Late Lessons, Jensen Huang and AI
Markdown version

Hindsight check: LL2-26 (Part E introduction; Ch 26 Science for precautionary decision-making)#

Source section: EEA, Late lessons from early warnings: science, precaution, innovation (EEA Report No 1/2013), Part E title page and contents (pp. 621–622) and Ch 26 by Philippe Grandjean (report pp. 623–640; PDF pp. 623–644). - Part E has no introductory essay, so there is nothing in it to check. Everything below concerns Ch 26. - Ch 26 is a cross-cutting position essay by an environmental epidemiologist who was also editor-in-chief of Environmental Health and a member of the report’s editorial team.

Check window: publication (2013) to late September 2026. The most recent data in the chapter are the author’s 2010–2011 extension of his bibliometric search (p. 627). Checked: 26 September 2026.

Method note. - Search. General web search was unavailable for this pass because the session’s search budget was exhausted. I retrieved sources directly from: - the Europe PMC REST API (abstracts, and open-access full texts for competing-interest statements); - the US Federal Register API and its full-text pages (EPA, FDA, and Executive Order 14303); - the EU Publications Office cellar: EUR-Lex texts of regulations, directives and CJEU judgments, plus a SPARQL query for acts on diuron, dichlofluanid, imidacloprid and triclosan; - SEC EDGAR (3M and DuPont 8-K filings; 3M’s 10-K for 2025); - the websites of EFSA, BfR, US FDA, US EPA, IARC, the NTP, NIH, the OSTP archive, cOAlition S, PARC, WHO and the Stockholm Convention. - Secondary sources, used only where no primary source was reachable and flagged where used: - Retraction Watch (the reasons for the 2025 glyphosate retraction); - a Guardian report, which I could not open. I located it through a Wikipedia citation, and its claim is unverified. - Own bibliometric check (Claims 1 and 6). - What I counted: Europe PMC hit counts (PubMed subset, SRC:MED; title or abstract terms) for 2000–2009 against 2015–2024, normalised to the growth of the whole database (×2.07: 6,436,003 records rising to 13,317,162). - Not comparable with the chapter’s data. Grandjean counted SciFinder CAS-number links in 78 environmental journals. My counts cover all biomedical literature, and terms are ambiguous: “arsenic” includes arsenic trioxide therapy, and lead could only be approximated through phrases. - So: read these as indicative of direction and scale only. - Access gaps. - ECHA’s website was blocked by a firewall, so I could not confirm the September 2026 status of the EU’s universal PFAS restriction proposal or of other ECHA processes. - The D.C. Circuit docket for the challenge to the 2024 US PFAS drinking-water rule (AWWA v. EPA, No. 24-1188) could not be read. - The full text of the ASA task-force statement (2021) was paywalled; it is characterised here only by title, venue and date. - EFSA’s journal site returned HTTP 403, so EFSA opinions were read through their Europe PMC abstracts and open full texts. - Scope. - I read the section digest, the section’s source extract and web sources. - For formatting only, I also looked at the structure of one sibling Strand A hindsight file (LL2-21). I rely on nothing from it: the one overlapping fact (Karipidis et al. 2024) was re-verified at source. - Page numbers are report pages of the 2013 volume. - Author’s later roles. These matter because the chapter’s own argument is about independence. - Grandjean remained editor-in-chief of Environmental Health, where the chapter’s core bibliometric study [S1] and several of his later essays were published [S13, S23, S100]. - He disclosed in 2018 that he had recently served as a health expert for the State of Minnesota in its lawsuit against a PFAS-producing company [S23]. - His Faroese cohort is one of the two critical human studies EFSA considered for its 2020 PFAS limit [S15]. - So the author is both a principal scientific source and a litigation expert in one of the chapter’s two flagship examples. On the other side, the main industry-funded review of the same evidence was paid for by 3M [S25].

Overview#

1. The chapter’s two flagship “safety claims later overturned” examples, perfluorinated compounds (PFAS) and bisphenol A (BPA), have been strongly vindicated in the EU and, for PFAS, in the US (Claims 2, 3). - EFSA: - set a PFAS tolerable weekly intake of 4.4 ng/kg bw on a vaccine-antibody endpoint in 2020 [S15]; - cut its BPA tolerable daily intake 20,000-fold in 2023 [S33]. - US EPA: - issued PFOA/PFOS health advisories of 0.004 and 0.02 ppt in 2022 [S16]; - set 4 ppt limits with zero health goals in 2024 [S17]. - IARC classified PFOA as a Group 1 carcinogen in 2023 [S21]. - Producers and bans: - 3M stopped all PFAS manufacture at the end of 2025 [S29]; - PFAS producers agreed water-system settlements of $10.5–12.5 billion (3M) and $1.185 billion (DuPont/Chemours/Corteva) [S27, S28]; - the EU banned BPA in food-contact materials from January 2025, with transitions [S36]. - What remains contested: - the US FDA still says BPA is safe at current levels [S37]; - Germany’s BfR and the European Medicines Agency rejected EFSA’s BPA endpoint [S35]; - in May 2026 the US EPA proposed rescinding four of its six PFAS drinking-water limits and delaying the other two [S19, S20]. That is a policy retreat, not a scientific reversal.

2. The descriptive core of the “research neglects poorly known hazards” argument held up (Claim 1). - Research on environmental chemicals remains concentrated on a small, familiar set [S2, S3, S4]. - Most chemicals in commerce are still barely studied or not publicly identified [S6]. - But the implied inertia was overstated. Several of the specific hazards the chapter named as neglected went on to attract rapidly growing research: PFAS, BPA, neonicotinoids and triclosan [S8]. - The shift followed controversy, litigation and regulation, not priority lists. There is no direct evidence that REACH, TSCA reform or the EU’s chemicals-risk research partnership (PARC) redirected academic attention.

3. The chapter’s choice of which emerging hazards were neglected looks prescient (Claim 6). - Triclosan was removed from US consumer antiseptic washes (2016) and refused EU biocide approval for human hygiene uses (2016) [S62, S63]. - Neonicotinoids were restricted in the EU (2013), banned outdoors (2018), upheld in court (2021), and targeted through import residue limits from March 2026 [S65–S69]. - The booster biocide diuron lost EU plant-protection approval in 2020 [S72]. - The exception: dichlofluanid was approved for antifouling use in 2017, and research on it declined [S73, S8].

