect of heavy drinking. That modest shift limits power to detect a small breast cancer effect. The ADH1B variant also alters acetaldehyde exposure specifically, which is well suited to sites where acetaldehyde is the carcinogen but may understate risk for hormonally mediated cancers. And a null in one well-powered study does not erase a large observational literature with a consistent dose-response.

Still, the pattern is informative. The sites where MR finds an effect — head and neck, colorectum, oesophagus — are exactly where ethanol and acetaldehyde contact tissue directly. The site where MR is null — breast — is where the mechanism is indirect and the per-drink observational effect is smaller than in the upper aerodigestive tract. That is what a real but modest causal effect looks like with an imperfect instrument — and also what a confounded literature would look like if the confounding were concentrated in breast cancer. The evidence cannot yet distinguish the two.

Two curves that look incompatible

The cancer literature and the all-cause mortality literature are often presented as if they disagree about whether moderate drinking is harmful. They do not, exactly. They describe different outcomes with different dose-response shapes.

The Global Burden of Disease 2020 Alcohol Collaborators estimated a J-shaped relationship between alcohol and all-cause mortality for people aged 40 and older, with a theoretical minimum risk exposure level between 0.114 and 1.87 standard drinks per day. For people aged 15–39, the minimum was between 0 and 0.603 drinks per day, and 59.1% of the world's harmful drinkers were in that younger group. The J-curve for older adults is the empirical basis for the idea that light drinking is neutral or beneficial.

That curve is fragile. A 2023 reanalysis of 107 cohort studies, with 4.8 million participants and 425,564 deaths, found that the apparent protective effect at low volume largely disappeared after adjusting for abstainer bias — the tendency of former drinkers to quit because they are already ill. Low-volume drinking (1.3–24 grams per day) was associated with a mortality relative risk of 0.93, but the result was not statistically significant (p = 0.07), and heavier drinking was clearly harmful (RR 1.19 at 45–64 grams per day; RR 1.35 at 65 or more). The authors concluded that abstainer bias explains much of the apparent J-curve.

Even if a small cardiovascular benefit at low doses survived every adjustment, it would not imply a cancer benefit. Cancer and cardiovascular disease have different dose-response relationships, latency periods, and mechanisms. An exposure that reduces clotting risk at low doses can still promote carcinogenesis over decades. The relevant question is whether the carcinogenic effect at low doses is large enough to matter against competing risks. The two literatures are compatible: a small cardiovascular benefit and a small cancer risk can coexist, with the net effect depending on age, sex, and baseline risk.

Mechanisms: why the gut and the throat before the breast

The mechanistic case explains the epidemiological pattern. Ethanol is not itself the primary carcinogen at most sites; acetaldehyde is. Alcohol dehydrogenase converts ethanol to acetaldehyde, which is highly reactive, forms DNA adducts, and is classified by IARC as a Group 1 carcinogen in the context of alcoholic beverages. Aldehyde dehydrogenase 2 (ALDH2) clears acetaldehyde. When ALDH2 is deficient, acetaldehyde accumulates, producing the flushing response and a markedly higher risk of oesophageal cancer.

The ALDH2 evidence is unusually direct. Brooks and colleagues estimated that people heterozygous for the null ALDH2 variant have an oesophageal cancer odds ratio of 3.7 to 18.1, depending on the comparison, and that the variant accounts for 58–69% of the excess risk in East Asian populations. About 540 million people carry the deficiency. The mechanism is not subtle: less acetaldehyde clearance, more DNA damage, more cancer. It also explains why the upper aerodigestive tract — the mouth, pharynx, larynx, and oesophagus — is the highest-risk territory. Those tissues receive the highest concentration of ethanol and acetaldehyde on first contact, and they have the least metabolic protection.

The colorectum and liver are next. The colorectum is exposed through systemic circulation and through local microbial metabolism of ethanol, which produces acetaldehyde. The liver is the primary site of ethanol oxidation and is exposed to both acetaldehyde and the oxidative stress of prolonged metabolism. The breast is different. The leading hypotheses involve estrogen metabolism, one-carbon metabolism, and DNA damage, but the exposure-response is weaker and the mechanism is less direct. That is consistent with a smaller relative risk per drink and with the difficulty of detecting it in genetic studies.

Alcohol also acts as a solvent for tobacco carcinogens, which is why the combined effect of smoking and drinking on head and neck cancer exceeds the product of their separate effects. A substantial fraction of alcohol-attributable upper aerodigestive cancers therefore occur in people who also smoke.

