COVID: What Actually Held Up?

I have been thinking about doing something that sounds straightforward but is not.

Take the things we were told during COVID and sort them into three piles:

• things that later evidence supported;
• things that later evidence disproved or showed were materially overstated; and
• things that remain unresolved, depend on context, or require a policy judgement.

I wanted the facts rather than the marketing. I also wanted to look at claims people were discouraged or prevented from making, because a platform removing a claim did not make it scientifically false.

The difficulty is that many public statements were compressed into slogans. "Masks work." "Vaccines stop transmission." "Stay two metres apart." "Natural immunity does not count." "It was just flu." Each sentence hides an outcome, a population, a date, a variant, a setting and a level of uncertainty.

This is not an audit of every sentence uttered during the pandemic. It is a UK-centred audit of 27 representative, consequential propositions, checked against primary evidence and major systematic reviews available by August 2026.

It is public information and commentary, not medical advice.

Scope and standard of proof

I have treated a proposition as supported or well established when good later evidence supports the claim within a clearly stated scope.

I have treated it as false, disproven or materially overstated when its ordinary public meaning is contradicted by better evidence, or when an important qualification changes the substance of the claim.

I have used mixed, unresolved or dependent on context when the evidence cannot sustain a single universal answer. Lockdown proportionality and school closure are explicitly labelled policy judgements because effectiveness is only one part of deciding whether a policy was justified.

The classifications are mine, not labels assigned by the sources. A reader may reasonably place a boundary differently. That is why the wording of each proposition matters more than the colour of its box.

The 27-claim scorecard

Every row links directly to the evidence used for its verdict. Where the literature disagreed, I have included both the studies that supported the claim and the larger or more rigorous evidence that changed the weight of the conclusion.

Prevention and transmission

The clearest later evidence is that SARS-CoV-2 was substantially transmitted through shared air. In May 2021, the CDC's scientific brief described inhalation of infectious respiratory particles, including transmission beyond six feet in enclosed, poorly ventilated spaces. It judged contaminated surfaces a possible route, but not a substantial contributor in most situations.

That does not make early handwashing foolish. It changes the hierarchy. Fresh air, ventilation, filtration, time, crowding and effective respiratory protection deserved more attention than repeatedly disinfecting low-risk surfaces.

People could also transmit before symptoms, or without ever developing recognised symptoms. A living systematic review found both presymptomatic and asymptomatic transmission, while generally finding lower secondary-attack estimates from persistently asymptomatic cases than from symptomatic ones.

Distance mattered, but there was no invisible wall at two metres. In June 2020, SAGE said the relationship was continuous and non-linear. It advised two metres where possible, while warning that two metres should not be treated as an absolute rule.

The public rule was simple. The underlying evidence was conditional.

Masks were not one thing

A loose cloth covering, a surgical mask and a well-fitting respirator are not interchangeable. Later CDC and NIOSH guidance placed loosely woven cloth at the lower end of protection and approved, tight-fitting respirators at the higher end. Fit and leakage matter as well as the filter material.

The policy evidence is harder. In Bangladesh, a complete village-level package provided masks, promoted their use and monitored behaviour. Proper mask use rose from 13.3% to 42.3%, observed distancing also rose modestly, and symptomatic seroprevalence was about 9% lower. The 95% confidence interval ranged from roughly no reduction to 18%.

That is useful evidence. It is not evidence of a universal, predictable effect from every mandate.

Cochrane's 2023 review described the randomised evidence as generally low to moderate certainty. For its principal medical or surgical mask comparisons it used moderate certainty, while emphasising high or unclear risk of bias, heterogeneous settings and low adherence. Cochrane later said the review had been widely misinterpreted as proving masks do not work. The Royal Society, looking across evidence types, concluded that mask wearing and mandates generally, though not universally, reduced infection.

The fair conclusion is that filtration works, better masks and better fit work better, and the real-world effect of a policy depends on what people actually wear, where, how consistently and alongside which other measures.

Lockdowns and school closures are not one-number questions

Reviews found that packages of non-pharmaceutical interventions reduced transmission. That is an empirical result. It does not by itself prove that every restriction was necessary or proportionate.

Restrictions arrived in bundles. Behaviour changed before and after the legal rules. Variants, immunity, season and compliance changed. Harms were unevenly distributed. The narrower the question becomes, the harder it is to isolate one measure from the rest.

