Antidepressants don't work. Statins wreck your memory. Hormone therapy causes cancer. Ozempic makes people suicidal. Here is what each study actually reported, what each scare cost in real patients, and the three times the alarm was right and nobody listened.
In October 2013 the BMJ published two papers questioning whether people at low cardiovascular risk should be on a statin. The British press ran with it for about five months. Researchers later went into UK primary-care records and counted the damage: more than 200,000 people stopped taking a statin they had already been prescribed.
The most-quoted number in those papers turned out to be wrong. They reported side effects in 18 to 20% of patients, a figure lifted from an observational study that did not support it. The authors withdrew the statement in 2014. An independent panel reviewed the papers and declined to retract them, so they still stand with a correction attached.
The correction ran seven months after the coverage. You have almost certainly never heard about it, and there is no version of you who should have. Corrections don't get shared, and the wrong number had already emptied a lot of pill organisers by then.
Nobody in this chain is the villain, so it is worth being precise about who does what. A researcher publishes a paper with a modest finding and a page of caveats. A university press office turns it into a release, because press offices are measured on coverage. A reporter turns the release into a story, because reporters are measured on readers. An editor writes the headline, because the headline is the only part most people read, and a headline that says "modest effect, wide confidence interval" is a headline that gets nobody promoted.
By the time it reaches you, four people have each made one small, individually defensible decision to sharpen it, and not one of them had to lie for the compounding to work.
Underneath that sits a structural gap nobody has solved: the alarming version of a study is news and the boring correction eighteen months later is not, so the alarm ends up with a hundred times the distribution of the retraction, forever, in every field. That is arithmetic rather than conspiracy, which is why complaining about it has never once helped.
These are the ones I have watched go past over about four decades. I picked them because in every case somebody afterward went and measured what the coverage did, which is rarer than it should be.
| The scare | The headline version | What the research actually reported | What it cost |
|---|---|---|---|
| Antidepressants2008, and again in 2018 | "Antidepressants don't work." Then ten years later, "It's official, antidepressants work." | Kirsch's 2008 analysis found the drugs did beat placebo, and judged the margin too small to be clinically meaningful. Cipriani's 2018 network meta-analysis of 522 trials found all 21 drugs studied beat placebo. The effect sizes in the two papers are closer than the two headlines suggest. | A decade of "they're just placebos" in the culture, and a lot of people quietly deciding not to raise it with a doctor. |
| Hormone therapy2002 | "HRT causes breast cancer." | The Women's Health Initiative found a hazard ratio of 1.26 for invasive breast cancer on combined estrogen plus progestin, reported everywhere as "26% higher". In absolute terms: 38 cases per 10,000 women per year versus 30. The estrogen-only arm showed fewer breast cancers than placebo. | US prescriptions fell roughly 40% within a year. A generation of women went without, on a relative number. |
| Statins, part oneOctober 2013 | "The statin side-effect epidemic." | Two BMJ papers questioned statins for low-risk patients and cited an 18–20% adverse-event rate. That figure came from an observational study that did not support it, and the authors withdrew it in 2014. A more cautious reading of the same data puts discontinuation from side effects nearer 9%. | An estimated 200,000+ UK patients stopped an existing statin, with 2,000–6,000 avoidable cardiovascular events projected over ten years. |
| Statins, part two2012 | "Statins give you memory loss." | The FDA added a cognitive warning to statin labels in February 2012, based largely on case reports of mild, reversible symptoms. A 2015 meta-analysis of 25 placebo-controlled trials, with cognitive test data pooled from about 28,000 participants, found no significant effect on any aspect of cognition. | Harder to measure, but "statins cause brain fog" is now folk knowledge, and it entered folk knowledge through a label change rather than a finding. |
| Antidepressants in teenagers2004 | "Antidepressants make kids suicidal." | This one had real trial data behind it: pooled trials did show more suicidal thoughts and behaviours on drug than placebo in under-18s, with no completed suicides. The FDA's black-box warning was a defensible reading of that evidence. | Adolescent antidepressant use fell 31% and young-adult use 24%, while psychotropic drug poisonings rose. A warning that was arguably correct still produced a worse outcome. |
| Heartburn pills and dementia2014–2016 | "Your acid-reflux pill will give you dementia." Same genre covered benzodiazepines. | The alarming studies were observational. A later meta-analysis pooling 642,305 patients found no association between proton-pump inhibitors and dementia, and a prospective cohort found none either. The benzodiazepine question is genuinely less settled, and the argument there is largely about reverse causation. | Nobody has counted this one properly, which is itself the point. Scares get measured; the follow-up rarely does. |
| GLP-1 drugs2023–2024 | "Ozempic makes people suicidal." | The signal came from spontaneous adverse-event databases, which anyone can file a report to and which cannot establish cause. The FDA's January 2024 review found no evidence the drugs cause suicidal thoughts or actions, and it later asked for the warning to be removed from labels. | Still running. Worth watching how much of the retraction reaches the people who saw the original. |
Every claim in this table is sourced at the foot of the page. Where I have written "reported", I mean the paper reported it, not that I am endorsing it.