4. The statistics critique was largely endorsed by the statistics profession, but practice and policy moved only partly (Claim 5). - The American Statistical Association (ASA)’s 2016 statement adopted the chapter’s core points [S52]. - A 2019 campaign to “retire” statistical significance followed [S54, S55]. So did a rival proposal to tighten the threshold to p<0.005 [S53], and a 2021 ASA task-force defence of p-values [S56]. - P-values remain entrenched: 94–98% of biomedical articles that report p-values report at least one ≤0.05 [S58]. - Bayesian methods did enter regulatory toxicology: EFSA’s 2022 benchmark-dose guidance [S60]. - The worst-case recommendation met direct political opposition. The chapter wanted worst-case scenarios scrutinised as carefully as the null hypothesis. A 2025 US Executive Order told agencies to rely on “highly unlikely and overly precautionary” scenarios only where required by law or otherwise pertinent [S61].

5. The claimed asymmetry of error is the chapter’s most weakened claim (Claim 4). - Design features that produce false negatives are real. But the replication crisis shows that low power and publication bias also produce many false positives [S43–S45]. - Environmental science had its own clear false positive: ocean acidification’s reported effects on fish behaviour [S46, S47]. - Within this report, the mobile-phone warning (Ch 21) has largely not been borne out [S50]. - “Most often” underestimation is overstated: non-differential misclassification does not guarantee underestimation [S48]. - Still arguably true: erroneous regulatory alarms remain uncommon. For data-poor chemicals, the dominant risk is still false negatives by default.

6. The exposure-limit ratchet held for the chapter’s flagship toxicants but is not universal (Claim 7). - Limits kept falling after 2013 for: - lead (US CDC [S75]; EU workplace limits [S76]; US drinking water [S77]); - fine particles (WHO, EU, US [S78, S79]); - PFAS [S16, S17, S23]; - BPA [S33]. - Evidence-driven counterexamples exist: - EFSA raised its nickel intake limit 4.6-fold in 2020 [S80]; - US EPA’s proposed perchlorate health goal (20 µg/L, 2026) is above its 2008 advisory level (15 µg/L) [S81].

7. Sponsorship bias is well supported for drugs and plausibly for chemicals, but institutions chose disclosure over the elimination of ties that the chapter preferred (Claim 8). - Drugs: the updated Cochrane review found industry-sponsored drug and device studies more often favourable, and the difference was not explained by standard risk-of-bias assessments [S83]. - Chemicals: the post-2013 evidence is mostly case-level and qualitative: - PFAS company documents [S24]; - chlorpyrifos raw data that did not match the test summary [S85]; - a 3M-funded immune-effects review [S25]; - the 2025 retraction, for undisclosed Monsanto authorship, of a 2000 glyphosate safety review [S87, S88]. It appeared in Regulatory Toxicology and Pharmacology, a journal the chapter had named (p. 637). - Institutions chose disclosure. The EU’s 2019 Transparency Regulation requires disclosure and notification of industry studies [S90].

8. Openness grew as the chapter hoped, and its warning that openness can be turned against research was borne out (Claim 10). - Growth: Plan S (2021), the US OSTP zero-embargo policy (2022), NIH public access (July 2025) and NIH data sharing (2023) [S91–S94]. - Openness used against research: US EPA’s 2018–2021 “transparency” rule would have given less weight to studies whose underlying dose-response data were not public. A federal court vacated it in 2021 [S96–S99]. Its logic returned in Executive Order 14303 (2025) [S61].

9. The recommendations were taken up in European research-programme design, but there is no evidence they sped up protection (Claim 9). - Regulator- and stakeholder-informed prioritisation now shapes the European human-biomonitoring programme HBM4EU and PARC, a €400m partnership running 2022–2029 [S10–S12]. - Timescales remain long: - the PFAS and BPA limits above came roughly a decade after the chapter; - the chapter’s author wrote in 2019 that better documentation “has resulted in little progress in protection” [S102]. - US federal research capacity was reportedly cut in 2025 [S103; unverified secondary].

Claim-by-claim#

Claim 1: Research effort concentrates on already well-known chemicals and neglects data-poor priority substances (pp. 625–629)#

What the chapter says. - The author’s bibliometrics (78 environmental journals, 2000–2009): - 119,636 articles and 760,056 CAS links; - the top 20 substances took about 12% of links, and the top 10 were all metals (pp. 626–627); - 13 data-poor high-production chemicals on a 2006 US EPA priority list had only 352 links, five had none, and there was no increase by 2010–2011 (pp. 627–628). - The chapter attributes this to a “Matthew” dynamic (p. 629; Grandjean et al. 2011 [S1]). - It also claims more broadly that “the majority of environmental chemicals are poorly documented” (p. 639).

Subsequent developments.

Evidence that concentration persisted: - Muir et al. (ES&T, June 2023) [S2] ran a 50-year CAS bibliometric survey of chemicals measured in environmental media. - It covered 19,776 CAS numbers and 9.25 million counts. The top 100 substances accounted for 34% of counts, “confirming previous studies showing a significant bias toward repeated measurements of the same substances.” - Substances on the EU, Chinese and US industrial inventories made up only about 5% of the substances measured. - Muir et al. attribute the bias partly to regulatory monitoring needs. That complicates the chapter’s contrast between academic habit and regulatory need. - Muir et al. also found that pharmaceuticals and current-use pesticides made up 50–60% of counts in 2000–2015. That suggests the chapter’s picture of metals dominating depends partly on its method. - Sobek et al. (2016) [S3]: 87% of chemical analyses of Baltic Sea fish in 2000–2012 concerned 20% of the substances, mostly already-regulated persistent organic pollutants. They warn of “prioritizations… biased based on the knowns of the past.” - Daughton (2014) [S5] applied the Matthew effect to environmental monitoring of pharmaceuticals. He found “absence of data” for many widely prescribed drugs. - Grandjean’s own follow-up (Human & Experimental Toxicology, 2015) [S4]. He sampled 530 articles from four volumes of that journal (1984–2014). Metals dominated, and none of the 13 US priority substances was covered at all. The inertia persisted in his own field. - Wang et al. (2020) [S6]: - over 350,000 chemicals and mixtures are registered for production and use worldwide; - more than 50,000 have identities claimed as confidential, and up to 70,000 are ambiguously described. - This strongly supports the p. 639 claim that most chemicals are poorly documented. - The regulatory agenda showed the same pull toward the familiar. TSCA’s 2016 reform required EPA to start with ten chemicals drawn from its existing 2014 Work Plan list [S9]. These were long-studied substances such as asbestos, trichloroethylene and methylene chloride, not data-poor ones. - Funding shapes topics. A bibliometric study of glyphosate research found output tracking market size, with Monsanto the second-largest publishing institution after the US Department of Agriculture [S104].

