In April 2025 the Secretary of Health and Human Services promised to know what caused autism by September. In September the administration named acetaminophen, and the President told pregnant women eighteen separate times not to take Tylenol, ending at "fight like hell not to take it." The notice his FDA sent doctors that same day told them to consider minimizing it for routine low-grade fevers, said a causal relationship "has not been established," and called it the safest over-the-counter painkiller in pregnancy. This page follows all four claims to what became of them, and sets out what a woman who takes that advice is actually risking.
The short version
The promise was to identify the cause of autism by September 2025; what arrived was an association already in the literature that the FDA's own notice described as not established as causal. Since then it has failed to survive in all three national cohorts that compare siblings to each other, including one in Hong Kong that recorded 43 percent of its pregnancies as exposed and still found nothing. The promised autism warning never reached the label, and in March 2026 the same FDA refused to approve leucovorin for autism. The part that did not stay at the podium was the instruction: the FDA notice told clinicians to consider minimizing acetaminophen for routine low-grade fevers, while the President told patients not to take it at all, for the whole pregnancy. Acetaminophen orders for pregnant emergency patients then fell about 10 percent. It is the only drug available over the counter that safely brings a fever down in pregnancy, and hyperthermia in early pregnancy carries a pooled odds ratio of 1.92 for neural tube defects, a larger and better-evidenced association than the one being avoided.
In April 2025, at a cabinet meeting, the Secretary of Health and Human Services made a specific, dated commitment.
"By September we will know what has caused the autism epidemic and we'll be able to eliminate those exposures."
Robert F. Kennedy Jr., April 2025. He later softened it to expecting "to begin to have answers."
Five months later, on September 22, the answer arrived on schedule. In the Roosevelt Room, with the FDA Commissioner, the NIH Director and the CMS Administrator standing behind him, the President named a culprit: acetaminophen, taken during pregnancy, carrying what he called "a very increased risk of autism." The FDA would notify physicians. A label change would begin. And pregnant women were told, repeatedly and in plain language, to stop taking it.
This page follows that sequence to what became of it. A deadline was set. A cause was named. An instruction was given. Each of those three is checkable, and this is what the record shows about each.
The announcement was made from the Roosevelt Room on September 22, 2025, with the Secretary of Health and Human Services, the FDA Commissioner, the NIH Director and the CMS Administrator present. The full transcript is public. Below is every passage in it where the President tells pregnant women to avoid Tylenol or to take less of it, in the order he said them, with the caveats he attached where he attached them. Eighteen in all.
Two of the eighteen contain the FDA's actual wording, at 04:33 and 05:34: limit use unless medically necessary, and talk to a doctor. In both he wraps it in his own harder verdict, so neither is a clean relay. Six carry a caveat of some kind, from the fever exception to "unless it's absolutely necessary" to "that's up to you and your doctor." The rest are the bare instruction.
The passage at 36:31 is the one that matters most, and not because it is the harshest. It is where he says the recommendation is "maybe stronger for me than from the group because they're waiting for certain studies." The group is the officials standing behind him, all of them heads of federal health agencies. He knew he was ahead of them, and said so, and then gave the harder version anyway.
Every quotation is from the Roll Call Factba.se transcript of the September 22, 2025 remarks, checked as an exact string against it. Each row is one timestamped segment, which is how the archive divides the remarks; consecutive segments can belong to a single unbroken stretch of speech, so this counts moments he returned to the subject rather than separate occasions. Ellipses mark omitted words inside a segment. Not included, because they are not advice about acetaminophen in pregnancy: his remarks about vaccine schedules, and his remarks about other countries not using Tylenol, which he made at 11:07 and again in the Q and A.
The press conference was announcing a real regulatory action, and that action exists as a document. The FDA issued a notice to physicians the same day, signed by Commissioner Martin Makary. It is the administration's own written statement of what it believed the evidence showed. It does not say what the podium said, and the most important difference is not a matter of degree. It is a matter of audience.
