Name: Polyphenols: what the evidence supports, and why the antioxidant story collapsed
Creator: Realised
Published: 2026-08-01

Summary

"Polyphenols" is a chemical category rather than a mechanism, the direct free-radical-scavenging story behind the whole category has been retired (the USDA withdrew its own antioxidant-capacity food database in 2012), and the evidence that survives attaches to specific foods and specific compounds — so this entry organises the class, kills the class-level claim, and routes every compound question to the entry that owns it.

Why Moderate

This entry is rated Moderate overall, and the rating describes the class-level entry as a whole rather than any individual compound.

Why not the tier above (Strong). A Strong rating would require consistent randomised human evidence supporting the class's core claim. There is no such thing, because there is no coherent class-level claim to support. The strongest individual result here is the COSMOS null, which is Strong at the level of a specific finding and is reported as such; the cocoa blood-pressure meta-analysis sits at Moderate because its own reviewers rated the evidence moderate certainty. But the entry's subject is the category, and a category cannot inherit the evidence rating of its best-studied member. Two further blocks: the human evidence that does exist is almost entirely on biomarkers rather than clinical outcomes, and the mechanism accounting for those biomarker effects is not established in humans.

Why not the tier below (Emerging). An Emerging rating would understate what is genuinely settled. The negative claims in this entry are Foundational, not tentative: the category is definitional rather than mechanistic; bioavailability is low and well-quantified across 97 human studies; circulating species are conjugates and microbial breakdown products rather than parent compounds; the USDA formally withdrew the antioxidant-capacity database with published reasoning; and the observational literature's confounding structure is well-characterised. There is also a Cochrane-level randomised finding, a large well-powered randomised trial with a clean null, and one formally adjudicated regulatory claim. That is more than an emerging field has.

Per-sub-area split, because the strength is genuinely uneven:
• What the category is — Foundational. Definitional and taxonomic, not an effect estimate.
• Bioavailability and metabolism — Foundational. A 97-study systematic review plus well-established conjugation and microbial-transformation biochemistry. The producer-phenotype prevalence figures are deliberately unstated and would be Emerging if included.
• The retired scavenging mechanism — Foundational as a negative claim. The institutional withdrawal is a documented fact and the concentration arithmetic is standard physiology.
• The hormetic and NRF2 replacement — Moderate. Coherent, convergent cell and animal data, no demonstrated human causal chain, and pervasive concentration and species overreach in the supporting literature.
• Microbiome mediation — Emerging. Composition shifts are demonstrable, outcome links are not, and fibre is a confound that most designs do not separate.
• Cocoa and tea flavanols on blood pressure — Moderate, matching the Cochrane reviewers' own moderate-certainty rating. Blinding is the structural weakness, the pooled effect is driven by hypertensive participants with no significant effect in normotensives, and the tier must not be read as describing effect size.
• COSMOS primary endpoint (null) — Strong. Large, well-powered, prespecified.
• COSMOS cardiovascular-death secondary — Emerging. Secondary endpoint in a trial whose primary failed, uncontrolled multiplicity, internally inconsistent with neutral heart attack and stroke results.
• Olive oil polyphenols on oxidised LDL — Moderate. Regulator-adjudicated, but a small short trial base and a surrogate endpoint.
• The isolated-antioxidant supplement lesson — Strong as evidence about vitamins, explicitly a labelled analogy rather than evidence about polyphenols.
• Exercise adaptation blunting — Strong for vitamins C and E, Emerging as read across to polyphenols.
• Food beats extract — Moderate. A synthesis across verified findings, limited by the fact that extract trials are few, so the comparison is partly an artefact of what has been studied.
• Observational confounding — Foundational as a methodological claim.

Confidence: moderate. We are highly confident in the negative claims and in the citation-level facts reported here. We are moderately confident that the hormetic account is directionally right, and deliberately uncertain about which specific compounds do anything worth paying for — which is why the entry routes rather than rules.

Practical takeaway

This entry does not tell you what to take. It tells you where the evidence sits and which entry to read next. Compounds are grouped by how strong the human evidence is, not by how popular they are.

Rule zero, before any of the routing below. Do not buy anything because it says "antioxidant," "high ORAC," or "polyphenol complex" on the label. Those are class-level claims and the class does not support claims. If a product cannot name the specific compound, the specific dose, and the specific outcome it moved in humans, it is selling the category, not the compound.

Where the food-level evidence is strongest — eat these, and expect small effects
• Cocoa and tea flavanols — lower blood pressure by about 2 mmHg pooled over weeks, at Moderate strength, with the effect concentrated in people whose blood pressure is already raised (around 4 mmHg systolic in hypertensive participants) and no significant difference in people with normal blood pressure. Small effect, no clinical-outcome evidence. Food-level intake, not extract. Tea catechins specifically: go to green_tea_extract_egcg_evidence, which also owns the extract safety question.
• Olive oil polyphenols — a polyphenol claim that survived formal European regulatory adjudication, at Moderate strength, and it reaches a laboratory marker of lipid oxidation rather than a disease outcome. Practical read: choose a genuine extra-virgin oil, expect a real but modest and marker-level benefit. No dedicated spoke; this hub owns the summary above.
• Whole fruit, vegetables, coffee and legumes — the food-level case is much better than the compound-level case, and part of the benefit is fibre, matrix and displacement of worse food rather than the polyphenols themselves. Go to whole_food_emphasis and fruit_whole_food_evidence_and_sugar.

Where compound-specific evidence exists and is worth reading, at varying strength
• Green tea catechins / EGCG — the best-studied single polyphenol, with genuine human data and a genuine extract-specific liver safety signal that makes the food-versus-extract distinction concrete. Go to green_tea_extract_egcg_evidence.
• Curcumin (turmeric) — huge literature, notoriously poor absorption, and an entire bioavailability-engineering industry built on that fact. Judge it on its own human endpoints. Go to turmeric_curcumin_evidence.
• Quercetin — a flavonol with plausible mechanisms and a mixed human record; commonly stacked into immune and longevity products on thin grounds. Go to quercetin_evidence_and_use.
• Spermidine — not a polyphenol at all, but routinely sold alongside them in the same "plant longevity compound" aisle; included here so the boundary is explicit. Go to spermidine_longevity_evidence.

Where the claims run far ahead of the evidence
• Resveratrol — the compound that carried the entire polyphenol-longevity narrative for a decade, on human evidence that never arrived at the level the marketing implied. Go to resveratrol_and_longevity_claims before spending anything.
• Anything sold on an ORAC score or "antioxidant blend" — no spoke, because there is nothing to route to. This is the category the USDA withdrawal was about.
• Methylene blue — not a polyphenol, but frequently marketed inside the same redox and mitochondrial story, so the same reasoning errors apply. Go to methylene_blue_cognition_and_mitochondria.