Evidence that attention can shift (own Europe PMC counts [S8]; database growth ×2.07 between the two decades):

Term (title/abstract) 2000–09 2015–24 Growth Relative to database growth
PFAS family (PFOS/PFOA/PFAS/perfluoro-/polyfluoro-) 1,495 10,761 ×7.2 ×3.5
neonicotinoid* 262 2,957 ×11.3 ×5.5
bisphenol A 2,139 10,191 ×4.8 ×2.3
imidacloprid 673 3,025 ×4.5 ×2.2
triclosan 784 2,644 ×3.4 ×1.6
cadmium 9,616 24,134 ×2.5 ×1.2
arsenic 7,408 18,738 ×2.5 ×1.2
mercury 9,135 16,112 ×1.8 ×0.85
diuron 439 629 ×1.4 ×0.69
lead (phrases: “blood lead”, “lead exposure”, “lead poisoning”) 2,606 3,513 ×1.35 ×0.65
dichlofluanid 45 34 ×0.76 ×0.37

Did REACH, TSCA reform or PARC cause the shift? - I found no study attributing academic attention shifts to these instruments. The timing fits controversy, litigation and regulatory action better. Examples: - PFAS: US litigation disclosures and the 2016–2024 US health advisories and limits; - neonicotinoids: the 2013 EU restrictions. - Institutional uptake of prioritisation by regulatory need is real, but recent: - HBM4EU (2017–2022): used nominations from EU and national ministries, agencies and a stakeholder forum to prioritise substances for biomonitoring (48 to 23 to 9 groups) [S12]. - PARC: launched 11 May 2022; €400m over seven years; about 200 institutions in 30 countries plus ECHA, EFSA and the EEA [S10]. It revises its project portfolio against ECHA’s 2023 “Key Areas of Regulatory Challenge” [S11]. - These programmes are designed to do what the chapter asked. Their effect on the wider academic literature is not yet measurable. - The author himself doubled down. In 2020 he and Ozonoff argued that treating replication as a criterion of validity “has resulted in a tendency toward inertia” in environmental health research, and proposed “stability” as an alternative [S13].

Verdict: partly held up. - The descriptive claim of concentration held and is now better documented [S2, S3, S4, S6]. - The chapter’s own caveat stands: counts are not content. The continued lead research he calls repetitive is also what revealed harm at lower doses (see the digest). - The implied inertia was overstated. Attention did move sharply to several named emerging hazards, though after controversy rather than in anticipation of it.

Implication for weight. - The lesson that research attention is path-dependent, and that listing a priority does not redirect it, remains strong as description. It fits the 2023 data and the regulator-driven repetition Muir et al. identify. - The causal story (academic reward structures) should carry moderate weight only. Regulatory monitoring requirements are an equally strong driver, and controversy can redirect attention quickly. - The more useful lesson for a general lens: attention follows salience, not hazard. Uncontroversial, data-poor agents stay dark until something makes them salient.

Claim 2: Perfluorinated compounds: a major US producer claimed for decades that little escaped and toxicity was negligible; current exposures “may be far from safe”; the chemicals “cannot be recalled” (p. 636; also p. 628)#

What the chapter says. - For decades a major US producer claimed little would escape and “essentially no toxicity occurred” (citing Lindstrom et al. 2011). - Only recently was it discovered that current exposures “may be far from safe” (Grandjean et al. 2012, the Faroese vaccine-antibody study [S32]). - The chemicals are globally dispersed and “cannot be recalled” (p. 636).

Subsequent developments.

What producers knew and said: - Gaber, Bero & Woodruff (Annals of Global Health, 2023) [S24] analysed previously secret industry documents. - Companies knew by 1970 that PFAS were “highly toxic when inhaled and moderately toxic when ingested”, “forty years before the public health community”. - The companies used strategies common to other industries, “most notably, suppressing unfavorable research and distorting public discourse”. - The same authors found no evidence in this archive of the companies funding favourable research. That qualifies the funding half of the pattern the chapter describes (pp. 636–637). - Grandjean (2018) [S23]: - early company studies were not released until after 2000; - the first PFAS risk assessments “ignored these reports and relied on scant journal publications”.

Regulatory and scientific assessments (strongly consistent with the chapter): - EFSA (17 Sept 2020) [S14, S15]: - a group tolerable weekly intake (TWI) of 4.4 ng/kg bw/week for PFOA, PFNA, PFHxS and PFOS; - critical effect: decreased antibody response to vaccination in children. EFSA considered two critical studies, the Faroese cohort and a German cohort of 101 infants. The limit was based on the latter (a benchmark-dose lower limit of 17.5 ng/mL in 1-year-olds); - EFSA concluded that “parts of the European population exceed this TWI, which is of concern.” - US EPA: - June 2022 [S16]: interim lifetime health advisories of 0.004 ppt (PFOA) and 0.02 ppt (PFOS), replacing the 2016 value of 70 ppt. The basis was “suppression of vaccine response… in children”. - April 2024 [S17]: enforceable limits of 4.0 ppt for PFOA and PFOS, with health goals of zero; 10 ppt for PFHxS, PFNA and GenX; and a Hazard Index for mixtures. - Grandjean’s 2018 table [S23] shows the decline in drinking-water limits. US EPA values for PFOS/PFOA were 200/400 ng/L in 2009 and 70/70 in 2016; his 2013 immune-based benchmark estimate was below 1 ng/L. The 2022 advisories are consistent with his estimate. - IARC (Nov 2023) [S21]: PFOA “carcinogenic to humans” (Group 1), on sufficient animal evidence, strong mechanistic evidence in exposed humans (including immunosuppression), and limited human evidence for kidney and testicular cancer. PFOS Group 2B. - US National Academies (2022) [S22]: recommended clinical PFAS testing and follow-up for highly exposed people. - Global and EU controls: - Stockholm Convention [S30]: PFOS listed in Annex B (restriction); PFOA, PFHxS and long-chain perfluorocarboxylic acids listed in Annex A (elimination). The listing decisions date from 2009, 2019, 2022 and 2025 respectively. Those years come from my knowledge of the COP decisions; the list page confirms only the annexes. - EU Drinking Water Directive [S31]: limits of 0.10 µg/L for a “Sum of PFAS” and 0.50 µg/L for “PFAS Total”, with compliance required by 12 January 2026. These are far less stringent than the US 2024 limits.

Producer exit and liability: - 3M: - announced on 20 Dec 2022 that it would exit all PFAS manufacturing by the end of 2025, citing “accelerating regulatory trends” [S26]; - its 10-K for 2025 states: “3M completed its exit of PFAS manufacturing at the end of 2025” [S29]. - Water-system settlements: - 3M (22 June 2023): $10.5–12.5 billion, payable 2024–2036, with no admission of liability [S27]; - DuPont, Chemours and Corteva (June 2023): $1.185 billion [S28]; - 3M’s 2025 New Jersey settlement: up to $450 million, also with no admission of liability [S29].