School closures show why the policy category matters. The UK's technical review found that attendance restrictions could reduce transmission as part of wider packages. The Department for Education and Ofqual evidence also found lost learning, impaired mental health, loss of support services and particularly severe consequences for disadvantaged and vulnerable children.

So "did it reduce transmission?" and "was it worth the harm?" are different questions. The second cannot be answered by an infection curve alone.

Vaccines, immunity and adverse events

The original vaccine results were genuinely impressive.

In the Pfizer-BioNTech randomised trial, the consistent analysis population contained 17,411 vaccine recipients and 17,511 placebo recipients at risk from seven days after dose two. There were 8 symptomatic cases in the vaccine group and 162 in placebo. That produced the famous 95% relative efficacy result against symptomatic disease during that trial period.

What the trial did not show was permanent sterilising immunity. SAGE had warned as early as June 2020 that true sterilising immunity from either vaccination or infection was unlikely.

Later evidence showed both parts of the story. Protection against infection and symptomatic disease waned, and Omicron changed performance materially. Protection against hospitalisation and death generally held up better. A 2022 systematic review and meta-regression found that protection against severe disease remained high at six months, although it had declined somewhat. An Israeli national surveillance study during the early 2021 vaccination campaign estimated 96.7% effectiveness against COVID-related death at least seven days after dose two, with a 95% confidence interval from 96.0% to 97.3%.

That means the statement "vaccines reduced severe disease" held up. The statement "vaccination permanently stops infection and transmission" did not.

Natural and hybrid immunity

Previous infection produced real protection. The UK SIREN study followed healthcare workers with regular testing and found infection-acquired protection that waned. In that pre-Omicron cohort, vaccination after infection produced strong additional protection.

A later systematic review focused on Omicron found that hybrid immunity generally gave the highest and most durable protection against hospital admission or severe disease. That does not make deliberate infection a sensible route to immunity. It means the immune system remembers both infection and vaccination.

If "natural immunity did not count" was meant as a biological claim, it was wrong. Sometimes, however, "did not count" described a documentation or policy rule. Biology alone does not resolve whether vaccination should have been compulsory for a particular previously infected person. That also involves age, risk, timing, product, ethics, documentation and the costs of the policy.

The same is true of repeated boosting. Later JCVI advice narrowed the UK programme towards people at higher risk of serious disease. That is evidence that the balance of incremental benefit changed with population immunity and risk. It is not evidence that every earlier dose was useless, or that every person benefited equally.

Safety required two truths at once

Rare serious vaccine side effects were real. MHRA identified an association between AstraZeneca vaccination and very rare thrombosis with thrombocytopenia. Myocarditis after mRNA vaccination was also real and concentrated in particular age and sex groups.

At the same time, two separate matched analyses in an Israeli population, each with 42-day follow-up, estimated an excess of 2.7 myocarditis events per 100,000 people after Pfizer vaccination and 11 per 100,000 after SARS-CoV-2 infection. Those population estimates do not settle every individual decision, but they show why "the vaccine had a risk" and "infection carried a larger set of risks" could both be true.

Treatments, mortality and statistics

Some treatments failed. One steroid worked in the right patients.

Ivermectin had laboratory plausibility, and it would be wrong to imply that no study ever reported a benefit. Several 2021 meta-analyses did. A review by Bryant and colleagues reported a mortality risk ratio of 0.38, while another by Zein and colleagues reported 0.39. The latter explicitly rated the evidence low certainty; other early reviews noted that many included trials had a high risk of bias.

The evidence changed as larger and more rigorous trials arrived. TOGETHER and ACTIV-6 did not find a meaningful clinical benefit for the outpatient regimens studied. Cochrane's updated review applied explicit trustworthiness criteria, excluded seven of the 14 trials in its earlier evidence set, included 11 trials with 3,409 participants, and concluded that the available evidence did not support using ivermectin to treat COVID. It found no eligible prevention trial.

That is why the audited proposition is carefully worded. The conclusion is not "ivermectin was proved useless in every possible dose and population." It is that ivermectin was not established as an effective routine COVID treatment by reliable clinical evidence.

Hydroxychloroquine followed a similar route. Early laboratory results and uncontrolled studies created excitement. The large UK RECOVERY randomised trial found no mortality or recovery benefit in hospitalised patients.

