Nucleus of the dietary-acrylamide cluster and the wiki’s opening on a heavily-hyped foodborne
toxin whose human cancer evidence is null — the archetype the symmetric-standards, streetlight,
and attention-is-an-anti-signal rules were written for. Acrylamide forms when starchy foods are
cooked above ~120 C in low moisture (frying, baking, roasting) — crisps, chips, bread crust, biscuits,
roasted coffee — via the Maillard reaction. It is a rodent carcinogen and IARC group 2A (probable)
on mechanistic/animal grounds, which drives recurrent public alarm (California Prop 65, EFSA advisories,
don’t burn your toast). This page holds the human dietary-outcome evidence against that alarm.
Layer-1 verdict, upfront: across 16 studies and 1.15M people, higher dietary acrylamide was not
associated with any of ~20 site-specific cancers, with no dose-response threshold — and the
one comparison that matters for hype (typical human intake vs the animal-toxic dose) spans ~4-5
orders of magnitude. For a reasonably-healthy person this is a near-zero lever: much-discussed,
small/uncertain-to-null effect — the anti-signal pattern exactly. confidence: moderate — a single
gold dose-response MA, its own binding constraint being exposure measurement error (below), so the
null is well-supported for the studied range but cannot fully exclude a small true effect masked by
attenuation.
The evidence: null at every site (Filippini 2022)
Filippini pooled 31 papers (16 distinct studies; case-control, case-cohort, and prospective cohort) totalling 1,151,189 participants, of whom 48,175 developed cancer over a median 14.9-year follow-up (range 7.3-33.9). Non-occupationally-exposed adults; dietary acrylamide assessed by food frequency questionnaire in every included study, mean/median intake 23 µg/day. Studies had to adjust at least for smoking. Female reproductive cancers (breast/endometrial/ovarian) were excluded here — carved off into a separate prior dose-response MA (ref 42, which reported a «relatively linear increased risk for ovarian and endometrial cancer» — a source not held here) (Filippini et al., 2022).
Summary conclusion: «high dietary acrylamide exposure was not associated with an increased risk of site-specific non-gynecological cancer.» (Filippini et al., 2022)
Per-site pooled RR, highest vs lowest intake (all subjects)
Every site’s confidence interval crosses 1.0 — not one site reaches significance in the pooled all-subjects analysis. Do not read these as protective or harmful; they are null. N = number of risk estimates pooled (Filippini et al., 2022):
| Site | N | RR (95% CI) | Site | N | RR (95% CI) |
|---|---|---|---|---|---|
| Oral cavity | 2 | 0.99 (0.67-1.46) | Liver | 1 | 1.08 (0.87-1.34) |
| Oro/hypopharynx | 1 | 0.61 (0.33-1.12) | Head and neck | 1 | 0.74 (0.50-1.09) |
| Esophageal | 5 | 1.05 (0.85-1.29) | Laryngeal | 2 | 1.10 (0.79-1.54) |
| Stomach | 2 | 0.92 (0.82-1.05) | Lung | 2 | 0.91 (0.64-1.28) |
| Colorectal | 7 | 0.94 (0.87-1.02) | Thyroid | 1 | 1.33 (0.70-2.53) |
| Colon | 4 | 0.96 (0.85-1.09) | Bladder/urothelial | 3 | 0.89 (0.74-1.07) |
| Rectal | 4 | 0.99 (0.84-1.18) | Renal | 7 | 1.08 (0.93-1.26) |
| Pancreatic | 5 | 0.88 (0.77-1.02) | Prostatic | 7 | 1.00 (0.93-1.07) |
| Lymphoma | 2 | 1.08 (0.95-1.22) | Multiple myeloma | 2 | 0.97 (0.58-1.64) |
| Melanoma | 1 | 1.18 (0.72-1.96) | Brain | 1 | 0.87 (0.54-1.41) |
The classically-flagged sites are the ones to watch, and they are null here. Renal (kidney) — the site earlier MAs flagged — is 1.08 (0.93-1.26), null. This actively overturns the prior landmark: Pelucchi’s 2015 MA found renal-cell RR 1.20 (1.00-1.45); Filippini nulls it by adding three recent US and Japanese cohorts (heterogeneous, with the two US cohorts carrying ~10% weight and pulling the estimate down) (Filippini et al., 2022). Kidney, endometrial, and ovarian are the sites the mechanism (hormonal/genotoxic) predicts; the non-gynecological ones here do not confirm it, and the gynecological signal lives in a separate source not yet held.