Contrary or complicating evidence: - Industry-funded review. A 3M-funded review (Chang et al. 2016; the authors report 3M could review drafts but had no control) concluded the epidemiologic evidence was “insufficient to reach a conclusion about a causal relationship” between PFOA/PFOS and immune-related conditions [S25]. - Its caution that immunodeficiency “should not be presumed” without clinical abnormality is the main scientific objection to the vaccine-antibody endpoint. - Mainstream assessments (EFSA 2020, US EPA 2022–2024, IARC 2023, NASEM 2022) went the other way. - The US policy retreat: - on 14 May 2025 EPA announced that it would keep the PFOA/PFOS limits but revisit the others [S18]; - on 20 May 2026 EPA proposed rescinding its limits for PFHxS, PFNA, GenX and the Hazard Index [S19]; - it also proposed allowing systems to delay compliance with the PFOA/PFOS limits from 2029 to 2031 [S20]. - These proposals are legal and procedural (framed around Safe Drinking Water Act requirements). They do not reassess toxicity. - The author’s role. As noted above, the immune endpoint is based partly on the chapter author’s own cohort, and he has served as a litigation expert against a producer [S23]. The evidence has since been independently reproduced (the German cohort in EFSA [S15]) and endorsed by agencies. That reduces, but does not remove, the concern that the chapter’s authority and its evidence are intertwined.

Verdict: strengthened. - The historical claim about producer knowledge is now documented from company records [S24]. - “May be far from safe” became the regulatory conclusion on both sides of the Atlantic [S15, S16, S17]. - “Cannot be recalled” is borne out by multi-billion-dollar remediation settlements [S27, S28, S29].

Implication for weight. - This is one of the strongest cases in the report for these lessons: persistent, dispersed agents make delay irreversible, and producer knowledge can run decades ahead of public science. - It also shows two further mechanisms: - (a) Limits keep falling as endpoints become more sensitive. Values dropped by about four orders of magnitude between 2016 and 2022 [S16]. - (b) Late regulation can be reversed for reasons unrelated to the science (US, 2026). - Caveat for use as a lens: the key endpoint (antibody response) is an intermediate marker. Its clinical significance is still disputed by industry-funded reviewers.

Claim 3: Bisphenol A was said to be safe at low consumer exposures, but after decades of expanding use independent research found evidence of health risks (p. 636)#

What the chapter says. - BPA was “said to be safe at the very low exposures that consumers were likely to receive”. After “several decades of expanding use, independent research eventually uncovered evidence of health risks” (p. 636, citing Myers et al. 2009). - BPA ranked 62nd in the author’s bibliometrics (952 links, 2000–2009; Table 26.2, p. 627).

Subsequent developments.

EU (strongly consistent): - EFSA (19 April 2023) [S33, S34]: - set a TDI of 0.2 ng/kg bw/day, 20,000 times lower than its 2015 temporary TDI of 4 µg/kg; - critical effect: an increase in Th17 (inflammation-related T-helper) cells in mice. EFSA called the immune system the “most sensitive”; - estimated mean and 95th-percentile exposures exceeded the TDI “by two to three orders of magnitude”, so there is “a health concern”. - Commission Regulation (EU) 2024/3190 (19 Dec 2024) [S36]: - bans BPA (and hazardous bisphenols) in food-contact materials; - takes effect 20 days after publication, i.e. January 2025, with transition periods of 18 and 36 months and some repeat-use articles allowed until January 2029; - its recitals rely directly on EFSA’s 2023 TDI. - EU Drinking Water Directive [S31]: a BPA limit of 2.5 µg/L from January 2026.

Divergence (the claim is still contested outside EFSA): - BfR (Opinion 018/2023, 19 April 2023) [S35] “does not support the new TDI derived by EFSA”. - It says there is “currently no evidence” that the Th17 increases cause adverse effects, and that the human relevance is “questionable”. - It proposes 0.2 µg/kg bw/day (200 ng) instead. - The European Medicines Agency also registered methodological disagreement [S35]. - Note, though: BfR’s value is itself 20 times lower than EFSA’s 2015 temporary TDI. BfR, too, accepts that effects occur at lower doses than previously assumed. - US FDA [S37] maintains that “BPA is safe at the current levels occurring in foods”. The page was last updated 20 April 2023, but its substantive content dates from 2014. - A 2022 food-additive petition (Environmental Defense Fund, the Endocrine Society and others) asked FDA to remove or restrict BPA authorisations [S38]. - I found no Federal Register notice of a decision on it through September 2026. - The US CLARITY-BPA programme (NTP/FDA/NIEHS) complicates “independent vs industry”: - The guideline-compliant core study was run by FDA’s own laboratory, not industry. It found “no BPA-related effects” on in-life and non-histopathology endpoints, and only possible effects at the top dose (25,000 µg/kg/day) [S39]. - The academic grantee studies, on blinded samples from the same animals, reported effects in brain, prostate, urinary tract, ovary, mammary gland and heart. “Many effects” were at the lowest dose (2.5 µg/kg/day) and “many… were non-monotonic” [S40]. - The NTP compendium (2021) collates both without integrating them [S41]. - So the split is less about funding than about guideline tests versus investigator-designed endpoints. That is a different mechanism from the one the chapter emphasises. - Substitution. Under PARC (2026), eight BPA alternatives are being tested for “regrettable substitution” [S42]. This carries the chapter’s concern about the replacements for a restricted chemical into current research.

Verdict: strengthened (in the EU); still contested in the US and among some EU agencies. - The EU risk assessor and legislator adopted the chapter’s view. - The specific critical endpoint remains disputed [S35], and FDA has not moved [S37].

Implication for weight. - BPA supports the lesson that claims of safety at low doses can be revised dramatically. The size of the revision (20,000-fold at EFSA) is unusual. - The CLARITY-BPA results [S39, S40] suggest a lesson the chapter did not state. Standardised guideline tests and investigator-designed studies can reach opposite conclusions on the same animals. Which endpoints count becomes the site of controversy. - This is at least as important as sponsorship, and more transferable.

Claim 4: Error asymmetry: standard designs and conventions bias toward false negatives; uncertainties “most often” lead to underestimation; erroneous alarms are “fairly rare” (Table 26.4, p. 635; pp. 636, 638)#

What the chapter says. - Table 26.4 lists ten design features whose main direction of error is toward false negatives, including: - low power; - exposure misclassification; - insensitive outcomes; - disregard of vulnerable subgroups; - the 5% and 20% error conventions; - “pressure to avoid false alarm”. - Against these it lists three that push toward false positives: incomplete confounder adjustment, post hoc hypotheses and publication bias (p. 635). - The chapter says uncertainties “most often” cause underestimation (p. 638), and, relying on Ch 2, that “erroneous alarms are fairly rare” (p. 636). - It dismisses publication bias toward scares by pointing to how little is published on emerging hazards (pp. 635–636).