In this one the loud, unqualified, deeply annoying internet was right, and the institution was wrong for five years.
Wellbutrin has form. Burroughs Wellcome pulled it in March 1986 after a study in bulimic patients showed an increase in seizures, worked out that the risk was dose-dependent, and brought it back in 1989 with a capped dose. That is the system doing its job in three years, mostly out of public view.
Then in 2006 generic versions of the 300 mg extended-release form arrived, and from 2007 patients started saying the same thing everywhere online: this is not the same drug. Depression coming back. Headaches, fatigue and anxiety getting worse after the pharmacy switched them.
This is exactly the profile of a story you learn to discount. Self-reported, on the internet, about a psychiatric drug, from people with a known nocebo mechanism available, in a country where brand loyalty is manufactured for a living. I would have discounted it. Most doctors did.
The generic had never been directly tested at 300 mg. The FDA had approved it by testing the 150 mg strength against the brand and extrapolating upward. When the agency finally ran the study on the 300 mg version, in results available in August 2012, Budeprion XL 300 mg failed to release bupropion into the blood at the same rate and extent as Wellbutrin XL 300 mg.
Approval was withdrawn in October 2012. The patients had been right for five years.
That case cured me of treating skepticism as a direction. Being reflexively dismissive is the same intellectual move as being reflexively alarmed, run backwards, and it fails for the same reason: in both cases you decided before you counted.
A page that only listed overblown scares would be doing the identical thing it complains about, pointed the other way. So here is the other column.
| The alarm | Who raised it, and when | How long it was waved off | What it turned out to be |
|---|---|---|---|
| Vioxx and heart attacks | The drug's own VIGOR trial, published November 2000, showed a large excess of heart attacks. It was explained away as the comparison drug being protective. Cardiologists pulled the full data off the FDA website and challenged that in print in 2001. | Four years. | Withdrawn September 2004 after a separate trial showed the risk directly. An FDA drug-safety reviewer estimated roughly 88,000 excess heart attacks in the United States, about 38,000 of them fatal. |
| Generic Budeprion XL 300 mg | Patients, on message boards and in adverse-event reports, from 2007. | Five years. | FDA testing confirmed it was not therapeutically equivalent to the brand. Approval withdrawn October 2012. |
| Wellbutrin and seizures | The manufacturer's own study in bulimic patients, 1986. | Not waved off. Voluntarily withdrawn within the year. | The risk was dose-dependent. Reintroduced in 1989 with a 450 mg daily cap and contraindications for patients with seizure risk. Included here because this is what the process looks like when it works, and it made no headlines at all. |
Two of these three came from inside the data rather than from a press release, which is roughly the opposite of how the scares in the first table reached you.
Four questions, about ninety seconds, and no statistical training required.
A shop near me could truthfully advertise 300% more free parking. They went from one space to four. The claim is accurate, the percentage is real, and it will not help you park.
"Doubles your risk of X" behaves the same way. If X happens to two people in 10,000, doubling it costs two more people in 10,000. If X happens to one person in three, doubling it is a catastrophe. The percentage alone cannot tell you which world you are in, which is precisely why it makes such a good headline. The hormone-therapy row above is the cleanest example on this page: 26% and "8 more women in 10,000 each year" are the same fact, and they produce completely different decisions.
These sound similar and are not remotely the same. A randomised trial measures: two groups, one gets the drug, somebody counts what happens. A spontaneous adverse-event database collects: patients, doctors and lawyers file reports of anything that happened to anyone taking anything, with no comparison group and no verification.