How to think about dose, if you are going to take something anyway
• If the mechanism really is mild-stress signalling, dose is not monotonic — more is not more, and a high-dose extract may sit past the useful window. See hormesis_and_adaptive_stress.
• Do not take high-dose antioxidant supplements around training if you are training for adaptation. This is established for vitamins C and E and inferential for polyphenol extracts, but the inference points one way and the cost of caution is zero.
• Extract dose labels are frequently uninterpretable — "500 mg turmeric" and "500 mg curcuminoids" are different things by a factor of about twenty. Go to supplement_form_elemental_dose_and_bioavailability before comparing any two products.
• Individual response is genuinely variable and partly determined by your gut bacteria. A compound that does nothing for you may do something for someone else, and neither result generalises.

What to do with this entry as a reader. Use it to disqualify class-level claims quickly, then go to the spoke for anything you are actually considering. If this hub and a spoke disagree about a compound, the spoke is the authority.

Evidence detail

Why This Entry Exists

Almost every supplement bottle that says "antioxidant" is making a class-level argument: this plant extract contains polyphenols, polyphenols are antioxidants, antioxidants fight free radicals, free radicals cause ageing and disease, therefore this bottle slows ageing and disease. Each link in that chain is either wrong or unproven, and the chain as a whole is the single most successful piece of nutritional marketing of the last thirty years.

The problem starts at the first link. Polyphenols are defined by chemical structure — multiple phenolic hydroxyl groups — not by biological action. The category spans flavanols like the catechins in tea, anthocyanins in berries, flavonols like quercetin, isoflavones in soy, phenolic acids like the chlorogenic acid in coffee, stilbenes like resveratrol, and lignans. These are thousands of distinct compounds with radically different absorption, metabolism, molecular targets and potency. Grouping them and then making one health claim about the group is like grouping arsenic and sodium under "elements" and drawing a conclusion about both.

The second link failed publicly. The USDA Agricultural Research Service published an antioxidant-capacity database for common foods — Release 1 in November 2007 covering 277 foods, Release 2 in May 2010 covering 326 — and then withdrew it on 16 May 2012, stating that the values had no relevance to the effects of specific bioactive compounds, including polyphenols, on human health, that they were being routinely misused by food and supplement companies to promote products and by consumers to guide their choices, and that non-antioxidant mechanisms — still undefined — may be responsible for whatever benefits these compounds have. Fourteen years later those scores are still printed on packaging.

But the opposite error is just as common and just as wrong. A certain kind of sceptic reads the low absorption figures and concludes that polyphenols are inert, that it all passes through unabsorbed and gets excreted, and that the entire field is a scam. That overshoots. Low bioavailability is entirely compatible with real signalling effects at low concentrations, some human endpoints genuinely do move, and one polyphenol claim has survived formal European regulatory adjudication. The honest position sits between the two, and it is uncomfortable for both sides: the foods are worth eating, the class-level supplement story is marketing, the mechanism is more likely mild-stress signalling and gut-microbial transformation than radical mopping, and every claim has to be judged compound by compound.

What bad advice this protects against, in all directions:
• "Polyphenols are powerful antioxidants that fight free radicals and slow ageing." — This rests on an obsolete mechanism and an assay the US government withdrew. Plasma concentrations after normal dietary intake sit in the nanomolar to low-single-digit micromolar range, while your own urate runs at hundreds of micromolar and intracellular glutathione at millimolar levels alongside dedicated enzymes. A trace-level, heavily chemically-modified metabolite cannot meaningfully out-compete that system as a stoichiometric radical sponge. The arithmetic alone kills it.
• "This supplement has a high ORAC score, so it's more protective." — ORAC measures how well a food extract quenches a synthetic radical in a test tube. It was never validated against human outcomes, and the body that created it disowned it in 2012 specifically because it was being used this way. A high score tells you about a cuvette.
• "Polyphenols are useless — they're barely absorbed and you just pee them out." — Overshoots badly. Poor absorption rules out the sponge mechanism; it does not rule out signalling, and it does not rule out effects driven by the gut-microbial breakdown products, which is where the field's most interesting current work sits. Cocoa and tea flavanols genuinely lower blood pressure in randomised trials, by a small and replicated amount that is concentrated in people whose blood pressure is already raised.
• "The evidence for polyphenols is strong, so a concentrated extract must work even better." — Two errors stacked. The evidence that survives scrutiny is overwhelmingly about foods and dietary patterns, and the one large randomised trial of a concentrated extract with hard cardiovascular outcomes missed its primary endpoint. Separately, if the mechanism really is mild stress signalling, then more is not better — it is a window you can overshoot.
• "Cocoa flavanols were clinically proven to cut cardiovascular death by 27 percent." — That number is a secondary endpoint inside a trial whose primary endpoint failed, in a family of comparisons where multiplicity was not controlled, sitting next to neutral results for heart attack and stroke. It is a signal worth testing further, not a proven effect.
• "EFSA approved a health claim for olive oil polyphenols, so olive oil is proven to protect your heart." — The authorised claim is about protecting blood lipids from oxidative damage. That is a laboratory marker, not a disease outcome. The regulator adjudicated a biomarker, and the claim is routinely quoted as though it adjudicated heart disease.

What this entry OWNS: the class-level picture. Why "polyphenol" is not a meaningful unit of health evidence; why absorption is low and why what circulates in your blood is largely not what you ate; why the direct-antioxidant framing is obsolete and how the ORAC withdrawal happened; what the modern hormetic and signalling reading proposes and how strong it actually is; the food-versus-extract split; why the observational literature is severely confounded; and a routing map that sends every specific compound question to the entry that owns it.

What this entry DEFERS: every individual compound's evidence base, dosing, safety profile and controversies. Green tea catechins and the extract hepatotoxicity question belong to green_tea_extract_egcg_evidence. Quercetin belongs to quercetin_evidence_and_use. Resveratrol and the longevity claims belong to resveratrol_and_longevity_claims. Curcumin and its bioavailability engineering belong to turmeric_curcumin_evidence. Spermidine belongs to spermidine_longevity_evidence. The general theory of plant defence chemistry belongs to plant_defence_compounds_and_xenohormesis, and the general theory of adaptive stress to hormesis_and_adaptive_stress. This hub does not re-argue any of them, and where this entry and a spoke disagree on a compound, the spoke wins.

Evidence

Read the tiers, not the thesis. This entry's structure is deliberately lopsided: the negative claims (the class is not a mechanism, absorption is low, the antioxidant assay was withdrawn) sit at Foundational, while the positive claims (hormetic signalling, microbiome mediation, extract benefit) sit at Moderate or Emerging. That asymmetry is the finding. It is much better established what polyphenols do not do than what they do.

Sub-area 1: What the category actually is

1. [Foundational] "Polyphenols" names a chemical structure, not a biological action, and so cannot carry a health claim. The category is defined by multiple phenolic hydroxyl groups and spans flavonoids (flavanols and catechins, anthocyanins, flavonols, flavanones, isoflavones), phenolic acids, stilbenes and lignans — thousands of compounds with different absorption, metabolism, targets and potency. Manach C, Scalbert A, Morand C, Rémésy C, Jiménez L. "Polyphenols: food sources and bioavailability." Am J Clin Nutr. 2004;79(5):727-747. Cui bono: the class framing is worth a great deal to supplement marketers and "superfood" brands, because it lets evidence earned by cocoa flavanols or olive oil transfer free of charge to any plant extract containing a phenolic ring. Nobody profits from the compound-level framing, which is part of why it is stated so rarely — but it also hands sceptics a way to dismiss the entire field in one move, which is its own overreach. Note: we do not attach a compound count here. The figure of roughly eight thousand polyphenols circulates widely and is often bolted onto citations that do not contain it; "thousands" is the defensible statement.