Subsequent developments.

Evidence that false positives are common in published research: - Psychology (Open Science Collaboration, Science, Aug 2015) [S43]: 97% of 100 original studies were significant; 36% of the replications were. Replication effects were “half the magnitude”. - Preclinical cancer biology (Errington et al., eLife, Dec 2021) [S44]: for positive effects, replication effect sizes were 85% smaller (median). 40% of positive effects replicated on three or more of five criteria. - Low power cuts both ways (Button et al., 2013) [S45]. Low power “also reduces the likelihood that a statistically significant result reflects a true effect” and inflates effect sizes. Low power therefore works in both directions, not only toward false negatives as Table 26.4 implies. - An environmental example of a false alarm in the research literature (not a regulatory one): - Clark et al. (Nature, 2020) [S46] found reported effects of ocean acidification on coral-reef fish behaviour “not reproducible”. Their simulations showed the original effect sizes and variances were “highly improbable”. - A meta-analysis of 91 studies (Clements et al., PLoS Biology, 2022) [S47] found an “extreme decline effect”. Large early effects came from small studies in high-impact journals. - Within this report: the WHO-commissioned systematic review (Karipidis et al., Environment International, 2024) [S50] found mobile-phone use not associated with glioma (mRR 1.01, 0.89–1.13), meningioma or acoustic neuroma. Ch 21’s early warning has largely not been borne out. - Multiplicity in exposome-scale epidemiology: Patel & Ioannidis (2014) [S49] argued that multiplicity is often undisclosed. They proposed placing each new association within the distribution of effects in the dataset.

Evidence against “most often underestimation”: - Jurek et al. (IJE, 2005; before the check window, cited because the chapter overlooks it) [S48]: non-differential misclassification “does not justify claims that the observed estimate must be an underestimate”. Individual estimates can be overestimates by chance even when the bias is toward the null. - Differential recall bias pushes away from the null. It is not in Table 26.4.

Evidence consistent with the chapter: - The author extended the claim in 2015 [S4]: biases toward the null “are generally not considered”, and toxicology suffers from “a vast number of toxic hazards which by default are considered innocuous due to lack of documentation”. - That default is well supported: more than 350,000 registered chemicals, with tens of thousands of confidential identities [S6]. - For untested chemicals, regulatory inaction is a false negative by construction. - For the report’s flagship cases, the errors ran toward underestimation. Later assessments of PFAS, BPA and lead repeatedly tightened limits (Claims 2, 3, 7). - Regulatory false alarms remain uncommon. The alarms that did not pan out are mostly research-level findings (ocean acidification and fish behaviour) or unconfirmed classifications (mobile phones, still IARC 2B). They are not costly regulatory interventions that had to be reversed. - I found no post-2013 systematic count of regulatory false positives to test Ch 2’s figures directly. This check therefore cannot confirm or refute “fairly rare” at the regulatory level. - A related retraction: Séralini et al. 2012 (an alarming GM-maize/Roundup rat study) was retracted in 2014 by Food and Chemical Toxicology [S51]. - I did not read the notice itself. The retraction is widely reported to have been for inconclusive results (small group sizes, rat strain) rather than fraud, and the study was republished elsewhere in 2014. Treat that characterisation as unverified. - The episode shows that “false alarm” labels are themselves contested.

Verdict: contested. - The chapter is right that many conventional features push toward false negatives, and that for data-poor agents non-action by default dominates. - It is wrong or overstated in three places: - that low power is only a false-negative problem; - that uncertainty “most often” means underestimation; - that publication bias toward scares is answered by the neglect of emerging hazards (a non sequitur). - The replication crisis literature, which postdates the chapter, shows false positives are common at the level of published findings.

Implication for weight. - Two claims should be kept apart when this is used as a lens: - At the level of regulatory defaults for untested agents: inaction is a false negative by construction. Strong. - At the level of individual research findings: errors run both ways, and small, noisy, high-profile studies are prone to exaggeration. The chapter’s claim here is weak. - A sound restatement: under low power and high uncertainty, both false alarms and false reassurance become likely; which error is costlier depends on irreversibility and scale. It is better supported than the chapter’s one-directional version and keeps its precautionary point.

Claim 5: Statistics: replace the p<0.05 dichotomy with confidence intervals, treat the upper confidence limit as a plausible worst case, use sensitivity/worst-case and power analyses; Bayesian use is “gaining support” (pp. 632–635, 638)#

What the chapter says. - The 5% threshold is “almost sacrosanct” (p. 632). - Non-significant results are wrongly read as “no risk” (p. 635). - The chapter recommends: - confidence intervals, with the upper limit as “a plausible worst case” (p. 633; Figure 26.1); - worst-case scenarios deserving “as careful scrutiny as the null hypothesis” (p. 638); - power analysis (p. 635). - It notes that “empirical use of Bayesian statistics is gaining support” (p. 633).

Subsequent developments.

Endorsement by the statistics profession: - ASA Statement on p-values (7 March 2016) [S52]. Six principles, including: - (3) conclusions and policy decisions “should not be based only on whether a p-value passes a specific threshold”; - (5) a p-value “does not measure the size of an effect or the importance of a result”. - This is the chapter’s core critique, now stated by the professional body. - Principle 2 also corrects the chapter. P-values “do not measure the probability that the studied hypothesis is true”. The chapter’s definitions of p-values (“The probability that their results are significant is usually expressed as p values”, p. 632) and of confidence intervals (p. 633) are both technically wrong (as the digest noted), though its practical conclusion is not. - 2019 “retire statistical significance”: - a Nature comment with more than 800 signatories (Amrhein, Greenland & McShane, 20 March 2019) [S54]; - an American Statistician special issue whose editorial recommended no longer using “statistically significant” (Wasserstein, Schirm & Lazar 2019) [S55].

A counter-movement: - “Redefine statistical significance” (Benjamin et al., Nature Human Behaviour, 2018) [S53] proposed p<0.005 for claims of new discoveries. This is the opposite remedy: it guards against false positives, which suits the replication-crisis diagnosis (Claim 4) and runs against the chapter’s concern about false negatives. - ASA President’s Task Force (Annals of Applied Statistics, Sept 2021) [S56] issued a statement defending properly used p-values and significance tests. It was prompted by concern that the 2019 editorial was being read as ASA policy (full text not accessed).