Report databases are genuinely useful for spotting things worth investigating. They cannot show that a drug caused anything, and they are strongly affected by how much news coverage a drug is getting, which produces a lovely feedback loop. The GLP-1 suicidality story came out of report databases. So did a great deal of what you have read about every drug that has ever been popular.
Sleeping pills are associated with later dementia. That association is real and it has been found repeatedly. The difficulty is that early dementia wrecks sleep and raises anxiety years before anyone gets a diagnosis, so a person in the earliest stage is more likely to be prescribed something to sleep. The pills would show up in the data as a warning sign even if they did nothing at all.
Researchers try to handle this by ignoring prescriptions written in the years just before diagnosis, and the honest state of the benzodiazepine question in 2026 is that doing so shrinks the excess risk without entirely removing it. Nobody has settled it. When you see a drug associated with a disease that has a long silent build-up, that ordering question is usually the whole argument.
This is the cheapest and most useful one. Search the drug name with "meta-analysis" or "review" and a year later than the scare, and see whether the finding survived. The 2012 statin memory warning is the model case: the warning went on the label, then trial data pooled from 28,000 participants found nothing, and the label change is the part people remember, because it is the part that got covered.
The scare gets the front page and the correction gets a paragraph on page 14, and then both of them sit in the search results with identical-looking authority when you go and check three years later. Google cannot tell which one survived.
Less than you might expect, and none of it on your own.
Do not stop a prescription because of an article, including this one. Every measured harm on this page came from people stopping a drug in response to coverage, and in the two cases where somebody counted properly, the stopping caused more damage than the drug ever did.
Write the headline down, with the drug name and the date, and take it to whoever prescribed it. That is the entire action. It converts a piece of anxiety you cannot resolve at midnight into a two-minute conversation with somebody who knows your other conditions, your other medications, and why you were put on it in the first place.
If you want to look it up first, run the four questions — whether the number is relative or absolute, whether the effect was measured or merely reported, which of the two things came first, and what happened to the finding eighteen months later. That is enough to tell an actual result from a well-optimised sentence.
One caveat I want to leave standing rather than resolve: none of this tells you whether any specific drug is right for you. The four questions sort good evidence from bad evidence. They cannot weigh an 8-in-10,000 risk against a symptom that is wrecking your life, because that trade depends on you, and the research does not contain you.
Our anti-aging guide collects everything that holds up under real human research, with the studies behind each one and no supplement hype.
See the Full GuideOften, and one case was measured. After the October 2013 BMJ statin papers were covered heavily in the UK press, researchers using primary-care records estimated more than 200,000 patients stopped a statin they were already on, with 2,000 to 6,000 avoidable cardiovascular events projected. The 18–20% side-effect figure central to those papers was withdrawn by the authors in 2014. The correction got a fraction of the coverage.
Both headlines you remember were overstated. Kirsch's 2008 analysis found the drugs beat placebo but judged the margin clinically unimportant, which the press turned into "they don't work". Cipriani's 2018 analysis of 522 trials found all 21 drugs beat placebo, which the press turned into "it's official". The underlying numbers are not far apart. Modest average effect, wide variation between people, and neither headline said that.
Nothing on its own, because it is relative to a starting number the headline usually leaves out. In the Women's Health Initiative, combined hormone therapy carried a hazard ratio of 1.26 for invasive breast cancer. The absolute figures were 38 cases per 10,000 women per year against 30 on placebo, so 8 extra cases per 10,000, roughly one woman in 1,250. Whether that trade is worth making is a real conversation. It is a different conversation from the one "26%" starts.
Yes, and those tend to arrive quietly. The Vioxx heart-attack signal was in the published VIGOR data in 2000 and argued in print in 2001; the drug came off the market in September 2004. Patients reported from 2007 that generic Budeprion XL 300 mg was not working like the brand, and were dismissed until FDA testing confirmed it in 2012. Both were under-covered rather than over-covered.
Not on the strength of an article, and not without the person who prescribed it. Every measured harm described on this page came from people stopping medication after coverage. Write the headline and the date down and take it to your prescriber or pharmacist, who knows what else you are taking and why you started.
— Scott Covert, 60, skeptic, not a physician. I have been reading drug scare stories since the mid-eighties, which is long enough to watch several of them come round twice. Think I have read one of these wrong? Tell me and I'll dig in.