Dose-response: no threshold, no knee — a genuinely informative null
The one-stage dose-response MA (cubic splines, knots at the 10th/50th/90th percentiles; reference 23 µg/day; linear trend fitted alongside) found «a null association with no thresholds for all site-specific cancer» and, as «a novel finding, we found no thresholds between different levels of dietary acrylamide and the risk of any of the site-specific cancers considered.» (Filippini et al., 2022)
This is the right question asked and answered (-> The U-Shaped Association Artifact vocabulary): where is the knee / threshold / upper bound? There isn’t one located over the studied intake range 6.8-44.1 µg/day (up to ~0.7 µg/kg body weight/day). Per the corpus rule, a measured-flat curve is weak evidence of a true flat curve — measurement error can hide a knee — so this is no threshold shown, not no threshold exists; but the direction of that caveat only weakens an already-null signal, it cannot manufacture harm (-> Measurement Error in Dietary Assessment).
The two subgroup exceptions — both fragile
- Lung cancer in ever-smokers: RR 1.16 (1.03-1.31), I2 0.0 — the one significant estimate in the whole paper. Almost certainly residual confounding by smoking: acrylamide is in tobacco smoke (smokers carry 3-4x higher acrylamide hemoglobin adducts), so dietary-acrylamide adjustment cannot fully separate the two, and smoking is the dominant lung-cancer cause. Filippini flag it as needing cautious interpretation (Filippini et al., 2022).
- Lymphoma, Western studies only: RR 1.12 (0.99-1.22) — CI touches 1.0, based on a single study combining clinically distinct lymphoma subtypes; the authors call it a signal to interpret «with caution» (Filippini et al., 2022). Not a finding.
Why the null is credible: the dose gap and the mechanism-vs-outcome split
The exposure sits 4-5 orders of magnitude below where acrylamide is a carcinogen in animals. Human dietary intake is ~0.7 µg/kg bw/day at the top; the relevant toxicological anchors are far above it: «compared to the levels of acrylamide observed to be toxic in animal studies (50 mg/kg body weight/day) … acrylamide does not generate any toxicologically detrimental effects when male rats were administrated three low oral doses of acrylamide (20, 40, and 90 µg/kg body weight/day)» and «the EFSA CONTAM Panel selected BMDL10 value of 0.17 mg/kg body weight/day for neoplastic effects in mice … i.e., much higher level compared with those generally experienced by humans.» (Filippini et al., 2022) The animal-toxic dose (50 mg/kg) is ~70,000x the top human dietary dose; the EFSA neoplastic BMDL10 (0.17 mg/kg) is ~240x it. This is the mechanism-directional rule in action: a rodent carcinogen at rodent doses says nothing about a human at human doses — animal != human, and dose is the whole argument (net-effect-not-intended: whole-organism metabolism, food-matrix co-exposures, and dose all sit between the pathway and the outcome).
The mechanism is real but speaks only to direction, not to a human outcome. Acrylamide «is classified as probably carcinogenic to humans (class 2A) by the International Agency for Research on Cancer» (Filippini et al., 2022); its metabolite glycidamide is genotoxic (a glycidamide-related mutational signature appears in ~1/3 of ~1,600 human tumor genomes across 19 cancer types), with proposed hormonal (non-genotoxic) pathways too (Filippini et al., 2022). Filippini reports that «in vitro and animal studies [that] have consistently shown … genotoxic, mutagenic, and carcinogenic» effects stand against inconsistent human epidemiology (Filippini et al., 2022). The source does not state what evidence the IARC 2A classification itself rests on; the reading below is the wiki’s inference: the mechanistic/animal carcinogenicity is what a 2A (probable) hazard identification turns on, whereas the human dietary-outcome data — exactly what Filippini supplies and finds null — is a different question (hazard identification vs dietary-dose risk), so the two do not conflict.