Sub-area 2: What actually reaches your tissues

2. [Foundational] Absorption is low, and the concentrations achievable in human blood sit well below those used in the laboratory studies that built the antioxidant story. The systematic review of 97 human bioavailability studies reports that plasma concentrations of total metabolites ranged from 0 to 4 µmol/L after an intake of 50 mg aglycone equivalents, with urinary recovery of 0.3 to 43 percent of the ingested dose depending on the compound. Absorption varies enormously by compound: gallic acid and isoflavones are the best absorbed, followed by catechins, flavanones and quercetin glucosides, while proanthocyanidins, the galloylated tea catechins and anthocyanins are the least well absorbed. The mechanistic and antioxidant literature overwhelmingly uses concentrations of 10 to 100 µmol/L. A test-tube result at 50 µmol/L is not evidence about a human at 1 µmol/L — but state the gap accurately: it is typically one to two orders of magnitude, not the three sometimes claimed, and at the top of the achievable plasma range it is less than one. The argument does not need the exaggeration. Manach C, Williamson G, Morand C, Scalbert A, Rémésy C. "Bioavailability and bioefficacy of polyphenols in humans. I. Review of 97 bioavailability studies." Am J Clin Nutr. 2005;81(1 Suppl):230S-242S. Citation fidelity note: the widely repeated formulation "plasma concentrations rarely exceed 1 µmol/L after consumption of 10 to 100 mg of a single phenolic compound" is routinely attributed to this group's reviews, but was not located in the abstract of either the 2004 or the 2005 paper during verification for this entry. We state the figures the abstract does carry and do not reproduce the unlocated sentence. Cui bono: the low-absorption framing is commercially useful to firms selling "enhanced bioavailability" formulations — phytosome, liposomal, piperine-boosted — who need the absorption problem to be real and solvable by their product. It is inconvenient for whole-extract sellers, since it undercuts dose-response marketing. And it is over-used by reductionist critics to declare the whole field void.

3. [Foundational] What circulates in your blood after eating polyphenols is largely not what you ate. Absorbed polyphenols are extensively glucuronidated, sulfated and methylated in the gut wall and liver, so the free parent molecules are scarce or absent in plasma. The larger fraction is never absorbed intact at all — it reaches the colon and is broken down by gut bacteria into smaller phenolic acids and specific breakdown products. Two well-characterised examples: the ellagitannins in pomegranate, walnuts and berries are converted by gut bacteria into urolithins, and only some people carry the microbial capacity to make urolithin A; soy daidzein is converted to equol only in a minority of Western adults. Manach C, Williamson G, Morand C, Scalbert A, Rémésy C. Am J Clin Nutr. 2005;81(1 Suppl):230S-242S. Cui bono: companies selling the pre-formed breakdown products (urolithin A supplements, S-equol) benefit directly — this framing is the entire commercial rationale for their product category. Brands selling the parent food or extract prefer the simpler "you eat it, it works" story. The metabolite science is real and also happens to have been productised. We state the producer-phenotype prevalences qualitatively ("a minority of people") rather than numerically; the commonly quoted percentages are not in the cited paper and were not independently verified for this entry.

Sub-area 3: The mechanism that died

4. [Foundational] The direct free-radical-scavenging story is obsolete, and the body that built its most famous measurement withdrew it. ORAC (Oxygen Radical Absorbance Capacity) is an in-vitro chemical assay measuring how well a food extract quenches a synthetic radical in a cuvette. The USDA Agricultural Research Service published ORAC values for common foods from November 2007 (Release 1, 277 foods; Release 2 in May 2010, 326 foods) and withdrew the database on 16 May 2012, stating that the values have no relevance to the effects of specific bioactive compounds, including polyphenols, on human health, that ORAC values are routinely misused by food and dietary supplement companies to promote their products and by consumers to guide their food and supplement choices, that evidence had mounted against the assay's biological significance, and that undefined non-antioxidant mechanisms may be responsible for any benefits. USDA Agricultural Research Service, Nutrient Data Laboratory, "Withdrawal of the USDA Database for the Oxygen Radical Absorbance Capacity (ORAC) of Selected Foods, Release 2," 16 May 2012; database compiled by Haytowitz DB, Bhagwat S. Cui bono: ORAC scores remain commercially valuable because they are cheap to generate and produce a large, impressive number — they still appear on açaí, mangosteen and "antioxidant blend" packaging fourteen years after the government that created them disowned them. The USDA had no commercial stake in the withdrawal and took reputational damage for retracting a database it had published and updated for five years, which is a costly signal in favour of the withdrawal being honest. What the withdrawal does not say is that polyphenols are useless — only that in-vitro antioxidant capacity does not index human benefit. Over-reading it in the sceptical direction is the mirror-image error.

5. [Foundational] The arithmetic independently rules out the sponge model. Plasma polyphenol concentrations after food sit in the nanomolar to low-single-digit micromolar range, while endogenous antioxidant systems operate far above that — urate at roughly 200 to 400 µmol/L, intracellular glutathione at millimolar concentrations, plus dedicated enzyme systems. A trace-level, heavily conjugated molecule cannot meaningfully out-compete that as a stoichiometric radical scavenger. These physiological concentrations are standard textbook values stated generically here; they are not drawn from the USDA statement or from any polyphenol paper, and should not be attributed to either. Cui bono: no commercial party benefits from the arithmetic being stated, which is roughly why it is absent from every product page in the category.

Sub-area 4: The mechanism that replaced it

6. [Moderate] The leading modern account is that dietary "antioxidants" act as mild stressors and signalling molecules rather than as scavengers — transiently and mildly oxidising cellular sensors, which activates the KEAP1-NRF2 pathway and induces the cell's own antioxidant and detoxification enzymes (glutathione S-transferases, NAD(P)H quinone dehydrogenase 1, heme oxygenase-1). A related proposal is that polyphenols auto-oxidise in cell-culture medium to generate low levels of hydrogen peroxide, which inhibits protein tyrosine phosphatases and drives the same signalling. Forman HJ, Davies KJA, Ursini F. "How do nutritional antioxidants really work: nucleophilic tone and para-hormesis versus free radical scavenging in vivo." Free Radic Biol Med. 2014;66:24-35. Species and concentration overreach is the dominant problem in this literature and must be stated plainly: the great majority of polyphenol-NRF2 evidence is cell culture at 10 to 100 µmol/L — above achievable human plasma levels — or rodent gavage at doses that do not scale to human diets. A rodent NRF2 result is never human proof. The hydrogen-peroxide route in particular is partly a culture artefact, since polyphenols auto-oxidise in bicarbonate-buffered media in ways they may not in a living body. Human evidence that NRF2 activation is the operative mechanism for any measured human outcome is thin. Moderate here means mechanistic plausibility plus convergent cell and animal data — not a demonstrated human causal chain. Cui bono: hormesis is a rescue narrative for an industry whose original mechanism was retired — "the antioxidants work, just differently" preserves the product line while conceding the science, and it keeps academic funding alive in a field that would otherwise be closing. Against that, the framing is genuinely inconvenient for supplement dosing, because if the mechanism is mild stress then high-dose extracts may overshoot the useful window, which is exactly what the antioxidant-supplement failures suggest.