"So ideally, you don't take it at all." / "don't take Tylenol. Don't take it. If you just can't -- I mean it's fight like hell not to take it."
Clinicians should "consider minimizing" acetaminophen in pregnancy for "routine low-grade fevers."
"To be clear, while an association between acetaminophen and autism has been described in many studies, a causal relationship has not been established and there are contrary studies in the scientific literature."
The consideration "should also be balanced with the fact that acetaminophen is the safest over-the-counter alternative in pregnancy among all analgesics and antipyretics."
A notice to physicians is an instruction to a professional who has the patient in front of her. She knows the temperature, the gestational age, the infection, the history. "Consider minimizing for routine low-grade fevers" is a sentence she can apply, because she is the one who can tell a routine low-grade fever from a dangerous one, and because the same notice reminds her that this drug is still the safest option she has.
None of that transfers. A pregnant woman at home with a thermometer is not in a position to make that judgement, and she was not asked to. She was told not to take it, and to fight like hell not to. Strip the clinical judgement out of clinical guidance and what is left is not a weaker version of the same advice. It is different advice, aimed at someone who cannot act on the part that made it safe.
Two further gaps follow from that one. The notice covers routine low-grade fevers; the speech covered the entire pregnancy. And the notice states plainly that causation has not been established, while the speech described "a very increased risk of autism." Audience, scope, certainty. The document he was announcing differed from him on all three.
Ask the desk "Can a president overrule what the FDA tells doctors?"
The rest of this page turns on figures like 0.98 and 1.92. They are ratios, and the scale is not the intuitive one. It does not run from zero to a hundred, and zero is not the bottom of it. The number that means "no difference" is 1.00.
A ratio compares two groups: how often something happened in the exposed group, divided by how often it happened in the unexposed one. Divide a number by itself and you get 1. So 1.00 means the two groups came out the same, and the exposure made no difference at all. Above 1.00 the exposed group had more of the outcome. Below 1.00 they had less.
Tylenol and autism
compared within families, three cohorts
Tylenol and autism
compared between unrelated families
Fever and neural tube defects
pooled across 15 studies
Tylenol measured in cord blood, and ADHD
highest exposure third, 996 births
Every ratio comes with a range attached, the confidence interval, which says how much the estimate would wobble if the study were repeated. The rule for reading it is simple and it does a lot of work on this page: if that range includes 1.00, the study has not shown a difference. The Danish sibling figure of 1.09 sounds like a nine percent increase, but its range crosses 1.00, so it is a result consistent with no effect at all. That is what "not statistically significant" means, and it is why the page keeps saying it.
The last thing these numbers do not tell you is how likely anything is in absolute terms, and that gap is where most health scares live. Neural tube defects occur in roughly 6.5 of every 10,000 births in the United States. Nearly doubling a risk that small still leaves it small: it is the difference between about 6 or 7 in 10,000 and about 12 or 13. Both the risk this page says was accepted and the risk it says was avoided are, for any individual pregnancy, unlikely events. What separates them is not that one is common. It is that one of them survives when you control for the family, and the other does not.
Ask the desk "Why do small relative risks so often turn out to be confounding?"
"Tough it out" has a physiology. Acetaminophen is not primarily a comfort drug in pregnancy; it is the drug used to bring a temperature down, and there is no other one available over the counter that does the job safely. So the instruction to endure a fever rather than treat it is a medical instruction, and it lands on a body doing something specific.
Heat is a teratogen. A raised core temperature interferes with the folding of the neural tube, the structure that becomes the brain and spinal cord. Laboratory work has traced part of the mechanism: heat activates a family of ion channels in the neural crest cells that build the face and heart, and stimulating those channels artificially reproduces fever-pattern defects in animal embryos. The developmental window that matters most is early. The neural tube closes in the third and fourth weeks after conception, which for many women is before they know they are pregnant.