Dexamethasone was different. RECOVERY found lower 28-day mortality among hospital patients who were receiving oxygen or invasive mechanical ventilation, but not among those receiving no respiratory support. WHO therefore recommended systemic corticosteroids for severe or critical COVID, not for non-severe disease.

There is an important statistical detail here. RECOVERY's primary endpoint was all-cause mortality within 28 days of randomisation. Technically, if a participant died during that period after a fatal accident, from unrelated sepsis, or from another cause, that death still counted in the endpoint. The trial did not label every such death as caused by COVID.

That is deliberate. The patients had been admitted to hospital with clinically suspected or laboratory-confirmed COVID and were randomly allocated to dexamethasone or usual care. Randomisation is designed to make the risk of unrelated deaths comparable between groups in expectation. The trial then compared all-cause mortality rates, not raw death totals, without having to make a potentially subjective decision about the cause of each death. RECOVERY also recorded cause-specific mortality, but its main treatment comparison was all-cause mortality.

"Steroids work" was too vague. The evidence-supported statement was: dexamethasone helped selected hospital patients at the inflammatory, oxygen-requiring stage of disease.

The ventilator story was more complicated than the slogan

There was also advice about when people should be intubated and put on invasive mechanical ventilation. That advice is often remembered as, "COVID patients were put on ventilators too early." There is a real issue underneath that memory, but the simple version goes too far.

The March 2020 Surviving Sepsis Campaign guidance did not recommend invasive ventilation for everyone whose oxygen level fell. It suggested high-flow nasal oxygen when conventional oxygen failed, a trial of non-invasive support when high-flow oxygen was unavailable, close monitoring, and early controlled intubation if the patient worsened. Infection-control concerns and uncertainty about whether non-invasive support increased aerosol exposure also influenced early practice.

Later evidence widened the route before intubation. In the UK RECOVERY-RS randomised trial, the combined outcome of tracheal intubation or death within 30 days occurred in 36.3% of patients assigned to CPAP and 44.4% assigned to conventional oxygen. The difference was driven by fewer intubations; CPAP did not show a statistically significant mortality reduction, and it caused more adverse events. High-flow nasal oxygen did not significantly improve the primary outcome in that trial.

Studies comparing early and later intubation were observational, used inconsistent definitions and produced conflicting results. One 2022 meta-analysis of four cohorts found no statistically significant mortality difference. The responsible conclusion is not that ventilators killed people, nor that delaying intubation was always safer. It is that early invasive ventilation was not established as the universally superior first escalation. Selected patients could benefit from closely monitored non-invasive support, while people with urgent indications still needed prompt intubation.

Science changing its mind after a better trial is not proof that science failed. It is what the correction mechanism is meant to look like.

The mortality burden was real and unequal

ONS estimated 167,356 excess registered deaths in England and Wales from March 2020 to December 2022. The figures included non-residents and the 2022 data were provisional. ONS used 2015 to 2019 as the comparison for 2020 and 2021, and 2016 to 2019 plus 2021 for 2022. When deaths with COVID as the underlying cause were removed, the remaining excess fell to 17,288. When every death mentioning COVID was removed, deaths were 10,964 below the comparison baseline.

That does not mean every excess death was caused directly by infection. It does show that "the deaths were mostly invented by incidental positive tests" is not a credible description of the mortality record.

The burden was also radically unequal. OpenSAFELY analysed 17,278,392 adults registered with English general practices from 1 February to 6 May 2020 and found strong gradients by age, sex, deprivation and clinical condition. People aged 80 or over had an adjusted hazard of COVID-related death 20.60 times that of people aged 50 to 59, with a 95% confidence interval from 18.70 to 22.68.

A population-wide survival percentage hides that distribution. It can be numerically true and still be a poor description of the danger faced by an older or vulnerable person.

Deaths were counted in several different ways

The UK dashboard's "within 28 days of a positive test" count was designed for speed. It was not the same thing as a doctor certifying COVID as the underlying cause of death.

ONS distinguished deaths due to COVID, where it was the underlying cause, from deaths involving COVID, where it appeared anywhere on the certificate. In 2021, 77,727 registered deaths in England and Wales involved COVID, and 67,350 of those were due to it.

So, yes: the rapid dashboard measure could include somebody who died after a fatal accident, complications of a broken leg, unrelated sepsis, cancer, or another cause if they had tested positive in the previous 28 days. That did not mean the death certificate said COVID caused the death. Conversely, the dashboard could miss somebody whose COVID-related death occurred more than 28 days after their positive test.