7. [Emerging] Polyphenols act as substrate for the gut microbiota and reshape its composition, while the microbiota in turn determines which bioactive breakdown products are produced. This is a coherent and increasingly favoured explanation for how a poorly absorbed class could still exert systemic effects. Composition shifts are demonstrable in human intervention studies; the chain from a composition shift to a health outcome that matters is not established for any polyphenol. The microbial-transformation half of this claim is supported by Manach et al. 2005 above; the prebiotic composition-shift half is stated here without a specific citation because it was not independently verified for this entry, and should be verified before any stronger statement is made. Cui bono: the postbiotic and urolithin supplement sector and academic microbiome laboratories both benefit from this being the leading mechanism; proponents of the old direct-antioxidant story lose their explanatory monopoly. Two cautions. Microbiome mechanisms currently function as the field's universal explanatory solvent — when a mechanism is unknown, "it's the microbiome" fills the gap without adding evidence. And polyphenol-rich diets are fibre-rich diets, with fibre the far better-established driver of microbial change, so attributing a shift to the polyphenols specifically needs a design that separates them, which most studies do not have.

Sub-area 5: The human evidence, precisely

8. [Moderate] Cocoa and tea flavanols lower blood pressure by a small amount — pooled at about 2 mmHg, and concentrated in people whose blood pressure is already raised. A Cochrane review pooled 35 trials contributing 40 treatment comparisons in 1,804 mainly healthy adults given 30 to 1,218 mg flavanols per day (mean around 670 mg) for 2 to 18 weeks (mean nine weeks). Pooled effect: systolic −1.76 mmHg (95% CI −3.09 to −0.43, P = 0.009) and diastolic −1.76 mmHg (95% CI −2.57 to −0.94, P < 0.001). (The two point estimates really are identical to two decimal places in the published abstract, with different confidence intervals — this is a rounding coincidence, not a transcription error, and should not be flagged as one.) Tier reasoning stated openly: the review's own certainty rating is moderate, downgraded from high for unexplained heterogeneity, so this entry rates the finding Moderate rather than Strong. Calling a source's moderate-certainty conclusion Strong would be tier inflation. The subgroup structure matters more than the headline. In hypertensive participants (9 comparisons, 401 participants) systolic pressure fell by about 4 mmHg. In pre-hypertensive participants (8 comparisons, 340 participants) it only tended to be lowered. In normotensive participants (23 comparisons, 1,063 participants) there was no significant difference. So the pooled ~2 mmHg is not a number a person with normal blood pressure should expect to see in themselves, and the interaction with baseline blood pressure was borderline, so treat the subgroup split as hypothesis-generating rather than settled. Ried K, Fakler P, Stocks NP. "Effect of cocoa on blood pressure." Cochrane Database Syst Rev. 2017;4:CD008893. Read this tier carefully: Moderate describes confidence that a small effect exists, not the size or importance of the effect. That is the single most likely misreading of this whole entry. Two mmHg is roughly what a modest reduction in salt intake achieves, and roughly a fifth to a quarter of what a single antihypertensive drug typically delivers. Blinding is the structural weakness — cocoa tastes like cocoa, and many trials used low-flavanol chocolate as the control, an imperfect placebo. All trials were short, and none reported clinical outcomes. Cui bono: Mars Inc. has funded a very large share of the cocoa flavanol literature and holds cocoa-flavanol processing patents, so the shape of this field reflects that. On the other side, pharmaceutical antihypertensive interests benefit from cocoa looking trivial, and public-health bodies are institutionally wary of anything that lets chocolate be reframed as medicine given its sugar and energy load.

9. [Strong] The largest cocoa flavanol trial ever run missed its primary endpoint. COSMOS randomised 21,442 US adults (12,666 women aged 65 and over, 8,776 men aged 60 and over) in a two-by-two factorial design to cocoa extract (500 mg cocoa flavanols daily, including 80 mg epicatechin) and/or a multivitamin. Over a median 3.6 years, the primary endpoint — a composite of total cardiovascular events including heart attack, stroke, coronary revascularisation, cardiovascular death, carotid artery surgery, peripheral artery surgery and unstable angina — occurred in 410 cocoa-extract versus 456 placebo participants: hazard ratio 0.90 (95% CI 0.78 to 1.02, P = 0.11). Not statistically significant. Sesso HD, Manson JE, Aragaki AK, et al. "Effect of cocoa flavanol supplementation for the prevention of cardiovascular disease events: the COcoa Supplement and Multivitamin Outcomes Study (COSMOS) randomized clinical trial." Am J Clin Nutr. 2022;115(6):1490-1500. A null on a well-powered primary endpoint is strong evidence against a large effect, not proof of no effect — the confidence interval remains compatible with up to a 22 percent relative reduction, follow-up was short for a prevention question, and adherence declines over time. Cite this trial for what it found: cocoa extract did not significantly reduce total cardiovascular events. Do not cite it as showing that cocoa prevents cardiovascular disease. Cui bono: nobody benefits commercially from the null being the headline, which is why it almost never is. The trial was funded by Mars Edge, which supplied the cocoa extract and provided infrastructure support, by Pfizer Consumer Healthcare for the multivitamin arm, and by the National Institutes of Health — state all three rather than only the industry pair, since selective reporting of the funding is its own bias. Press coverage overwhelmingly led with the secondary result. To the investigators' credit, the paper states plainly that the primary outcome was not significantly reduced.

10. [Emerging] The famous COSMOS cardiovascular-death result is a secondary endpoint inside a trial whose primary failed. Among prespecified secondary outcomes, cardiovascular death fell: hazard ratio 0.73 (95% CI 0.54 to 0.98), a 27 percent relative reduction. The other cardiovascular secondaries were neutral — heart attack 0.87 (0.66 to 1.16), stroke 0.91 (0.70 to 1.17), all-cause mortality 0.89 (0.77 to 1.03). A per-protocol analysis censoring for non-adherence gave 0.85 (0.72 to 0.99) for total cardiovascular events. The authors themselves call for cautious interpretation and longer follow-up. Sesso HD, Manson JE, Aragaki AK, et al. Am J Clin Nutr. 2022;115(6):1490-1500, secondary and per-protocol analyses. Three problems compound here: when a primary endpoint fails, secondary endpoints are hypothesis-generating by convention rather than confirmatory; multiple secondary comparisons inflate the chance of one crossing significance; and it is mechanistically odd that cardiovascular death fell 27 percent while heart attack and stroke — the events that cause cardiovascular death — did not move. That internal inconsistency is the strongest reason for caution. The per-protocol analysis is not a fix, because adherent participants differ systematically from non-adherent ones in ways that independently predict survival. Cui bono: the cocoa extract supplement market was built substantially on this one hazard ratio, and "clinically proven to reduce cardiovascular death" language traces directly to it. But over-correcting to "COSMOS was negative, cocoa does nothing" also overshoots — the point estimates sat below 1.0 across outcomes, which is what a real-but-underpowered small effect looks like.