Competing explanations and counterevidence

The strongest argument against the causal interpretation is confounding. People who drink differ from people who do not in income, education, diet, smoking, physical activity, and access to health care. Adjusting for measured confounders helps, but residual confounding can produce a modest association even when none exists. Mendelian randomization has settled it for some sites and not others.

The strongest argument against the low-dose claim is measurement. Surveys underreport alcohol consumption, often by 30–50% relative to sales data. If the true intake of "one drink per day" is actually higher, then the observed dose-response at the low end is biased upward — the people classified as light drinkers are drinking more than they report. That would mean the true risk at one drink per day is smaller than the observational estimates suggest, not larger.

The strongest argument against the J-curve is abstainer bias, as the 2023 reanalysis showed. The strongest argument against the MR null is power and instrument validity. The strongest argument against the whole enterprise is that the outcome is rare, the exposure is lifelong, and the relevant effect size is small enough that no single design can resolve it. Together they explain why the field's estimates have not converged.

There is also a funding asymmetry. Alcohol research has been financed by both public health agencies and industry-linked groups, and the historical record includes attempts to shape the evidence on moderate drinking. That does not invalidate the current literature, but it argues for weighting systematic reviews and preregistered analyses more heavily than single studies.

Evidence quality by site

The evidence is not equally strong across the seven cancers, and the differences are large enough to matter for how the claim should be stated.

Oral cavity, pharynx, larynx, and oesophagus. Strong. Consistent observational dose-response, mechanistic evidence from acetaldehyde and ALDH2, genetic support, and a plausible synergy with smoking. The oesophageal effect is the best-understood because the ALDH2 variant provides a natural experiment.

Colorectum. Strong. Consistent observational associations and genetic support in the 2025 MR analysis. The mechanism is plausible through local acetaldehyde production, though less direct than in the upper aerodigestive tract.

Liver. Strong for heavy drinking and cirrhosis-related risk; moderate for low-dose risk. The observational literature is consistent, but the confounding by viral hepatitis and metabolic disease is substantial, and the dose-response at low doses is difficult to isolate.

Breast. Moderate. The observational literature is large and internally consistent, with a dose-response that survives adjustment for reproductive factors. The mechanistic case is plausible but indirect. The 2025 MR null is a genuine challenge, though not a refutation. The best summary is that a small causal effect is likely, but its size at low doses is uncertain.

Pancreas, stomach, and other sites. Not established at the same level. Some studies report associations, but the evidence is weaker and more inconsistent. The 2025 MR analysis did not find significant effects at most of these sites.

Verdict and calibrated confidence

The causal claim is secure in its broad outlines and weaker in its quantitative details than the public messaging suggests.

Confidence that alcohol is a cause of cancers of the oral cavity, pharynx, larynx, oesophagus, colorectum, and liver: ~95%. The combination of observational dose-response, mechanistic evidence, genetic support, and the ALDH2 natural experiment is about as strong as cancer epidemiology gets without a randomized trial.

Confidence that a causal effect on breast cancer exists at low doses: ~65–70%. The observational literature is consistent, but the genetic evidence is null and the mechanism is indirect. A small effect remains the best explanation, but the size is uncertain.

Confidence that the dose-response is approximately linear with no threshold for the upper aerodigestive tract and colorectum: ~80%. The data are consistent with linearity down to the lowest measured doses, but the confidence intervals at the bottom of the range are wide, and "no threshold" is an inference from the absence of evidence rather than a measured biological fact.

Confidence that "no safe level" is a useful public-health statement: high, with a caveat. It is defensible as a statement that no threshold has been demonstrated and that the precautionary default is to avoid unnecessary exposure. It is not a statement that one drink per day produces a measurable increase in any individual's risk. The distinction matters because conflating the two invites both false reassurance and false alarm.

The population burden is the least uncertain part of the story. Roughly 3–4% of global cancer cases and deaths are attributable to alcohol, with the largest shares in oesophageal, liver, colorectal, and breast cancers. A substantial fraction arises below two drinks per day, and a smaller but non-negligible fraction below one. They depend on models, but they are large enough to justify prevention efforts.

What would change the answer

Four kinds of evidence would materially move the estimates.

First, a better-instrumented Mendelian randomization of breast cancer, with variants that capture peak exposure and total lifetime dose rather than average intake, and with enough power to detect an OR of 1.05. The current null is compatible with a small effect, and a well-powered replication could resolve it.

Second, natural experiments with cancer endpoints. Minimum unit pricing in Scotland, changes in Russian alcohol policy in the 2000s, and dry-law variations in India and Brazil provide exogenous shifts in consumption that could be linked to cancer registries. These designs avoid the self-report problem and the genetic-instrument limitations, though they require long follow-up.

