By Keith Devlin @KeithDevlin@fediscience.org, @profkeithdevlin.bsky.social

This past month, Tylenol went from being a red-and-white bottle most of us keep in a cupboard for dealing with physical pain—everything from a morning-after headache to post surgery recovery—to being questioned in the US as a potentially dangerous drug (for some people, pregnant women in particular).

True, the current US administration is significantly anti-science, even being skeptical of the vaccines that have controlled, and in some cases totally eliminated, terrible illnesses, that would otherwise have led to millions of deaths (and in the past did so). But questioning this particular drug seemed to come out of the blue.

It’s been around a long time and is close to being available globally. Discovered in the UK in 1889, and given the drug name acetaminophen (APAP), it has been available since 1956. At least in my experience, living in health-conscious California, it’s the number one pain-killer recommended by physicians. In most countries outside the US, it’s sold as Paracetamol.

Some countries do restrict or regulate its sale, mostly for certain formulations or dosages, due to overdose risks. And a few countries (Japan is one of them) ban sales of APAP products because of additional ingredients. 

The episode aroused my mathematician’s interest, because events like this often stem from a misunderstanding of the numerical evidence that all medical treatments amass (for good reason). For example, misunderstanding surely lies behind people fearing vaccination, because of the published statistics on adverse reactions (including deaths).

With any medical treatment, there are two numbers you should consider. The likelihood of sickness or death if you don’t have the treatment, and the likelihood of sickness or death resulting from the treatment. With all of the common vaccines we have available, those two numbers are far apart, usually several orders of magnitude. As a result, I, and many Americans, line up annually for a booster shot for influenza. The percentages State-by-State can be seen here. (That’s based on 2023 data, when we had a pro-science administration, so I assume it’s reliable.)

Similar data for COVID-19 seems less easy to find. A news report in December 2022 (when the pandemic was almost over) stated that over 80% of the US population had received an initial shot. [The WHO declared the pandemic stage of the viral infection was over in May 2023. If you want to remind yourself of what it was like to live through it, Northwestern University has a timeline here.]

Uptake on booster shots has been surprisingly low, given that the illness is worse than flu. How much worse? COVID-19 has a higher mortality rate (0.1%) than flu (0.002%). It is more likely to require hospitalization than flu. As many as 20% of people infected with COVID-19 may require hospitalization, compared to around 1% of people with flu. And COVID-19 can cause a number of long-term complications, whereas flu typically does not cause long-term complications. That’s pretty compelling data.

Again (and this is the main focus of this month’s post) the numbers provide a reliable basis for making your own decision when it comes to preventative medicine. (Financial capacity is another numerical factor to consider in the US.)

One group that does have a relatively higher rate of COVID-19 boosters are the over-75. As of December 14, 2024, about 49% of adults 75 years and older had received the latest COVID-19 vaccine. (See here. Again, this is pre-2025 CDC data, so can be viewed as reliable.) This disparity surely reflects widespread knowledge of the statistical data regarding the dangers of COVID-19 for this age-group—which I am in. (Yes, I got my flu and COVID-19 boosters last month.)

Getting back to Tylenol, when I started to look for scientific data on which the government was (presumably) basing its recent decision, my expectation was that I would find that it was a result of reading causation into correlation—a very common mistake. For example, on its own, a study that showed more women who took APAP had children with autism than those who did not, comes nowhere close to showing the drug is a cause of autism. Indeed, a more likely explanation is that women who have difficult pregnancies are more likely to reach for the Tylenol bottle, and it was whatever was causing their difficulties that resulted in neurological changes in their fetus. In other words, both the Tylenol use and the neurological changes were the result of that factor.

To my initial surprise, what I discovered was (mathematically) more  interesting, as I’ll explain.

But before I start, I should say that mine was by no means an exhaustive examination. On the contrary, the overall point I want to make is that a fairly basic ability to gather (and assess the reliability of) information, and get enough data to make a sensible choice, is a critical skill for life in today’s data-driven world.

My search tool was Google, and I deliberately did not pursue threads more than one deep. My goal was to see what I could find in a short space of time—a couple of hours in this case—knowing virtually nothing about autism or pharmaceuticals.

The first thing I wanted to learn: is there indeed a scientific question about possible dangers of APAP regarding the growth of embryos? Having grown up during the Thalidomide tragedy of 1957-61, where a drug believed safe for pregnant women turned out to cause horrendous birth defects, I would think any mother-to-be should be cautious about anything they put into their body.

Thalidomide was a new (over-the-counter) drug, introduced in Germany to prevent sickness, anxiety, and difficulty in sleeping. It had not been designed for use during pregnancy, but physicians in Europe quickly took to prescribing it for them, believing it to be safe. It was anything but. In the US, the FDA refused to certify the drug. The courageous pharmacologist who persuaded them to do so, Dr Frances Oldham Kelsey, was subsequently honored by President Kennedy.

That kind of danger is definitely not the case with APAP. It’s been around a long time and been used by millions of people around the world.

Still, I googled for “scholarly articles on acetaminophen and autism”. I used those keywords because the government’s announcement mentioned autism specifically. 

The first paper that came up was titled  Acetaminophen’s Role in Autism and ADHD: A Mitochondrial Perspective.

Having no expertise in the domain, I relied on meta-information to assess the paper’s scientific value. Being listed on the National Library of Medicine website told me nothing; as the site says, “inclusion does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health”.