The binding constraint and the confounder — and why they cut toward the null, not away
Dietary acrylamide carries a double measurement error (-> Measurement Error in Dietary Assessment). The exposure = FFQ-reported food intake x a food-acrylamide-content database — two error layers stacked: «the summary estimates may be affected by misclassification of exposure related to the self-reported dietary intake … [and] the acrylamide food database used in the separate studies, will not fully capture the variations in acrylamide levels between brands of a given food and in different food categories, as well as the different cooking methods used at home by the participants.» (Filippini et al., 2022) Acrylamide content varies enormously with the specific food, batch, and browning — a database mean cannot see how dark your toast is.
- But two facts stop this from rescuing a hidden effect. (1) A biomarker validation: dietary-acrylamide estimates correlate significantly with acrylamide hemoglobin adducts (an objective internal-dose marker), so the FFQ x database estimate is not pure noise. (2) The one study using the biomarker directly — adducts, not FFQ — for prostate cancer «showed no association in a Swedish population» (Filippini et al., 2022). So even removing the FFQ error layer, the signal stays null. Measurement error attenuates toward null in the simple case, so it is the reason to hold a small residual gap open — not evidence of a masked harm.
- Confounding runs one clear direction: smoking. All included studies were rated moderate risk of bias for confounding («All studies had a moderate risk of bias for confounding.») (Filippini et al., 2022). High-acrylamide diets (fried/processed starchy foods) also correlate with smoking and a generally less-healthy dietary pattern; smoking adjustment was mandatory but imperfect (it produced the spurious-looking smoker lung signal above). Residual confounding here would bias toward a false positive, so the observed null is if anything conservative.
Synthesis
A symmetric-standards test the fad fails — demote to no-meaningful-effect for the general population
Acrylamide is the worked instance of the demarcation rule: a fad exposure gets the same bar as any other, and is demoted to no meaningful effect when the human evidence does not hold. The mechanistic and animal case (genotoxic metabolite, IARC 2A, rodent tumors) is real but sits in the directional / not-yet bucket at human dietary doses — and the human dose is ~240x below even the EFSA neoplastic benchmark. The epidemiology, when finally pooled with a dose-response model, is null across every non-gynecological site and shows no threshold. Believing a protective or harmful arm here would require ignoring the null; the correct state is no meaningful effect for the general population over the studied intake range, with a small residual insufficient-evidence margin from measurement error and from the un-held gynecological-cancer source.
Layer-1 sizing — near-zero, and a case study in attention-as-anti-signal
For a reasonably-healthy person the acrylamide lever ranks near the bottom of Layer 1 - Ranking Interventions for a Stratum: a null-to-tiny, uncertain effect on a low-variance ubiquitous exposure, dwarfed by the actual big rocks in the same cancer table (smoking, adiposity, alcohol -> Diet Physical Activity and Cancer Prevention, Body Fatness and Cancer Risk). The public volume around acrylamide (Prop 65, burnt toast scares) runs inversely to its decision value — the settled big rocks are boring, the contested tiny toxin generates content. The one honest decision-change this page licenses: someone worried about acrylamide from cooking can stop worrying and reallocate attention to a lever that moves the outcome. It is the same shape as Disinfection Byproducts and Bladder Cancer — a much-discussed foodborne/environmental chemical whose individual behavioural lever is negligible — except that acrylamide’s individual signal is null rather than merely small.
Open threads / gaps
- G (source not held): the gynecological-cancer dose-response MA that Filippini carved out (its ref 42; a separate group MA, author not resolved from this source) reported a «relatively linear increased risk for ovarian and endometrial cancer» — the one place the hormonal mechanism may bite. This is the decision-relevant open question (a female stratum), and it is excluded from every number on this page — acquire it to close the gynecological arm.
- G: no MR / genetic-instrument evidence exists for dietary acrylamide (there is no clean natural experiment
for a diet-formed contaminant), so causal identification rests on confounded observational data plus the adduct
biomarker — the residual uncertainty the
moderateconfidence encodes. - The coffee intersection: coffee is a major acrylamide source yet is net-neutral-to-beneficial for most outcomes (-> Coffee Consumption and Health) — a standing illustration that a toxin-bearing food’s matrix effect swamps the isolated contaminant.