Third, better biomarkers of exposure and effect. Acetaldehyde-DNA adducts, ethyl glucuronide in hair, and one-carbon metabolites could distinguish light from heavy drinkers more accurately than surveys and could test whether the low-dose association is real or a measurement artifact.

Fourth, longer follow-up of existing cohorts. Cancer latency is measured in decades, and most low-dose estimates rely on follow-up that is still short. A twenty-year extension of the Million Women Study or EPIC would sharpen the absolute risk estimates where the policy debate lives.

Until then, the evidence supports a narrower claim than either side prefers: alcohol is a proven carcinogen where it and its metabolite contact tissue, the risk rises with dose, no threshold has been demonstrated, and the individual risk from light drinking is small but not zero. "No safe level" is a reasonable precautionary conclusion. It is not a substitute for the numbers.

References

1. International Agency for Research on Cancer. Personal habits and indoor combustions. IARC Monographs on the Evaluation of Carcinogenic Risks to Humans, Volume 100E. Lyon: IARC; 2012. https://publications.iarc.who.int/Book-And-Report-Series/Iarc-Monographs-On-The-Identification-Of-Carcinogenic-Hazards-To-Humans/Personal-Habits-And-Indoor-Combustions-2012
2. International Agency for Research on Cancer. Alcohol: a major preventable cause of cancer. IARC Evidence Summary Brief No. 6. 8 October 2025. https://www.iarc.who.int/wp-content/uploads/2025/10/IARC_Evidence_Summary_Brief_6.pdf
3. Rumgay H, Shield K, Charvat H, et al. Global burden of cancer in 2020 attributable to alcohol consumption: a population-based study. Lancet Oncology. 2021;22(8):1071–1080. doi:10.1016/S1470-2045(21)00279-5. PMID 34270924.
4. Fink H, Langselius O, Vignat J, et al. Global and regional cancer burden attributable to modifiable risk factors to inform prevention. Nature Medicine. 2026;32(4):1306–1315. doi:10.1038/s41591-026-04219-7. PMID 41634393.
5. Danpanichkul P, Pang Y, Díaz LA, et al. Alcohol-attributable cancer: update from the Global Burden of Disease 2021 study. Alimentary Pharmacology & Therapeutics. 2025;62(1):33–47. doi:10.1111/apt.70163. PMID 40287931.
6. Office of the Surgeon General. Alcohol and Cancer Risk. Washington, DC: US Department of Health and Human Services; 2025. https://www.hhs.gov/sites/default/files/oash-alcohol-cancer-risk.pdf
7. Allen NE, Beral V, Casabonne D, et al. Moderate alcohol intake and cancer incidence in women. Journal of the National Cancer Institute. 2009;101(5):296–305. doi:10.1093/jnci/djn514. PMID 19244173.
8. Anderson BO, Berdzuli N, Ilbawi A, et al. Health and cancer risks associated with low levels of alcohol consumption. Lancet Public Health. 2023;8(1):e6–e7. doi:10.1016/S2468-2667(22)00317-6. PMID 36603913.
9. Larsson SC, et al. Alcohol consumption and risk of cancer: a Mendelian randomization analysis of four biobanks and consortia. BMC Medicine. 2025;23:676. doi:10.1186/s12916-025-04543-8. PMID 41398582. PMCID PMC12707013.
10. Holmes MV, Dale CE, Zuccolo L, et al. Association between alcohol and cardiovascular disease: Mendelian randomisation analysis based on individual participant data. BMJ. 2014;349:g4164. doi:10.1136/bmj.g4164. PMID 25011450.
11. Zhao J, Stockwell T, Naimi T, et al. Association between daily alcohol intake and risk of all-cause mortality: a systematic review and meta-analyses. JAMA Network Open. 2023;6(3):e236185. doi:10.1001/jamanetworkopen.2023.6185. PMID 37000449.
12. GBD 2020 Alcohol Collaborators. Population-level risks of alcohol consumption by amount, geography, age, sex, and year: a systematic analysis for the Global Burden of Disease Study 2020. The Lancet. 2022;400(10347):185–235. doi:10.1016/S0140-6736(22)00847-9. PMID 35843246.
13. Brooks PJ, Enoch MA, Goldman D, et al. The alcohol flushing response: an unrecognized risk factor for esophageal cancer from alcohol consumption. PLoS Medicine. 2009;6(3):e1000050. doi:10.1371/journal.pmed.1000050. PMID 19320537.
14. Rovira P, Rehm J. Estimation of cancers caused by light to moderate alcohol consumption in the European Union. European Journal of Public Health. 2021;31(3):591–596. doi:10.1093/eurpub/ckaa236. PMID 33338220.