Practice: - Epidemiology moved partly (Stang, Deckert, Poole & Rothman, 2017) [S57]. Across 89,533 abstracts (1975–2014), reporting of confidence intervals only rose steadily. By 2014, confidence intervals alone were as common as significance testing in four epidemiology journals (e.g. Epidemiology 79%). JAMA, the NEJM and the Lancet remained dominated by significance testing. - Biomedicine overall stayed entrenched (Choi, Chavalarias, Ioannidis et al., preprint, Jan 2026) [S58]: - across 22.7 million PubMed abstracts (1990–2025), the share reporting p-values rose from 7.5% to 18.3%; - since 1998, 94–98% of those articles report at least one p≤0.05; - the authors call this “pervasive entrenchment”. - The author’s own field (2015) [S4]: “reliance on p values was standard, and non-significant findings were often called ‘negative’.”

Regulatory risk assessment (partial uptake of the chapter’s direction): - EFSA: - its Guidance on Uncertainty Analysis (2018) [S59] makes uncertainty analysis part of every assessment; - its 2022 benchmark-dose guidance [S60] recommends “a change from the frequentist to the Bayesian paradigm”, with model averaging and credible intervals. This confirms “Bayesian… gaining support” in regulatory toxicology. - The idea was not new to regulators. The benchmark-dose lower limit (BMDL) and upper-bound cancer slopes already built a confidence-limit worst-case logic into regulatory toxicology before 2013. The chapter’s proposal was new mainly for how epidemiological and academic results are reported. - Direct political opposition to the worst-case component (US Executive Order 14303, “Restoring Gold Standard Science”, 23 May 2025) [S61]: - Sec. 4(e): “Highly unlikely and overly precautionary assumptions and scenarios should only be relied upon… where required by law or otherwise pertinent.” - Sec. 1 cites a worst-case whale-population projection and a high-emissions climate scenario as examples of misleading science. - The same order also requires agencies to “acknowledge and document uncertainties” (Sec. 4(c)). So it uses the chapter’s language of uncertainty while rejecting its worst-case emphasis.

Verdict: partly held up. - The diagnosis was strengthened by the statistics profession’s endorsement [S52, S54, S55] and by Bayesian uptake at EFSA [S60]. - The remedies were adopted only unevenly: - dichotomous testing remains entrenched [S58]; - a rival reform (stricter thresholds [S53]) gained traction; - the “upper limit as plausible worst case” became politically contested in the US [S61].

Implication for weight. - The lesson “ask how large an effect a study could have missed” remains strong and is now mainstream statistical advice. - Whether worst-case framing is legitimate is a political variable, not a settled technical one. A lens built on this chapter should expect the same statistical tools to be labelled either prudent or “overly precautionary” depending on who governs. - Anyone quoting the chapter’s statistics text should not reproduce its definitions (pp. 632–633).

What the chapter says. - Triclosan and “perfluorinated octanoic sulfate” (a garbled name, probably PFOS) each got about two dozen articles a year, against about 35 on lead for each one (p. 628). - Booster biocides and Gaucho® (imidacloprid) received “only a little attention in independent research”. The chapter calls this unfortunate given the commercial interests involved (p. 636).

Subsequent developments.

Research volume (own Europe PMC counts [S8]; see table under Claim 1): - Grew sharply: the PFAS family (×7.2), neonicotinoids (×11.3), imidacloprid (×4.5) and triclosan (×3.4), against database growth of ×2.07. - Lagged: diuron (×1.4). - Fell: dichlofluanid (×0.76).

Regulatory action: - Triclosan: - US FDA final rule (6 Sept 2016) [S62]: triclosan and 18 other ingredients in consumer antiseptic washes are “not generally recognized as safe and effective”. Either “no additional data were submitted” or the data were insufficient. The missing data covered long-term safety, “potential hormonal effects” and bacterial resistance. - This is a burden-of-proof decision: a data gap led to removal, as the chapter would want. - EU: - a 2014 cosmetics amendment restricted triclosan’s uses [S64]; - Decision (EU) 2016/110 refused approval for biocidal hygiene products (product-type 1) on environmental risk grounds [S63]. - Neonicotinoids (imidacloprid, clothianidin, thiamethoxam): - EU Implementing Regulation 485/2013 (24 May 2013) [S65]: restricted uses and banned the sale of treated seed. - 2018/783 (and companion acts) [S66]: confined imidacloprid to permanent greenhouses. - Imidacloprid’s EU plant-protection approval expired on 1 December 2020 [S70]. - The CJEU dismissed Bayer’s appeal on 6 May 2021 (C-499/18 P) [S67], upholding the 2013 restrictions under the precautionary principle. - The CJEU ruled on 19 January 2023 (C-162/21) [S68] that member states may not use emergency authorisations for banned neonicotinoid-treated seeds. - Regulation (EU) 2023/334 [S69] lowered clothianidin and thiamethoxam residue limits, including on imports, from 7 March 2026. It explicitly cites global pollinator decline. This is an unusual use of food residue limits for an environmental purpose. - Imidacloprid remains approved as a biocide (product-type 18 insecticide), with the expiry postponed again in September 2025 [S74]. - US: EPA issued proposed interim registration-review decisions in February 2020 [S71]. I found no Federal Register notice of final decisions through September 2026. - Diuron: EU plant-protection approval “expired on 30 September 2020 and no application for renewal… had been submitted”. Residue limits were set at detection limits in 2023 [S72]. - Dichlofluanid: approved in 2017 for antifouling products (product-type 21), from 1 November 2018 to 31 December 2025, with conditions and a note that risk acceptability “should however be further confirmed” [S73]. I did not verify whether it was renewed. - PFOS/PFAS: see Claim 2.

Verdict: held up (for triclosan, PFAS and imidacloprid/neonicotinoids). Mixed for the booster biocides: diuron lapsed, dichlofluanid was approved.

Implication for weight. - The chapter’s judgement about which hazards were neglected was good. Most of its named emerging hazards became subjects of intense research and regulation within 3–10 years. - This supports the claim that neglect was a real gap, not a sign of low risk. - The mechanism, though, was mostly not the chapter’s recommended one (stakeholder-driven research agendas). Attention came from regulatory burden-of-proof reviews (FDA, the EU biocides and pesticides regimes) and public controversy (bees). - Several decisions turned on absence of data rather than proof of harm: the triclosan rule, diuron non-renewal. That supports the chapter’s broader point that governance need not wait for proof.

Claim 7: Exposure-limit ratchet: “nearly all exposure limits for hazardous agents have decreased” as evidence showed harm at lower levels; so standards set on incomplete evidence are too lenient (p. 624)#

What the chapter says. - “With time, nearly all exposure limits for hazardous agents have decreased as new evidence documented that harm occurred at lower exposure levels… Thus, when scientific evidence is incomplete, environmental standards are more lenient” (p. 624).