This limitation was known. For England, UKHSA reported that during 2020 and 2021, between 80% and 90% of deaths in the 28-day measure also mentioned COVID on the death registration. The measures diverged as population immunity rose and incidental infections became more common; UKHSA stopped recommending the 28-day count as the leading measure in 2023. The right response is to name the measure, geography and period, not to pretend dashboard, certificate and excess-death statistics were interchangeable or fraudulent.

Relative and absolute risk can both be true

Pfizer's 95% result was a relative reduction in symptomatic cases. Using the same people-at-risk population from the published table, the crude risks over that follow-up window were approximately 0.046% in the vaccine group and 0.925% in placebo, an absolute difference of about 0.88 percentage points.

Put more simply, imagine 10,000 people in each trial group under the same short trial conditions. The published figures correspond to roughly 93 symptomatic cases per 10,000 people in the placebo group and about 5 per 10,000 in the vaccine group. The difference is about 88 fewer cases per 10,000 people. That is an absolute reduction of about 0.88 percentage points.

Now compare the two groups rather than the whole population: 5 is about 95% lower than 93. That is the relative reduction of about 95%. It did not mean 95 out of every 100 vaccinated people were saved from an infection they would otherwise certainly have had. Nor does the 0.88-point figure mean the vaccine was "only 0.88% effective". The two percentages answer different questions.

Both figures are true. Neither should travel alone. The relative figure describes the contrast between groups. The absolute figure depends on how much virus was circulating, who was enrolled, the follow-up period and which outcome was counted.

This article does not turn that short trial window into a number needed to vaccinate. Doing so without the endpoint, population, exposure period and changing background risk would manufacture more certainty than the calculation deserves.

A model scenario is not a forecast

COVID models were often reported as if they were predictions. SAGE explicitly said its reasonable worst-case scenarios were not predictions or forecasts. The later UK technical report explained that scenarios show possible futures under assumptions about behaviour, policy and biology.

A scenario can be useful precisely because it changes a decision and therefore does not come true. It can also be wrong because an assumption was wrong. The honest communication is: "If these conditions hold, this is one possible outcome." It is not: "This is what will happen."

Moderation, origins and unresolved questions

The origin of SARS-CoV-2 remains unresolved.

In 2025, the WHO's independent SAGO group said the weight of available evidence suggested zoonotic spillover, directly from bats or through an intermediate host. It also said critical evidence had not been provided and that all hypotheses, including a laboratory incident, had to remain on the table.

That is not a satisfying answer. It is the honest one.

Removal was not a scientific verdict

Meta's own policy record is instructive. In February 2021, it listed the claim that COVID-19 was "man-made or manufactured" among claims it would remove. In May 2021, while investigations continued, it stopped removing that claim.

"Man-made or manufactured" is not synonymous with an accidental laboratory incident involving a naturally occurring virus. The policy reversal showed that Meta changed its moderation boundary while the origin evidence remained incomplete.

It proved neither origin hypothesis.

A removal is evidence that a platform applied a policy at a particular time. It is not evidence that the claim was false. A later decision to permit the claim is not evidence that it became true.

Some removed claims were plainly dangerous. Others touched live scientific questions. A responsible platform has to manage immediate harm, but it should preserve the exact claim, the date, the evidence used and a route for challenge and correction.

Being censored does not make a claim true. Being permitted does not make it true either.

What should be communicated differently next time

The lesson is not that every official statement was a lie. It is not that every dissenter was right. It is that emergency communication repeatedly collapsed conditional evidence into certainty.

1. Publish the exact claim. Say which outcome, population, period and variant it covers.
2. Separate evidence from policy. "This reduces transmission" is different from "therefore it must be compulsory".
3. Show relative and absolute numbers. Include the denominator and follow-up period.
4. Name the measure. Dashboard, death certificate, excess mortality, detected case and estimated infection answer different questions.
5. Label scenarios. Show assumptions and distinguish a possible future from a forecast.
6. Date every conclusion. The virus, immunity, treatments and evidence change.
7. Preserve disagreement. Record why a claim was rejected and what evidence would change the decision.
8. Correct publicly. Update the claim rather than quietly deleting the old certainty.

I do not think trust is rebuilt by insisting that institutions were always right.

Trust is rebuilt by showing the evidence, the uncertainty, the corrections and the boundaries of what we know.

That is slower than a slogan.

It is also much closer to the truth.