So I looked at the journal in which it was published: International Journal of Molecular Sciences. Knowing nothing about that, I searched to find out about the publisher, MDPI. It seems they charge authors to publish their papers. That’s always a red flag. One academic in the Netherlands who used to review for them wrote (on researchgate.net) a scathing review. Like him, I no longer respond to the many requests for reviews I receive from pay-to-publish journals (mostly in mathematics, but sometimes in fields about which I know nothing). I am sure those journals receive many good papers, particularly from younger researchers struggling to establish themselves in their field. But for my purposes (writing this article), this paper was a nonstarter. It may be perfectly good science; but I wouldn’t know. On the other hand, the paper starts off by complaining about a judge’s ruling in a lawsuit, and that rang alarm bells. This looked like a propaganda piece.

Still, the paper’s literature review started off with reference to a 2021 paper published by Nature, titled Paracetamol use during pregnancy — a call for precautionary action. The association with Nature gave me more faith in this article, so I checked it out. Apparently, there was an outstanding question about the safety of APAP during pregnancy. The authors state that

“… increasing experimental and epidemiological research suggests that prenatal exposure to APAP might alter fetal development, which could increase the risks of some neurodevelopmental, reproductive and urogenital disorders.”

That’s a call for further research, not government intervention. And there’s no mention of autism specifically. As the paper goes on to say,

“Here we summarize this evidence and call for precautionary action through a focused research effort and by increasing awareness among health professionals and pregnant women. APAP is an important medication and alternatives for treatment of high fever and severe pain are limited. We recommend that pregnant women should be cautioned at the beginning of pregnancy to: forego APAP unless its use is medically indicated; consult with a physician or pharmacist if they are uncertain whether use is indicated and before using on a long-term basis; and minimize exposure by using the lowest effective dose for the shortest possible time. We suggest specific actions to implement these recommendations.” 

That paper was published back in 2021. With publication in Nature, it would have been widely read, and for sure others had pursued such a study. I dug around a bit more and fairly quickly came up with what I was looking for: a more recent, large-scale study focusing on autism in particular, carried out by a reputable group of scientists in Sweden, and published in a prestigious, peer-reviewed journal.

The paper’s title is Acetaminophen Use During Pregnancy and Children’s Risk of Autism, ADHD, and Intellectual Disability. It was published by JAMA, a collection of peer-reviewed, monthly open-access medical journals, published by the American Medical Association. Written by five Swedish researchers and two Americans, the paper reported on a rigorous 2024 study of 2.48 million Swedish children. The authors describe the study as a

“nationwide cohort study with sibling control analysis [that] included a population-based sample of 2,480 ,797 children born in 1995 to 2019 in Sweden, with follow-up through December 31, 2021.”

[That “sibling control” factor was important, since siblings often have features in common, and that could confound the data from a study. See presently.]

They were not looking for evidence of causality; that’s invariably extremely hard to establish. (Correlation between smoking and lung cancer was known for many years before causation was finally established.) Rather, they looked for an actual correlation between

“use of acetaminophen during pregnancy prospectively recorded from antenatal and prescription records”

and

“autism, ADHD, and intellectual disability based on ‘International Classification of Diseases, Ninth Revision and International Classification of Diseases, Tenth Revision’ codes in health registers.”

Their conclusion:

“Acetaminophen use during pregnancy was not associated with children’s risk of autism, ADHD, or intellectual disability in sibling control analysis. This suggests that associations observed in other models may have been attributable to familial confounding.”

In other words, studies that reported correlation made the error of not controlling for siblings. And that’s just for correlation. Forget about causation; that hypothesis never gets off the ground.

It’s all in the math. It’s one of the best tools we have to make good decisions.

Anyway, the overall point I want to make is that, if we educated our future citizens with the very basic reasoning skills I used to write the above, we would all be in much better shape. Though I do have expert understanding of some of the math involved, I know nothing of the biology. But in the Age of Google, you don’t need any of that domain expertise; just basic reasoning, together with knowing how to use a search engine and vet and judge articles (by experts) on which you rely. Those are basic human reasoning skills that just require a bit of teasing out and developing.

FOOTNOTE: AI as it currently stands will leave you no better than when you start on a quest like mine, since it bases its answers on everything written on the subject that it can access. Here is what I got on OpenAI GTP-4:

Q: Is there an established correlation between acetaminophen use during pregnancy and having children with autism?

A: Research on the potential link between acetaminophen use during pregnancy and the risk of autism spectrum disorder (ASD) in children has produced mixed results. Some observational studies have suggested a possible association, indicating that prenatal exposure to acetaminophen may be linked to a higher risk of ASD, while other studies have found no significant correlation.

It's important to note that these studies often face limitations, such as confounding factors, recall bias, and difficulty establishing causality. Currently, most health authorities, including the CDC and WHO, recommend that pregnant women use acetaminophen only as directed and discuss any medication use with their healthcare provider.

If you're pregnant or planning to become pregnant, it’s best to consult your healthcare provider for personalized advice regarding medication use.

That answer is not wrong. I found all those points out in my brief search. But the AI (a LLM) doesn’t have the ability to make the kinds of social-based, common sense assessments I was using. In particular, it cannot identify the study that finally provided the answer. Sure, a conversation with the AI would likely get you there, but to make that happen you would have to lead it through the narrowing down process that I went through. And in that case, you have your answer anyway.