Six weeks ago I wrote about the texts people on GLP-1 drugs were getting from worried family members after a wave of vision-loss headlines. At that point the United States hadn’t acted. Europe had labeled NAION a very rare side effect of semaglutide in June 2025, the UK followed in February, and in July Australia updated every GLP-1 label, with a formal warning for semaglutide and a listed side effect for the rest.^1-3

On Monday, November 9, FDA acted. Every GLP-1 receptor agonist sold in the United States must add information about NAION (non-arteritic anterior ischemic optic neuropathy) to the Warnings and Precautions section of its label.⁴ In plain terms, NAION is a sudden drop in blood flow to the optic nerve, sometimes called an eye stroke, and the vision it takes often doesn’t come back. The order covers all 12 products on FDA’s list, from exenatide (now sold only as generics) and liraglutide through Ozempic, Wegovy, Mounjaro, and Zepbound to the 2 insulin combinations and the new oral orforglipron (Foundayo). Europe’s action covered semaglutide alone, and Australia’s was graded by drug. FDA’s puts NAION in Warnings and Precautions for the whole class.

Expect the texts to start again, and this time some will come from colleagues. Before answering, it’s worth knowing where FDA’s own numbers land, because it isn’t where the headlines do. I’ve written this one for patients as much as clinicians: if you take one of these drugs, the box below and the patient list further down are yours, and the statistics come with a plain-language translation.

The conclusion is one sentence, and its qualifiers carry the meaning: there is “a potential increased risk of rare NAION events in patients taking GLP-1 RAs.”⁴ The warning goes into Warnings and Precautions rather than a boxed warning, FDA added no contraindication, and it found NAION cases reported for every GLP-1 product on the market.

It weighed several kinds of evidence, unequally. It called its adverse-event reports hard to interpret, because NAION’s risk factors overlap with the conditions these drugs treat, and it noted the limitations of the 2024 study that started the inquiry. Rigorous observational studies in type 2 diabetes, it said, consistently found higher risk with GLP-1 drugs than with 2 other families of diabetes drugs, DPP-4 inhibitors (such as Januvia) and SGLT2 inhibitors (such as Jardiance or Farxiga). It also ran its own Sentinel analysis and pooled 126 randomized trials. In animal studies it found no mechanism.⁴

That last finding is worth sitting with. The best-known warning on these labels, the thyroid C-cell boxed warning, rests on rodent tumors that 16 years of human data haven’t confirmed, as I wrote in September. The newest rests on human data with no animal mechanism behind it. And in January the same agency took a warning off, asking that suicidal ideation and behavior language come off Wegovy, Zepbound, and Saxenda after a review of 91 placebo-controlled trials found no increased risk.⁵ A label that moves in both directions is following evidence. None of that tells you how often a harm happens, though. For that you need the numbers.

The study that started this came from a single neuro-ophthalmology service at Massachusetts Eye and Ear, which searched the records of 16,827 patients. Among those with type 2 diabetes, 17 of 194 semaglutide users developed NAION, against 6 of 516 on other diabetes drugs, a hazard ratio of 4.28, which means about 4 times the rate. Among patients with overweight or obesity it was 7.64, on 20 cases against 3 in the comparison group.⁶ FDA has said that paper is what made further assessment warranted.^4,7

Those figures were never going to describe a typical patient. Everyone in that registry had already been sent to a neuro-ophthalmologist, which helps explain why 8.9% of the semaglutide users with diabetes developed NAION within 3 years. In FDA’s Sentinel data, the rate in people with diabetes ran 26 to 32 per 100,000 person-years (a person-year is 1 person followed for 1 year), roughly 0.1% over 3 years.⁴ Small counts also made the estimate imprecise: the 4.28 had a confidence interval from 1.62 to 11.29, so the data fit anything from a modest increase to an 11-fold one.⁶

Larger datasets followed. Danish and Norwegian national registries, in a study published 3 months before Europe’s decision, compared semaglutide with SGLT2 inhibitors in type 2 diabetes and found a hazard ratio of 2.81, about 1.4 extra cases per 10,000 person-years.⁸ FDA’s Sentinel analysis followed adults with type 2 diabetes starting a GLP-1 drug, an SGLT2 inhibitor, or a DPP-4 inhibitor from 2019 to 2024. Against DPP-4 inhibitors the hazard ratio was 1.43 (95% CI 1.02–2.01). Against SGLT2 inhibitors it was 1.19 (0.91–1.54), which isn’t statistically significant: chance alone could explain it.⁴

The comparator matters, and the results don’t line up neatly. FDA’s signal held against DPP-4 inhibitors and faded against SGLT2 inhibitors; the Scandinavian registries found a strong one against SGLT2 inhibitors; and an August analysis of 98,082 matched pairs who started a GLP-1 drug or a DPP-4 inhibitor between 2006 and 2021 found no significant difference (odds ratio 1.11; 95% CI 0.96–1.29).⁹ Reviewing more than 40 studies, the American Academy of Ophthalmology (AAO) and the North American Neuro-Ophthalmology Society (NANOS) report that pooled meta-analyses put the semaglutide association at about 2-fold.¹⁰ The direction mostly points one way, and the biggest analyses have mostly landed well below the number that started the story.