11. [Moderate] A polyphenol claim has survived formal European regulatory adjudication — and what it reaches is a laboratory marker, not a disease. EFSA's panel gave a favourable opinion and the claim was authorised under Commission Regulation (EU) No 432/2012 in the form: olive oil polyphenols contribute to the protection of blood lipids from oxidative stress. Conditions of use require at least 5 mg of hydroxytyrosol and its derivatives (for example the oleuropein complex and tyrosol) per 20 g of olive oil, with the consumer told that the effect is obtained at a daily intake of 20 g of olive oil. Commission Regulation (EU) No 432/2012; EFSA NDA Panel scientific opinion on olive oil polyphenols and protection of LDL particles from oxidative damage, EFSA Journal 2011. Three caveats travel with this. The endpoint is oxidised LDL — a surrogate whose causal role in cardiovascular events is not established and whose assays are contested; EFSA authorised a biomarker claim, not a heart-disease-prevention claim. The underlying trial base is small and short, and we state that qualitatively: the specific characterisation sometimes given for it (one well-conducted trial plus two smaller dose-response studies over roughly three weeks) was not independently verified for this entry and should not be reproduced as fact. And the bar EFSA applied is a regulatory bar, not a clinical-outcome bar. The trial commonly cited as the underpinning is EUROLIVE (Covas et al., Ann Intern Med, 2006), an attribution we have not independently verified and to which no sample size or effect size should be attached from memory. Cui bono: the European olive oil industry lobbied for and directly monetises this claim — high-phenolic oils command a premium justified almost entirely by it, and an analytical-certification sub-industry exists to serve it. Refined-olive-oil and seed-oil producers lose, since refining strips the phenolics. Note the asymmetry: the claim's existence tells you more about EFSA's process than about olive oil's power, and it is routinely quoted as if EFSA endorsed cardiovascular protection, which it did not.

Sub-area 6: What the supplement era already taught us

12. [Strong] High-dose isolated antioxidant supplementation has repeatedly failed, and in some trials caused harm. The SELECT trial randomised over 35,000 healthy men and found that 400 IU/day of vitamin E alone increased prostate cancer incidence by about 17 percent relative — in absolute terms, 76 cases per 1,000 men on vitamin E versus 65 per 1,000 on placebo over seven years. This was the opposite of the hypothesised effect, in a trial designed and powered to demonstrate prevention. Klein EA, Thompson IM Jr, Tangen CM, et al. "Vitamin E and the risk of prostate cancer: the Selenium and Vitamin E Cancer Prevention Trial (SELECT)." JAMA. 2011;306(14):1549-1556. Critical fidelity point: vitamin E is not a polyphenol. SELECT is evidence about the strategy of extracting a chemical from food, raising the dose far above dietary levels and giving it for years. It is not evidence about polyphenol extracts specifically, and framing it as "antioxidant supplements including polyphenols cause cancer" would be citing a trial for a conclusion it does not reach. Used here as a labelled prior and analogy. The beta-carotene trials in smokers are the other canonical example of the same pattern; we have not verified their details for this entry and so name no author, journal or effect size for them. Cui bono: pharmaceutical and mainstream-medical interests benefit whenever supplements look useless. The supplement industry's standard rebuttal is that these trials used the wrong isomer, wrong dose or synthetic form — not entirely wrong, since synthetic all-rac-alpha-tocopherol is not food vitamin E, but a rebuttal deployed against every negative trial for twenty-five years, which makes it unfalsifiable in practice.

13. [Strong] as evidence about vitamins, [Emerging] as evidence about polyphenols: high-dose antioxidant supplementation blunts exercise adaptation, which is the clearest practical demonstration that oxidative stress is a signal rather than only damage. Ristow and colleagues showed that vitamin C (1 g/day) plus vitamin E during four weeks of training reduced exercise-induced expression of PGC-1alpha and of endogenous antioxidant enzymes, and prevented the exercise-induced improvement in insulin sensitivity. Paulsen and colleagues, in a double-blind randomised controlled trial, found that vitamin C plus vitamin E during endurance training blunted the training-induced rise in mitochondrial markers including COX4 and cytosolic PGC-1alpha, though whole-body performance measures moved less clearly. Ristow M, Zarse K, Oberbach A, et al. "Antioxidants prevent health-promoting effects of physical exercise in humans." Proc Natl Acad Sci U S A. 2009;106(21):8665-8670. Paulsen G, Cumming KT, Holden G, et al. "Vitamin C and E supplementation hampers cellular adaptation to endurance training in humans: a double-blind, randomised, controlled trial." J Physiol. 2014. Two fidelity warnings. These are vitamin C and E trials, not polyphenol trials — the read-across to polyphenol supplements is inferential and is labelled Emerging when stated about polyphenols. And there are two separate Paulsen 2014 papers in J Physiol from the same group, one on endurance and one on strength training, with different conclusions; the strength paper found altered signalling without an effect on muscle growth, and the two must not be merged. We have deliberately omitted volume and page numbers for Paulsen rather than risk assigning the wrong ones, and we do not print an exact vitamin E dose for the Paulsen trial because secondary sources disagree. Cui bono: nobody benefits commercially from this finding — it is inconvenient for the entire sports-antioxidant category, which is a point in its favour. Against that, trainers over-extend it into "never take any antioxidant," including during illness or genuine deficiency, which is not what these trials show; they concern chronic high-dose supplementation around training, not dietary polyphenol intake from food.

Sub-area 7: Food versus extract, and why the observational data cannot carry the weight

14. [Moderate, synthesis rather than single study] The evidence that survives scrutiny attaches overwhelmingly to polyphenol-rich foods and dietary patterns rather than to concentrated extracts. Every durable finding above attaches to a food matrix or a food-level intake — cocoa and tea flavanols at food-realistic doses, olive oil at 20 g/day, patterns rich in fruit, vegetables, tea, coffee and legumes. The one large extract trial with hard outcomes missed its primary endpoint. Isolated high-dose extracts also strip out the fibre, the matrix effects, the slower release and the accompanying nutrients that may be doing part of the work, and concentrate compounds to doses no diet delivers, which is where toxicity appears. The canonical example — a compound that is safe as a beverage and has caused liver injury as a concentrated extract — is owned in full by green_tea_extract_egcg_evidence; route there rather than re-litigating it here. This is a synthesis of the verified findings above, not a single-study result, and it carries an honest limitation: "food beats supplement" is partly an artefact of what has been studied, since food-based epidemiology is vast and extract trials are few, so absence of extract evidence is partly absence of extract trials. Cui bono: the food framing suits whole-food producers, dietitians and public health bodies, and disadvantages the extract and "standardised polyphenol complex" industry, which depends on transferring food-derived evidence to capsules. It is also convenient for Realised's own philosophical lean toward real food, which is a reason to scrutinise it harder rather than less — some individual extracts do have real compound-specific evidence, and the spokes say so.

