Hypothesis

Ultralow-mass axion-like dark matter would induce an oscillating neutron electric dipole moment or an axion-wind spin-precession effect visible in the ratio of ultracold-neutron and ¹⁹⁹Hg precession frequencies.

What failed

No signal consistent with dark matter was observed for axion masses from 10⁻²⁴ to 10⁻¹⁷ eV.

Broken assumption

Axion-like dark matter in this mass range couples to nucleons or gluons strongly enough to show up in spin-precession measurements.

Why

No signal was present at the experiment's sensitivity; the null result sets the first laboratory limits on axion–gluon coupling, up to 1,000 times tighter than astrophysical limits.

From the published paper: Abel et al. reroDoc Digital Library, 2017. ↗CC BY

Hypothesis

Faces perceived as less dominant and more trustworthy reach awareness faster because of their social-emotional meaning.

What failed

Dominance- and untrustworthiness-related slowing replicated for upright faces but was equally strong for inverted faces, where social characteristics are hard to read; the reported correlation between that slowing and self-reported propensity to trust did not replicate; and the dominance effect persisted when only the eye region was shown, inverted and with reversed contrast polarity, conditions in which the traits were no longer discernible. Experiment 3 reproduced this under a preregistered protocol.

Broken assumption

Faster access to awareness for certain faces reflects their social meaning rather than low-level image properties.

Why

The authors conclude that physical differences in the eye region, unrelated to perceived dominance when shown in isolation, parsimoniously explain the effect.

From the published paper: Stein et al. Journal of Experimental Psychology General, 2018. ↗Public domain

Hypothesis

A brief mindfulness exercise or facial emotion recognition training would immediately improve empathic accuracy — how well people infer others' emotions.

What failed

Across three conditions of 23 psychology students each (mindfulness, emotion-recognition training, neutral lecture control), there was no significant overall difference in empathic accuracy, so neither target intervention had an effect.

Broken assumption

Empathic accuracy is acutely malleable by a single brief intervention.

Why

The paper reports only one qualification: among participants with limited understanding of the task's Dutch-language material, accuracy was better after emotion-recognition training than control. The authors list a small convenience sample, a mindfulness lecture as control, no baseline measure and no formal language assessment as limitations.

From the published paper: Rot et al. Journal of Trial and Error, 2023. ↗CC BY

Hypothesis

Short instructions sent to authors alongside peer-review comments would reduce spin in the conclusion of a revised abstract.

What failed

Of 184 manuscripts randomized at BMJ Open, 108 revisions could be assessed; spin was present in 57% (31/54) of intervention manuscripts versus 63% (34/54) of controls, a 6% difference with a confidence interval from 24% lower to 13% higher, and the counts for each prespecified spin type were similar between arms.

Broken assumption

Authors will drop overstated conclusions when a journal simply asks them to.

Why

The authors note the confidence interval rules out a large effect and call for different interventions against spin.

From the published paper: Ghannad et al. Journal of Clinical Epidemiology, 2020. ↗CC BY

Hypothesis

A submillimetre magnet levitated in a superconducting trap would detect the oscillating force from ultralight dark matter coupled to baryon-minus-lepton number.

What failed

With a force sensitivity of 0.2 fN/√Hz, the search found no evidence of a signal in the mass range 1.10360–1.10485 × 10⁻¹³ eV/c², giving a most stringent coupling limit of g(B−L) ≲ 2.98 × 10⁻²¹.

Broken assumption

Ultralight B–L dark matter in this mass window couples strongly enough to move a levitated magnet detectably.

Why

No signal was present at this first experiment's sensitivity and narrow mass range; the authors propose staged upgrades (POLONAISE) to widen both.

From the published paper: Amaral et al. Physical Review Letters, 2025. ↗CC BY

Hypothesis

The neutrophils that flood the airways in severe RSV infection would contribute to controlling the virus, to lung inflammation, or to how sick the host becomes.

What failed

In RSV-infected mice, antibody depletion of neutrophils did not substantially change type I interferons, IL-6, TNF-alpha or IL-1beta, nor viral load. Neither depleting neutrophils nor boosting them with CXCL1 changed peak disease severity measured by weight loss.

Broken assumption

Abundant airway neutrophils in severe RSV are a driver of disease rather than a bystander.

Why

The authors conclude that in this mouse model lung neutrophils neither help antiviral defence nor drive disease; their abundance in severe infant disease is not explained.

From the published paper: Kirsebom et al. Scientific Reports, 2020. ↗CC BY

Hypothesis

Immunostimulatory agonistic CD40 antibodies, in development for solid tumours, would stimulate antitumour immunity in glioma and support responses to checkpoint inhibitors.

What failed

In preclinical glioma models, systemic agonistic CD40 led to cytotoxic T-cell dysfunction and impaired the response to immune checkpoint inhibitors, even while it induced tertiary lymphoid structures.

Broken assumption

An immunostimulatory CD40 agonist can only add to, not subtract from, checkpoint-blockade efficacy.

Why

The impairment was associated with accumulation of suppressive CD11b+ B cells. In human glioma, the tertiary lymphoid structures correlated with greater T-cell infiltration, which the authors read as pleiotropic rather than purely harmful effects.

From the published paper: Hooren et al. Nature Communications, 2021. ↗CC BY

Hypothesis

Metabolic syndrome components would shape the clinical picture of LRRK2- and GBA-associated Parkinson's disease and the prodromal burden in unaffected carriers.

What failed

Across 562 participants (idiopathic, LRRK2 and GBA Parkinson's, non-manifesting carriers and non-carrier controls), the number of metabolic syndrome components did not differ between Parkinson's groups (p = 0.101) or between carrier groups (p = 0.685).

Broken assumption

Metabolic syndrome modifies how genetic Parkinson's disease presents.

Why

Only isolated components differed — higher triglycerides and more prediabetes in LRRK2 carriers — and the authors conclude metabolic syndrome as a whole does not appear to modify the genetic Parkinson's phenotype.

From the published paper: Thaler et al. Scientific Reports, 2020. ↗CC BY

Hypothesis

Long-term daily avocado-soybean unsaponifiables would limit cartilage and subchondral bone damage in a surgically induced rabbit model of knee osteoarthritis.

What failed

After 24 weeks of treatment, avocado-soybean unsaponifiables did not reduce hypertrophic cartilage response, surface fibrillation, trabecular bone loss or collagen remodelling markers compared with placebo, although osteoarthritis was successfully induced.

Broken assumption

Avocado-soybean unsaponifiables modify structural joint damage in osteoarthritis.

Why

The paper reports limited structural efficacy in this model under the conditions tested; it does not propose a mechanism for the lack of effect.

From the published paper: Sabucedo-Suárez et al. Frontiers in Veterinary Science, 2026. ↗CC BY