1.92
The pooled odds ratio for neural tube defects in the offspring of women who had hyperthermia in early pregnancy, across 15 studies and roughly 39,600 participants. A 2024 umbrella review grades this "highly suggestive" evidence, its second-highest tier. Moretti et al., Epidemiology, 2005
Neural tube defects are anencephaly and spina bifida. Anencephaly means a skull and brain that do not form; it is not survivable. Spina bifida means a spinal cord left exposed, and its consequences run from a repairable lesion to lifelong paralysis, hydrocephalus and incontinence. These are the outcomes on the far end of an untreated fever in the wrong week.
The rest of the picture is consistent. A separate systematic review of fever in pregnancy found roughly a doubling of oral cleft risk and about a 54 percent increase in congenital heart defects. Influenza in pregnancy, which produces fever among other things, carries an odds ratio above 3 for neural tube defects. Maternal infection that progresses to sepsis is associated with sharply higher odds of preterm birth and low birth weight, and untreated infection is among the documented routes to stillbirth. ACOG's own advisory, written for this exact controversy, states that fever in pregnancy is "associated with increased risk of neural tube defects and other birth defects such as oral clefts and cardiac defects," and that inadequately treated pain "can destabilize maternal physiology, with potential downstream effects on fetal well-being."
And the alternative is not a different pill. The FDA recommended in October 2020 that pregnant women avoid ibuprofen and other NSAIDs at about 20 weeks or later, because of fetal kidney injury and dangerously low amniotic fluid. Aspirin carries its own restrictions. That is why the FDA's own notice called acetaminophen the safest over-the-counter option among painkillers and fever reducers: not as a compliment to the drug, but because the shelf is otherwise empty.
Ask the desk "How high does a fever have to be in pregnancy to be dangerous?"
What happened next can be measured, and it was. Researchers led by an emergency physician at Brigham and Women's Hospital looked at acetaminophen ordering for patients in emergency departments before and after September 22, 2025.
About 10%
The fall in acetaminophen orders across 88,857 emergency department visits by pregnant patients in the months after the announcement, reaching as much as 20 percent in the third week. Across 853,216 comparable visits by patients who were not pregnant, there was no statistically significant change. Faust et al., The Lancet, March 5, 2026
The comparison group is what makes the first number mean anything. A drop specific to pregnant patients, with no significant movement among everyone else in the same departments over the same months, rules out a general shift in how these departments were prescribing.
What it does not do is tell you which message caused it, and that limit matters on this page more than most. The thing being measured is what clinicians ordered, and clinicians received two messages that day: the press conference, and the FDA notice telling them to consider minimizing acetaminophen for routine low-grade fevers. The study cannot separate them, and neither can this page. What can be said is narrower and still substantial: after September 22, pregnant patients in emergency departments got less of the drug that the same day's FDA notice called the safest over-the-counter option among painkillers and fever reducers.
What a pregnant woman does instead is the subject of the section above, and the short answer is that there is no good substitute. That is the exchange the drop represents.
Studies of acetaminophen in pregnancy really do find a link. That is not a misreading, and anyone who says the association was invented is wrong. The argument is about what the association means, and it turns on a single question: what are you comparing these mothers to?
Autism is among the most heritable of common conditions. Twin and family studies put the genetic contribution somewhere around 80 percent. Parents differ in their genes, and autistic traits run in families. They also differ in why they reached for a painkiller: infection, migraine, chronic pain, inflammation, and the willingness to medicate at all. Every one of those things can travel with both the pill and the diagnosis without the pill causing the diagnosis.
Compare a mother who took acetaminophen with an unrelated mother who did not, and all of that comes along for the ride. Compare her with her own sister's pregnancy, or better, with her own other pregnancy, and the shared genetics and the shared household hold still. Whatever difference remains is closer to the thing you actually wanted to measure.