15. [Foundational] The observational evidence linking polyphenol-rich diets to better health is heavily confounded and cannot carry the weight placed on it. Cohort studies consistently associate higher intake of tea, coffee, berries, cocoa, olive oil and vegetables with lower cardiovascular and all-cause mortality. But polyphenol intake is among the most socially patterned dietary exposures there is: people who eat more of these foods are wealthier, more educated, less likely to smoke, more physically active, more likely to attend screening, and eat less ultra-processed food. Residual confounding after statistical adjustment is expected to be substantial and its direction is predictable — it inflates apparent benefit. The healthy-user effect is not a technicality here; it is plausibly most of the association. This is a methodological claim rather than a single study; it is best supported by the pattern in this entry, where large observational benefit signals for cocoa, tea and olive oil sit next to a null primary endpoint in the one large randomised extract trial (Sesso et al., COSMOS, Am J Clin Nutr 2022). There is a second, field-specific problem: polyphenol intake is usually estimated by mapping food-frequency questionnaires onto composition databases, while the polyphenol content of any given food varies enormously by cultivar, ripeness, storage, processing and brewing — so exposure misclassification is severe before confounding even enters. Cui bono: everyone selling a polyphenol-rich product benefits from observational hazard ratios being reported as effects, and so does science journalism. Acknowledging the confounding is costly to the field and to Realised's own instinct that these foods are good — the discipline point being that "eat the foods" can be correct advice for reasons other than the polyphenols in them.

Mechanism

The retired mechanism: stoichiometric radical scavenging. The original story was chemically simple. Reactive oxygen species damage lipids, proteins and DNA; polyphenols donate a hydrogen atom from a phenolic hydroxyl group, neutralising the radical; therefore eating more polyphenols means less oxidative damage. This is genuinely true in a test tube, which is why the ORAC assay produced impressive numbers. It fails in a human for two independent reasons: the concentrations achievable in blood are typically one to two orders of magnitude below those used in the test tube, and the body already runs a scavenging system operating at concentrations hundreds to thousands of times higher, with dedicated enzymes on top. Adding a trace of a chemically modified plant metabolite to that system does not move it.

What actually happens to the molecule. Follow one catechin from a cup of tea. Some fraction is absorbed across the gut wall, where it is immediately conjugated — a glucuronide, a sulfate or a methyl group is attached — and further processed in the liver. The molecule that reaches your bloodstream is therefore a chemically altered derivative, usually less reactive than the parent, present at a low single-digit micromolar concentration at most and often far below that. Most of the dose is not absorbed at all and arrives in the colon, where gut bacteria dismantle it into smaller phenolic acids. Those bacterial products can be absorbed, can circulate at higher concentrations than the parent compound, and may be the actual bioactive agents — which means much of the mechanistic literature has been studying the wrong molecule.

The replacement mechanism: mild stress as a signal. The leading modern account inverts the original story. Rather than mopping up oxidants, polyphenols are proposed to act as mild electrophilic or oxidising stimuli that transiently modify cysteine residues on the sensor protein KEAP1, releasing the transcription factor NRF2 to enter the nucleus and switch on the cell's own battery of antioxidant and detoxification genes. On this reading a polyphenol is not an antioxidant delivered from outside; it is a mild irritant that makes the cell build more of its own defences. This is the same logic as hormesis_and_adaptive_stress and the same logic that explains the exercise-antioxidant findings — and it is why "more is better" is incoherent under this mechanism, because a stimulus that works by being mild stops working when it stops being mild. It is also why the plant-chemistry framing in plant_defence_compounds_and_xenohormesis matters: these compounds evolved as the plant's chemical defences, not as human nutrients, and mild toxicity is the point rather than a side effect.

The honest state of the replacement. The NRF2 account is mechanistically coherent, has convergent cell and animal support, and explains findings the old model could not. It is also demonstrated overwhelmingly at concentrations humans do not reach and in species that are not humans. There is no human study establishing that NRF2 activation is the operative mechanism for any measured human polyphenol outcome. A plausible replacement mechanism is not a proven one, and this entry does not treat it as such.

The third candidate: the microbiome as the real interface. Because most of the dose reaches the colon, an increasingly favoured account is that the meaningful biology happens there — polyphenols feeding and reshaping the microbial community, and the community producing the compounds that actually circulate. This elegantly explains why individual responses vary so much and why trial averages can be null while some participants respond. It is also the least established of the three, and the one most vulnerable to being invoked as a placeholder for "we don't know."

Risks And Contraindications

Concentrated extracts carry risks that the food does not. The general principle: a compound with a long history of safe consumption as a food or beverage has no established safety record at ten or fifty times that dose in a capsule. Liver injury from concentrated botanical extracts is the recurring pattern in this category, and the best-documented example within polyphenols is owned by green_tea_extract_egcg_evidence. Treat every "standardised extract" as a different exposure from the food it came from.

High-dose antioxidant supplementation around training may blunt adaptation. Established for vitamins C and E at high doses during endurance training; inferential for polyphenol extracts. If you are training to adapt, high-dose antioxidant supplementation is a plausible way to get less out of the training. Dietary polyphenols from food are not implicated.

Interactions with medication are real and compound-specific. Several polyphenols inhibit or induce drug-metabolising enzymes and drug transporters, and some affect platelet function or interact with anticoagulants. This entry does not enumerate them because they are compound-specific by definition; check the relevant spoke, and check with a pharmacist or prescriber if you take anything regularly. The category label tells you nothing about the interaction risk.

Iron absorption. Tea and coffee polyphenols reduce non-haem iron absorption when consumed with meals. This is a real and well-described effect that matters most for people with low iron stores, menstruating women, and those eating largely plant-based. The practical adjustment is timing — separate tea and coffee from iron-containing meals — rather than avoidance. See antinutrients_evidence_and_context for the broader framing, which also explains why this effect is routinely both over- and under-stated.

Pregnancy, breastfeeding and children. Food-level intake is not the concern; concentrated extracts are, and safety data in these groups is generally absent rather than reassuring. Absence of evidence is the default state here.

The framing risk. The largest practical risk in this category is not toxicity but substitution — spending money and attention on an antioxidant supplement in place of the things that reliably move health outcomes. A 2 mmHg blood pressure effect from cocoa is real; it is also smaller than what sleep, movement, alcohol reduction or salt reduction achieve, and none of those come in a bottle with a margin attached.