Related reading

• But for the Grace of God Go I
• The AI Screwed Up. Then What?
• What Has The NHS Ever Done For Us?

Sources and notes

• UK SAGE minutes, 4 June 2020. Distance-related risk was continuous, two metres was not an absolute boundary, and sterilising immunity should not be assumed.
• CDC archived scientific brief on SARS-CoV-2 transmission. Respiratory-particle transmission and the later assessment that surfaces did not contribute substantially in most circumstances.
• Living systematic review of asymptomatic and presymptomatic transmission. Transmission without recognised symptoms and differences in secondary-attack estimates.
• Lancet systematic review of distancing and protective equipment. Lower transmission risk with physical distancing, with important evidence limitations.
• CDC/NIOSH respirator and mask guidance. Filtration, fit and leakage differ across cloth coverings, surgical masks and respirators.
• Bangladesh cluster-randomised community masking trial. A complete promotion package increased use, modestly increased distancing and produced a modest, imprecisely estimated reduction in symptomatic seroprevalence.
• Cochrane physical-interventions review, Cochrane's statement on its interpretation, and the Royal Society NPI review. Used to distinguish filtration, individual use and real-world policy effects.
• UK technical report: NPIs in education settings, SPI-B and DfE evidence on remaining in education, and Ofqual's English learning review. Transmission effects and documented educational, health and inequality costs.
• Original Pfizer-BioNTech trial. Source for the consistent analysis population, 8 versus 162 symptomatic cases, 95% relative efficacy and the short blinded follow-up.
• Systematic review of vaccine protection over time and UKHSA Omicron and Delta hospitalisation study. Outcome-specific protection and the stronger durability of protection against severe disease.
• Lancet study of Israel's nationwide Pfizer vaccination campaign. Early 2021 adjusted effectiveness estimates for infection, hospitalisation, severe disease and COVID-related death, with outcome-specific denominators and confidence intervals.
• UK SIREN study and systematic review of previous infection and hybrid immunity. Infection-acquired and hybrid protection, with variant, time and population boundaries.
• JCVI advice for 2025 and spring 2026. Later UK policy narrowed as immunity, disease risk and programme objectives changed.
• Nationwide vaccine-safety study and MHRA AstraZeneca safety notice. Serious adverse events, their rarity and scope.
• Bryant et al. ivermectin meta-analysis and Zein et al. ivermectin meta-analysis. Early syntheses that reported apparent mortality benefit, included here so the positive literature is visible alongside its certainty and bias limitations.
• RECOVERY dexamethasone trial and WHO dexamethasone guidance. The trial used all-cause mortality within 28 days and found benefit in oxygen-requiring or ventilated hospital patients, not non-severe disease.
• March 2020 Surviving Sepsis Campaign guidance summary, the RECOVERY-RS randomised trial, a 2022 early-versus-late intubation meta-analysis, and WHO respiratory-support guidance. These distinguish non-invasive support, controlled intubation when deterioration occurs, and urgent intubation that should not be delayed.
• Cochrane ivermectin review, TOGETHER trial and ACTIV-6 trial. Higher-quality evidence for the studied outpatient regimens and populations.
• Final RECOVERY hydroxychloroquine trial report. Hospitalised patients and no meaningful mortality benefit.
• ONS excess deaths, March 2020 to December 2022. Excess mortality with and without deaths due to or involving COVID.
• OpenSAFELY risk-factor study. More than 17 million adults in England and strong gradients by age and other characteristics.
• ONS deaths due to COVID registered in 2021, ONS mortality methodology, UKHSA's 2020 explanation of the 28-day measure, UKHSA's 2023 reporting review, and Office for Statistics Regulation explainer. Distinguishes rapid surveillance, death-certificate and excess-mortality measures, including unrelated deaths after a positive test.
• ONS comparison of COVID with flu and pneumonia. Age distribution and underlying-cause comparisons.
• UK technical report on COVID modelling and SAGE minutes, 30 July 2020. Model scenarios, assumptions and the explicit statement that reasonable worst cases were not predictions or forecasts.
• WHO SAGO origin assessment, 2025. The weight of available evidence and the evidence that remained unavailable.
• Meta's COVID-19 misinformation-policy record. Evidence of the platform's February and May 2021 policy changes, not evidence for either origin hypothesis.

Evidence checked 3 August 2026. This article audits a selected 27-claim corpus and does not provide individual medical advice.