Subsequent developments.

Continued ratchet (consistent): - Lead: - US CDC blood lead reference value cut from 5 to 3.5 µg/dL (Oct 2021), with the statement that there is “no known safe blood lead level” [S75]; - EU Directive 2024/869 (13 March 2024) [S76]: - the occupational limit falls to 0.03 mg/m³; - the binding biological limit falls to 30 µg/100 mL until 2028 and 15 µg/100 mL from 2029, from 70 previously (still permitted during transition for workers exposed before April 2026); - it states it is “not scientifically possible to identify a level below which exposure to lead… would be safe”; - US EPA Lead and Copper Rule Improvements (30 Oct 2024) [S77] require replacement of lead service lines. They also lower the lead action level (to 10 µg/L from 15; the value is from my knowledge, not verified in the abstract I fetched). - Fine particles (PM2.5): - WHO Air Quality Guidelines (22 Sept 2021) [S105] lowered guideline levels because of health effects “at even lower concentrations than previously understood”. The annual PM2.5 guideline went from 10 to 5 µg/m³; those values are my knowledge, as the page summary did not state them. - EU Directive 2024/2881 [S78] cuts the annual PM2.5 limit to 10 µg/m³ by 2030, from 25. - US EPA lowered the annual PM2.5 standard from 12 to 9 µg/m³ (6 March 2024) [S79]. - PFAS: see Claim 2 [S16, S17, S23]. - BPA: see Claim 3 [S33, S35].

Counterexamples: - Nickel (EFSA, Nov 2020) [S80]: TDI raised from 2.8 to 13 µg/kg bw/day after a revised benchmark-dose analysis of the same critical effect. It was a methodological change, not new evidence of safety, but the limit went up. - Perchlorate (US EPA) [S81]: - 2008 interim health advisory: 15 µg/L; - 2019 proposed limit: 56 µg/L; - January 2026 proposed health goal: 20 µg/L, with enforceable-limit options of 20, 40 or 80 µg/L. - The later values rest on a more elaborate biologically based model of iodide uptake in susceptible pregnancies, and are higher than the 2008 value. - Policy reversals rather than evidential ones: - the US 2026 proposal to rescind four PFAS limits [S19]; - a 2025 US proposal to reconsider whether workers’ protective equipment may be assumed in TSCA risk evaluations [S82]. This is a default-assumption change of the kind listed in Table 26.3 (p. 630). - No systematic analysis. I found no post-2013 study of the direction of exposure-limit revisions across agents, so “nearly all” remains an untested generalisation built from salient cases.

Verdict: partly held up. - For the well-studied toxicants the report focuses on (lead, particles, PFAS, BPA), the direction held strongly after 2013. - “Nearly all” is not established. Limits can also rise with methodological changes (nickel, perchlorate), and they can be loosened for political reasons.

Implication for weight. - As a heuristic, “early limits set on incomplete evidence tend to be too lenient” deserves moderate–strong weight for agents with low-dose, developmental or immune effects. It deserves only moderate weight as a general law. - The selection bias the digest noted (Table 26.3 shows only failed assumptions) applies here too.

Claim 8: Sponsorship bias: industry-supported studies more often favour sponsors (pharma, Jørgensen 2006; “seems” also in toxicology, Myers 2009); eliminating financial ties is preferable to disclosure (pp. 636–637)#

What the chapter says. - Pharmaceutical studies sponsored by industry are “much more likely” to conclude safety and efficacy. “The same seems to happen in toxicology and environmental research” (p. 637). - Journals: it names Indoor and Built Environment and Regulatory Toxicology and Pharmacology as “trade magazines disguised as scientific journals” (p. 637). - Remedy: “complete elimination of financial ties may be the best way to secure trust-worthy research” (p. 636).

Subsequent developments.

Quantitative evidence: - Cochrane review update (Lundh et al., Feb 2017; 75 papers) [S83]: - industry-sponsored drug and device studies more often had favourable efficacy results (RR 1.27, 1.17–1.37; moderate quality) and favourable conclusions (RR 1.34, 1.19–1.51; low quality); - harms results were similar (RR 1.37, 0.64–2.93; very low quality); - the differences were not explained by standard risk-of-bias assessments. - Nutrition (Chartres, Fabbri & Bero, JAMA Intern Med, 2016) [S84]: favourable conclusions RR 1.31 (0.99–1.72), not significant, and no association with methodological quality. - I found no comparable post-2013 meta-analysis for environmental toxicology. The chapter’s “seems” remains appropriate.

Case evidence in chemicals: - PFAS documents [S24]: suppression of unfavourable research and distortion of public discourse. The authors found no evidence in this archive of funding favourable research. - 3M-funded immune review [S25]: reached a more exonerating conclusion than all later regulatory assessments (Claim 2). - Chlorpyrifos (Mie, Rudén & Grandjean, 2018) [S85]: - industry-funded developmental-neurotoxicity studies submitted to regulators contained brain-dimension changes “at all dose levels tested” that were “not… reported in the original test summary”; - the positive control failed to show neurobehavioural effects. - The EU did not renew chlorpyrifos’s approval in January 2020, citing genotoxicity and developmental neurotoxicity concerns [S86]. - Glyphosate review retracted (the journal the chapter named): In November 2025 Regulatory Toxicology and Pharmacology retracted Williams, Kroes & Munro (2000), a glyphosate safety review [S87]. - The reasons as reported by Retraction Watch (secondary) [S88]: “several critical issues” undermining its integrity, chiefly undisclosed contributions by Monsanto employees revealed in 2017 court documents. - The paper had about 614 citations.

Evidence that complicates the “industry vs independent” framing: - CLARITY-BPA [S39, S40]: the null core study was government-run. The disagreement tracked study design rather than funding (Claim 3). - Litigation money flows to both sides. The chapter’s author disclosed a litigation expert role for a state suing a PFAS producer [S23]. The authors of the 3M-funded review include a former 3M expert witness [S25].

Institutional response (disclosure and access, not elimination): - EU Transparency Regulation 2019/1381 (applied to applications from 27 March 2021) [S90]: - EFSA must make public the “scientific data, studies and other information supporting applications”; - business operators must notify EFSA of commissioned studies (Art. 32b); - the Commission may ask EFSA to commission verification studies in “serious controversies or conflicting results” (Art. 32d). - This directly addresses the chapter’s complaint that proprietary research is kept secret (p. 636). It does so through transparency and verification, not by eliminating industry funding. - US Executive Order 14303 (2025) [S61] lists “without conflicts of interest” among its “Gold Standard Science” tenets. It is being applied by an administration that is also relaxing several chemical limits [S19, S82]. The rhetoric of independence is available to either side.