Controversy

Nature of the dispute. The argument is not primarily about whether polyphenol-rich foods are good for you — almost everyone agrees they are, though for contested reasons. It is about whether "polyphenols" names anything real enough to sell, and about whether a mechanism can be retired without the whole edifice built on it collapsing. One side says the compounds are powerful and the mechanism is just being refined. The other says the field spent thirty years on a chemistry error and the human data has never justified the industry.

Position A: polyphenols are meaningfully beneficial and the mechanism has simply been updated.

Best evidence for it. The Cochrane cocoa review is a real, replicated, randomised finding on a hard biomarker. EFSA authorised an olive oil polyphenol claim through a genuinely hostile process that rejects most submissions. The COSMOS point estimates sat below 1.0 across every cardiovascular outcome, which is the signature of a small real effect in an underpowered design rather than pure noise. The hormetic and NRF2 account is mechanistically serious work by respected redox biologists, not a marketing retrofit. And the microbial-metabolite story is opening real biology that the old model could not have predicted.

Where it overreaches. It routinely transfers evidence across compounds — the cocoa blood-pressure result is not evidence about a berry extract. It quotes the pooled ~2 mmHg as though it applied to everyone, when the normotensive subgroup showed no significant difference and the effect is concentrated in people who already have raised blood pressure. It cites cell-culture NRF2 activation at concentrations humans never reach as though it were human proof, and rodent results as though species were a detail. It leads with COSMOS's secondary endpoint and omits that the primary failed. It quotes the EFSA claim as cardiovascular protection when the authorised wording is about blood lipid oxidation. And it treats the retirement of the scavenging mechanism as a technical refinement rather than as the collapse of the reason anyone started buying these products.

Position B: the polyphenol supplement field is built on an error and should be treated as such.

Best evidence for it. The absorption arithmetic is decisive against the original mechanism, and the USDA's withdrawal of its own published database is an unusually costly institutional admission. The one large extract trial with hard outcomes failed its primary endpoint. The high-dose isolated-antioxidant era produced a signal of harm in at least one large trial. The observational literature is confounded by the healthy-user effect to a degree that plausibly accounts for most of the association, and exposure measurement in that literature is very poor. Thirty years and enormous funding have produced no compound in this class with demonstrated hard-outcome benefit in humans.

Where it overreaches. Low bioavailability disproves the sponge mechanism, not all mechanisms — signalling effects work precisely at low concentrations, and the microbial metabolites can circulate higher than the parent compounds. "It's all excreted" is chemically wrong as stated, and so is the frequently repeated claim that plasma levels are three orders of magnitude below in-vitro concentrations; the real gap is usually one to two. A null primary endpoint with a confidence interval reaching 0.78 does not exclude a clinically meaningful effect. SELECT was a vitamin E trial, and using it to condemn polyphenols is exactly the class-level reasoning error this entry objects to, run in the sceptical direction. And "the foods are only beneficial because of confounding" ignores that some of the randomised evidence is positive, small though it is.

The funding and bias dimension — cui bono, both ways. On the pro side the incentives are enormous and structural. The global supplement and functional-food industry depends on class-level transfer: evidence earned expensively by cocoa flavanols becomes free marketing for any phenolic extract. Mars Inc. has funded a large share of the cocoa flavanol literature and holds relevant processing patents; COSMOS itself was funded in part by the extract's supplier, alongside NIH support. ORAC scores persist on packaging fourteen years after withdrawal because they are cheap to generate and impressive to read. The bioavailability-enhancement sector needs poor absorption to be true and solvable by their formulation. The urolithin and equol sector needs the metabolite story to be true. The European olive oil industry monetises the EFSA claim directly. Academic labs in the field need it to remain fundable, which favours the hormesis rescue narrative over the conclusion that the field was wrong.

On the anti side the incentives are quieter but real. Pharmaceutical interests benefit whenever a non-prescription intervention looks trivial — a 2 mmHg blood pressure effect is easy to dismiss when your product delivers eight to ten. Mainstream nutrition bodies are institutionally uncomfortable with anything that lets chocolate be reframed as medicine, and that discomfort can shade into under-reading real data. Professional scepticism has its own market: debunking is an audience business, and "it's all a scam" outperforms "it's real but small." And Realised has its own bias to declare — the food-over-extract conclusion is congenial to this platform's philosophy, which is a reason to hold it to a higher standard, not a lower one.

The clean signal in all this is the USDA withdrawal. An institution retracted a database it had published and updated for five years, took the reputational cost, and gained nothing commercially. That is the kind of evidence that is expensive to produce and therefore worth weighting heavily.

Realised Position: Polyphenols are not a thing you can have an opinion about. The category is chemistry, not biology, and any claim made at the category level — in either direction — is unsupported by construction. What we hold: the direct free-radical-scavenging story is dead on the arithmetic and was formally disowned by the institution that popularised its measurement, so any product sold on it is selling an obsolete mechanism. The replacement account — mild stress activating your own defences — is plausible, coherent with the exercise-antioxidant findings, and not yet demonstrated in humans, so we describe it as the leading hypothesis rather than the explanation. The human evidence that exists is real, small, mostly on biomarkers, and almost entirely about foods: cocoa and tea flavanols move blood pressure by about 2 mmHg pooled, with the effect concentrated in people whose blood pressure is already raised and absent as a significant finding in people with normal blood pressure; olive oil polyphenols move a lipid-oxidation marker; and the largest extract trial ever run did not meet its primary endpoint. We therefore recommend the foods without hesitation and without claiming to know that the polyphenols are why. We recommend against buying anything labelled by the category. And for any individual compound, we hold no position at the hub level — read the spoke, because the answer differs by compound and this entry's job is to get you there with the class-level marketing already stripped off.

Cross-Pillar Connections

• green_tea_extract_egcg_evidence — the best-studied single polyphenol and the canonical demonstration of the food-versus-extract split, including the extract-specific liver safety signal. This hub routes all green tea and EGCG questions there and does not re-argue them.
• quercetin_evidence_and_use — a flavonol commonly stacked into immune and longevity products; judged on its own human record, not on the category's reputation.
• resveratrol_and_longevity_claims — the compound that carried the polyphenol-longevity narrative for a decade and the clearest case study in class-level claims outrunning compound-level evidence.
• turmeric_curcumin_evidence — the compound whose poor absorption spawned an entire bioavailability-engineering industry; the practical illustration of why "higher plasma level" is not an outcome.
• spermidine_longevity_evidence — not a polyphenol, but sold in the same aisle under the same plant-longevity story; included so readers can see where the category boundary actually falls.
• methylene_blue_cognition_and_mitochondria — also not a polyphenol, also marketed through a redox and mitochondrial narrative; the same class-level reasoning errors recur, which is why it is worth reading against this entry.
• plant_defence_compounds_and_xenohormesis — the general theory of why plants make these compounds at all (chemical defence, not human nutrition) and why mild toxicity may be the mechanism rather than a side effect.
• hormesis_and_adaptive_stress — the general principle behind the replacement mechanism, and the reason "more is better" is incoherent for anything that works by being a mild stressor.
• mitochondrial_health — where the exercise-adaptation findings land: high-dose antioxidant supplementation blunting the training-induced rise in mitochondrial markers is the clearest human demonstration that oxidative stress carries signal.
• whole_food_emphasis — the practical conclusion this entry supports, with the caveat that "eat the foods" may be right for reasons other than their polyphenol content.
• fruit_whole_food_evidence_and_sugar — where most dietary polyphenol intake actually comes from, and the entry that handles the sugar objection to fruit.
• supplement_form_elemental_dose_and_bioavailability — the general framework for reading extract labels, where "500 mg turmeric" and "500 mg curcuminoids" differ by roughly twenty-fold; read before comparing any two products in this category.
• antinutrients_evidence_and_context — the other face of these same compounds, including tea and coffee polyphenols reducing non-haem iron absorption at meals, and the broader framing of plant compounds that are simultaneously beneficial and interfering.