This is why two systematic reviews published five months apart, drawing on almost the same body of literature, reached opposite conclusions. A review in Environmental Health in August 2025 examined 46 studies and concluded the evidence was consistent with an association. A review in The Lancet Obstetrics, Gynaecology & Women's Health in January 2026 examined 43 studies and found no association. The first discounted sibling-controlled studies. The second prioritised them. Almost everything else follows from that one methodological choice.
Answering this question needs a health system that links medication records to births to diagnoses across a whole population, and does it for long enough that siblings can be compared. Three have now done it. They were run by different teams, in different countries, under different health systems, and they found the same thing.
0.98, 0.98, 1.01
Sibling-matched hazard ratios for autism, ADHD and intellectual disability in a Swedish cohort of 2,480,797 children born 1995 to 2019. A ratio of 1.00 means no difference. Without sibling controls, the same data showed marginally increased risks. Ahlqvist et al., JAMA, April 2024
1.03 and 1.09
Population and sibling-matched hazard ratios for autism in a Danish cohort of 1,506,155 children born 1997 to 2022. Neither reached statistical significance, and the study found no dose-response pattern and no effect specific to any trimester. Prahm et al., JAMA Pediatrics, April 13, 2026
1.00 and 1.01
Sibling-matched hazard ratios for autism (95% CI 0.91 to 1.11) and ADHD (95% CI 0.93 to 1.08) in 708,020 Hong Kong mother and child pairs covering births from 2001 to 2023. Results held across dose, trimester and pattern of use. Before sibling matching, this study too reproduced the slight increase other work had reported. Wan, Tanuseputro et al., JAMA Internal Medicine, June 29, 2026
Three times, the same sequence. The association is there in the data. Then the comparison group changes, and it is not.
The Hong Kong study matters most, and not because it is the biggest, because it is not. It matters because of who it managed to count. Roughly 43.3 percent of those pregnancies had recorded paracetamol exposure, close to the share of pregnancies believed to use it worldwide. The most serious objection to the Nordic results is that they recorded far too few users to detect anything. That objection cannot explain a null result in a population where nearly half the pregnancies were captured as exposed.
These three are not the whole sibling literature, and it would be an overclaim to say so. An earlier Norwegian analysis (Brandlistuen 2013, 134 discordant pairs at 28 days or more of use) did find significant within-sibling associations, though for motor and communication delays at age three rather than autism; a later one on ADHD (Gustavson 2021, 34 discordant pairs) did not. The honest statement is narrower and still substantial: every national-scale sibling-controlled cohort assembled to date, across Sweden, Denmark and Hong Kong, has found nothing.
Ask the desk "What would it take to prove a prenatal exposure causes autism?"
Two things were promised in the Roosevelt Room. Neither has arrived, and in one case this administration's own FDA decided against it outright. The label change is unfinished rather than withdrawn, which is not the same as reversed.
Ask the desk "What happens to a label change once FDA starts and stops it?"
Three Republican senators who practise medicine responded within a day. All three had voted to confirm the Secretary standing beside him.
Marshall's version is the defensible one, and it sits a long way from "don't take it at all."
There is a serious version of this position, held by credentialed researchers, and it does not depend on anything said in the Roosevelt Room. It runs like this.
Ask the desk "How do regulators in Europe handle paracetamol in pregnancy?"
It does not claim acetaminophen is proven harmless at every dose and duration. Long-term heavy use in pregnancy is the scenario with the least good data, and it is the scenario Sen. Marshall (R) pointed at.
That gap deserves a sharper statement than it usually gets, because it is the strongest question a sceptical reader can ask here. Most of the literature reporting an association locates it in prolonged use, at thresholds around four weeks. All three sibling cohorts report their null results holding across dose, timing and pattern of use. But none of them publishes the number this question actually turns on: how many sibling pairs were discordant on a month or more of use. This page prints that figure for the two Norwegian analyses, 134 pairs and 34 pairs, because they published it. It cannot print it for Sweden, Denmark or Hong Kong, because they did not, and it should not be inferred. So the fair summary is that the null results cover prolonged use by the studies' own account, and that the precision behind that coverage is not public.