Verdict: partly held up. - The pharmaceutical claim held [S83]. - The toxicology extension is supported by accumulating case evidence [S24, S85, S87], but still lacks quantitative synthesis. - The preference for eliminating financial ties was not adopted anywhere I could find. The main institutional response was compulsory disclosure plus public access to industry studies [S90].

Implication for weight. - “Funding shapes what is studied and concluded” deserves moderate–strong weight. Post-2013 cases show the mechanism often works through what is reported and disclosed (summaries that differ from raw data, ghostwriting, suppression) more than through study results as such. - That points to access to raw data and authorship transparency as the practical lever. It is a more transferable lesson than the chapter’s call to eliminate ties.

Claim 9: Recommendations: stakeholder-informed research topics; innovative and complementary rather than repetitive research; communication of possible magnitude; open and independent research (p. 639)#

What the chapter says. Four “ways to improve scientific evidence for robust and precautionary decisions” (p. 639), within a “PATIO” research culture: Participatory, Accessible, Transparent, Inventive, Open-minded (Table 26.5, p. 638).

Subsequent developments.

Uptake in research-programme design (mainly EU): - Stakeholder- and regulator-informed prioritisation: - HBM4EU (2017–2022) [S12]; - PARC (2022–2029; €400m; with ECHA, EFSA and the EEA as partners; aligned to ECHA’s regulatory challenges) [S10, S11]. - Note that the “stakeholders” are mainly ministries and agencies. The chapter’s “Participatory” ideal was broader. - Innovative and complementary research: - PARC’s hazard work package focuses on new approach methods and regulatory “gaps” [S11]; - its BPA-alternatives testing is explicitly anticipatory [S42]. - Communication of magnitude: - EFSA’s 2018 uncertainty guidance [S59]; - the Bayesian benchmark-dose guidance with credible intervals [S60]. - Open and independent research: see Claim 10, and the EU Transparency Regulation [S90]. - Funding for health and safety research. Hansen & Gee (2014), prepared alongside this report, found environmental, health and safety research took just 0.6% of EU research and development funding since 1996. They proposed 5–15% [S7]. I found no evidence that such a share was adopted. - PATIO as a label has almost no uptake. A Europe PMC search found no substantive use of the acronym.

Uptake in the US (divergent): - TSCA reform (2016) created mandatory prioritisation and risk evaluation, but started with long-studied chemicals [S9]. - 2025: Executive Order 14303 restricts reliance on precautionary scenarios [S61]. - Research capacity: a secondary source reports that EPA eliminated its Office of Research and Development in July 2025 [S103]. I could not verify this against an EPA primary source.

Did it improve timeliness? - No evidence that it did. - Time from warning to action remains long: - PFAS: public evidence from about 2000 (and industry knowledge from 1970) to the EU TWI in 2020 and US limits in 2024 [S24, S15, S17]; - BPA: EFSA’s 2023 TDI and an EU ban with transitions to 2026–2029 [S36]. - The chapter’s author concluded in 2019 that the improved documentation of developmental toxicity “has resulted in little progress in protection” [S102]. - Etzel, Grandjean & Ozonoff (2021) described environmental epidemiology as “in a crossfire” of doubt-raising critiques [S100].

Verdict: partly held up. - The recommendations were taken up in European research-programme architecture (PARC, HBM4EU, the transparency rules). - The chapter’s implied promise, that better-targeted, open research would enable earlier protection, is unclear at best. No evaluation shows faster action, and the author’s own later assessments say it has not happened.

Implication for weight. - The recommendations are sensible and partly institutionalised, but untested as causes of faster protection. - A cautious lens: stakeholder-informed agenda-setting changes which questions are studied. Whether that shortens the path from warning to action depends on decision processes the chapter only gestures at (“transparent and democratic procedures”, p. 638).

Claim 10: Openness: access to research “is improving” through open access; data sharing is spreading but risks “hostile analyses” (p. 639)#

What the chapter says. - Access: access to information “is improving” through open access, the Commission’s 12-month mandate, Wellcome requirements and Dutch repositories (p. 639). - Data sharing: some funders now require data-sharing strategies. But “such further analyses are not entirely benevolent… Hostile analyses have occurred” (citing Pearce & Smith 2011; p. 639).

Subsequent developments.

Access expanded substantially: - Open-access share: at least 28% of all scholarly literature was open access, and 45% of 2015 articles (Piwowar et al. 2018) [S95]. - Plan S: announced 4 Sept 2018, effective 2021. It requires immediate open access for publications funded by cOAlition S members. A 2026–2030 strategy was published in November 2025 [S94]. - US OSTP memorandum (25 Aug 2022): no embargo on federally funded publications; agencies to implement by 31 Dec 2025 [S91]. - NIH: its public access policy took effect 1 July 2025 [S92]. Its data management and sharing policy took effect 25 Jan 2023 [S93]. - Regulatory studies: in the EU food chain, industry studies supporting applications are now published proactively [S90].

Openness demands used to exclude epidemiology: - US EPA “Strengthening Transparency in Regulatory Science”: - proposed 30 April 2018 [S96]; supplemental proposal 18 March 2020 [S97]; - final rule 6 January 2021 [S98]: in significant regulatory actions EPA “will give greater consideration to studies where the underlying dose-response data are available”. In practice this would have down-weighted epidemiology based on confidential health records; - vacated by the US District Court for Montana. EPA formally removed it on 2 June 2021 [S99]. - Executive Order 14303 (May 2025) [S61] revived related elements: - public availability of data and models behind influential scientific information; - “weight of scientific evidence” with “replicability” as a quality criterion; - limits on “overly precautionary” scenarios. - Soskolne et al. (2021) [S101] catalogued methods used “to manipulate epidemiological findings”, including doubt-raising reanalysis and demands for unattainable certainty. The same journal’s editors called it good news [S100]. This formalises the chapter’s warnings about “hostile analyses” and “sound science” (pp. 632, 639).

Balance: - Openness also exposed industry evidence. The chlorpyrifos reanalysis [S85] and the PFAS document archive [S24] depended on access to raw data or documents. Openness cut against industry positions as well as against epidemiology.

Verdict: strengthened. - Access improved far more than the chapter anticipated. - Its warning that openness can be turned against unwelcome research was borne out in a major regulatory episode [S96–S99], and that episode’s logic re-emerged in 2025 [S61].

Implication for weight. - The lesson that openness cuts both ways deserves strong weight. - A general lens should expect demands for transparency, replicability and “gold standard” rigour to be used as tools. The question is not whether openness is good but who can demand what, at what stage, and with what effect on which kinds of evidence (e.g. confidential health records versus proprietary test data).

Implications for the digest’s transferable insights#

Sources#