What would change our mind

Falsifiability: explicit upgrade/downgrade criteria from source

Toward the compounds being more useful than we currently hold:
• A large randomised trial of a defined polyphenol at a defined dose meeting a prespecified primary hard clinical endpoint — mortality, cardiovascular events, incident diabetes — with adequate follow-up. One such trial would move that compound (not the class) substantially.
• Direct human evidence that NRF2 activation, or any specific signalling pathway, mediates a measured human outcome at plasma concentrations actually achieved from diet. That would convert the mechanism from leading hypothesis to established, and would make cross-compound reasoning legitimate for the first time.
• A stratified trial showing that a producer phenotype — urolithin producers, equol producers — predicts response to a defined endpoint, confirming that null averages have been hiding real responders and that the microbial-metabolite account is doing causal work.
• Replication of the COSMOS cardiovascular-death signal as a prespecified primary endpoint in an independent, non-industry-funded trial, with the intermediate events (heart attack, stroke) moving in a direction consistent with it.
• A trial showing the flavanol blood pressure effect in normotensive people, where the current pooled subgroup shows no significant difference — that would widen the finding from "helps people with raised blood pressure a little" to something closer to the way it is marketed.
• Long-term blood pressure data showing the roughly 2 mmHg flavanol effect persists beyond a few months rather than adapting away. Sustained small effects compound; transient ones do not.

Toward the compounds being less useful than we currently hold:
• A well-designed, adequately blinded cocoa trial with a genuine flavanol-matched placebo, in people with raised blood pressure where the effect is supposed to be concentrated, showing the blood pressure effect disappears — which would suggest the Cochrane result is substantially a blinding artefact.
• Evidence that the EFSA-authorised oxidised-LDL effect does not translate to any clinical endpoint, or that the oxidised-LDL assay itself does not track what it is assumed to track.
• A large trial of a concentrated polyphenol extract showing harm at supplement doses, extending the pattern from the vitamin era into this class directly rather than by analogy.
• Demonstration that the observational associations vanish entirely under quasi-experimental designs — Mendelian randomisation on relevant metabolic traits, or negative-control-outcome analyses — confirming that the healthy-user effect is doing all the work.

What would NOT move us:
• More in-vitro antioxidant capacity data of any kind, at any concentration, on any assay. That question is settled and the assay was withdrawn.
• More rodent studies showing NRF2 activation or lifespan effects at doses that do not scale to human diets. These accumulate endlessly and have not once predicted a human result in this class.
• Another observational cohort finding that people who drink more tea live longer. The confounding structure is known and predictable, and the hundredth such study adds nothing the first did not.
• Mechanistic plausibility arguments unaccompanied by human outcome data — including elegant ones, including ours.
• A new secondary or post-hoc endpoint from an existing trial. If the primary failed, later analyses of the same data generate hypotheses; they do not confirm them.
• Industry-funded trials of proprietary "enhanced bioavailability" formulations showing higher plasma levels. Higher plasma concentration is not an outcome; it is an intermediate step whose relevance is precisely what is in dispute.
• Any argument that a negative trial used the wrong form, wrong dose or wrong population, offered after the fact. That rebuttal has been used against every negative antioxidant trial for twenty-five years and is unfalsifiable as deployed.

Industry bias note

Structural incentives the evidence base may reflect

The pro-polyphenol end. The commercial architecture of this category is built specifically on class-level transfer, and that is not incidental — it is the business model. If evidence had to be earned compound by compound, most products in the aisle would have nothing to say. Instead, a randomised trial on cocoa flavanols funded by a chocolate manufacturer becomes ambient support for a berry-extract capsule with no trials at all. Specific mechanisms of distortion worth naming: a single company has funded a large share of the cocoa flavanol literature and holds relevant processing patents, so the field's shape reflects one firm's research priorities; COSMOS was funded in part by the supplier of the extract under test (alongside NIH support), and its null primary endpoint was reported far less than its favourable secondary; ORAC scores survive on packaging because they are cheap, large and impressive, fourteen years after their creator withdrew them; the bioavailability-enhancement sector requires the absorption problem to be real and solvable only by their formulation; the urolithin and equol sector requires the metabolite account to be the true one; and the European olive oil industry directly monetises a regulatory claim about a laboratory marker that is routinely quoted as a claim about heart disease. Academic incentives compound this — a field whose founding mechanism was retired needs a replacement to stay fundable, which is exactly the position the hormesis account occupies, however good the science behind it is.

The anti-polyphenol end. The sceptical side is less funded but not disinterested. Pharmaceutical antihypertensive and lipid interests benefit whenever a food-based intervention looks negligible; 2 mmHg is easy to make sound worthless next to a drug. Public health and mainstream nutrition bodies have an institutional aversion to any framing that lets chocolate be discussed as medicine, given its energy and sugar load, and that legitimate concern can shade into dismissing real randomised data. Scientific debunking is itself an audience business with its own reward structure, and "the whole field is a scam" is a better-performing claim than "it is real but small and compound-specific." Finally, the reductionist dismissal — barely absorbed, therefore inert — is intellectually cheap: it lets a critic skip the entire signalling and microbial-metabolite literature while sounding rigorous, and it tends to travel with inflated arithmetic (the "three orders of magnitude" figure) that does not survive checking.

Realised's own bias, declared. The conclusion this entry reaches — eat the foods, distrust the capsules — is congenial to this platform's philosophy and to its scepticism of optimisation products. That congeniality is a reason for extra scrutiny, not less. So we state plainly: some individual extracts do have real compound-specific human evidence, the spokes say so, and "food beats supplement" is partly an artefact of food-based epidemiology being vast while extract trials are few.

The clean signal. Three things in this entry were expensive for their producers to say and are therefore weighted heavily. The USDA withdrew a database it had published and updated for five years, absorbed the reputational cost, and gained nothing commercially. The COSMOS investigators, funded in part by the extract's manufacturer, stated plainly in their own paper that the primary outcome was not significantly reduced. And the exercise-adaptation findings are commercially inconvenient to an entire supplement category with no constituency arguing for them. When a finding costs its own source something, it deserves more weight than a finding that pays.

Polyphenols: what the evidence supports, and why the antioxidant story collapsed