It does not claim the association was invented. It appears in dozens of studies and the researchers who report it are not cranks.
It does not claim the Second Circuit vindicated the science, or rejected it. That ruling was about admissibility.
It does not claim the rise in autism diagnoses is entirely explained by broader criteria and better ascertainment. Those account for a substantial share; whether they account for all of it is unsettled.
It does not predict what the FDA will finally do with the label, and it does not tell anyone what to take. That is a conversation between a patient and their clinician, which is roughly what two of the three Republican physicians quoted above said.
Set the September announcement against the April promise, and take each part in turn.
They said they would find the cause of autism by September. They did not. What they produced was an association already present in the literature, which the FDA's own notice described as not established as causal, and which has since failed to survive in every national cohort that compares siblings to each other.
They said they had identified the culprit. The evidence has moved the other way since. The Danish cohort in April, the Hong Kong cohort in June, and a systematic review in January all landed on the same answer, and the Hong Kong study answered the best objection to the earlier ones by capturing 43 percent of its pregnancies as exposed and still finding nothing.
They said the label would carry a warning, and that a treatment was at hand. Neither arrived. The label was updated in May 2026 without an autism warning. Leucovorin was refused for autism by this administration's own FDA in March.
And they told pregnant women to stop taking it. That is the part that did not stay at the podium. Acetaminophen orders for pregnant emergency patients fell about 10 percent, with no significant change among comparable patients who were not pregnant. Part of that is the FDA notice working as intended, and this page cannot separate the two messages. What the notice never asked for, and what has no support in the label or in any of the three cohorts, is the thing that was actually said out loud: don't take it at all, for the whole pregnancy, fight like hell not to.
The cost of that instruction is not a hypothetical about autism. It is a fever in the third week after conception, in a woman who does not yet know she is pregnant, going untreated because the President of the United States told her on television that there was no downside. The neural tube closes that week. Sometimes nothing happens. Sometimes what happens is anencephaly.
No one has yet documented a specific child harmed this way, and this page does not claim otherwise. Risk imposed is not the same as injury proven. But the imposition was real, it was measurable in emergency departments within weeks, and it was made in exchange for protection against a cause that the government's own letter said had not been established, and that the evidence since has not found.
Sen. Bill Cassidy (R-LA), a physician, named the other cost the same week. He said he did not want mothers "to sit around and blame themselves that if they took Tylenol and they have an autistic child, that they are to blame."
The link was never proven. The fever always was.
Quotations from the September 22, 2025 remarks are taken from the Roll Call Factba.se transcript and reproduced exactly, including the speaker's self-corrections. Study figures come from the journals themselves. Where a claim rests on a single outlet rather than a primary document, it is labeled on the page. Two figures could not be verified to a primary document and are therefore stated as the journals' published summaries describe them: the confidence intervals for the Danish cohort, which sits behind a paywall, and the exact publication date of the ACOG advisory beyond September 2025.
This page sets the September 22, 2025 White House remarks beside the FDA notice issued the same day, the three national-scale sibling-controlled cohort studies of prenatal acetaminophen exposure, the two systematic reviews that reach opposite conclusions from them, and the measured change in emergency department care that followed. It concedes that the association appears in dozens of studies, that sibling-comparison designs have documented weaknesses that push results toward the null, that exposure was measured poorly in both large null studies, and that credentialed researchers disagree in good faith. The findings are narrower: causation has not been established, the FDA's own notice said so, and the advice given from the podium went further than that notice, further than the guidance it was announcing, and further than the evidence supports. Litigation, regulatory decisions and new studies can change this picture; the label decision and the remanded cases were both pending at publication. Corrections welcome.