The randomized evidence is the part of FDA’s review I’d most want clinicians to see, because randomization is the only protection against the confounding that dogs everything else here. FDA pooled 126 trials: 134,253 participants, 288,879 person-years. It found 15 probable cases of NAION: 10 among 78,217 people taking a GLP-1 drug and 5 among 56,036 in comparison groups, all 5 of them on placebo.⁴

The difference works out to about 1 extra case for every 38,000 years of treatment. Put another way, 38,000 people would have to take the drug for a year to produce 1 extra case. FDA didn’t publish an interval for that comparison, so I ran a crude one from its counts: a rate ratio of about 1.7, with a 95% confidence interval from roughly 0.5 to 6. Fifteen events can’t tell no effect from a large one. What they can tell you is that the event is uncommon. Trials weren’t built to look for NAION, though, so these rates are more likely a floor than a ceiling; in FDA’s Sentinel study of people with diabetes, the rate ran several times higher in every drug group.⁴

For scale, NAION strikes 2.3 to 10.2 of every 100,000 adults over 50 each year in the general population.⁴ Europe’s regulator estimated about 1 additional case per 10,000 person-years of semaglutide,¹ and the Scandinavian registries found 1.4 per 10,000 person-years.⁸ Put the trial difference beside those and the point estimates run from about 0.3 to 1.4 extra cases for every 10,000 people treated for a year, though the trial figure’s interval includes zero and the registry’s reaches about 2.3.

You may see a much bigger number this week: 68.6 times the risk, from a 2025 meeting abstract. It came from the World Health Organization’s database of voluntary adverse-event reports,¹¹ which measures how often a side effect is reported, not how often it happens, and after 2 years of headlines about this exact side effect, reporting is the last thing to trust. FDA called its own version of those reports hard to interpret.⁴

Most of what we know is about semaglutide. Europe limited its action to that drug, and Australia’s advisers concluded that the evidence may support a signal for semaglutide but not for dulaglutide or tirzepatide.³ The AAO and NANOS consensus calls the evidence for other GLP-1 drugs, tirzepatide included, insufficient, while cautioning clinicians not to assume the concern stops at semaglutide.¹² FDA, like Australia, chose the class, though without Australia’s tiering. With cases reported for every product and no mechanism to say why one molecule would differ, I think that’s the defensible default.

Two consequences follow. Switching from semaglutide to tirzepatide to dodge the warning has no reliable evidence behind it: tirzepatide’s weaker signal so far reflects thinner data, not demonstrated safety, and the U.S. labels will no longer distinguish them. And compounded or “research-use” versions don’t carry the warning, because they don’t carry an FDA label at all. That doesn’t make them safer; it means no label will tell the person injecting them, the same gap I wrote about in October when the subject was recalls.

This is the gap I’d most like FDA to close. Its Sentinel analysis and the observational comparisons it leaned on were all in type 2 diabetes, and it reported no separate analysis for obesity alone,⁴ though the warning will sit on every GLP-1 drug approved for weight management.

What exists points two ways. The single-clinic study reported its largest hazard ratio, 7.64, in overweight or obesity, a figure that rests on 23 cases in all.⁶ A 2026 meta-analysis of 8 observational cohorts found about 1.9 times the risk with semaglutide overall, but in the overweight and obesity subgroup the estimate was 1.68, with an interval (0.72 to 3.91) that includes no effect.¹³

There’s also an arithmetic reason to expect a smaller absolute risk here: diabetes is itself a NAION risk factor, so people without it start lower, and the same relative risk on a lower baseline yields fewer extra cases. That’s a reasoned guess, not a measurement, and the measurement is what we need.

The risk factor that dwarfs the drug signal has nothing to do with drugs. It’s the optic disc, where the nerve leaves the eye. Some people are born with a small, tightly packed one, which eye doctors call crowded, or a disc at risk. In a study published in August, 16 neuro-ophthalmology centers reviewed records to compare the cup-to-disc ratio in the unaffected eye of patients with NAION (1,812 eyes from 2,169 consecutive patients) with 465 controls.¹⁴

The authors note this can’t tell us whether the drugs add independent risk.¹⁴ It does tell us that people who develop NAION on a GLP-1 look, anatomically, like people who develop it anyway.

That is where the eye-exam question gets interesting, because the professions disagree. The American Optometric Association recommends a comprehensive eye exam within 12 months before starting a GLP-1 drug or within a month after.¹⁵ The AAO and NANOS don’t recommend routine exams solely to assess NAION risk, saying current evidence doesn’t support it; people with diabetes should keep their regular diabetic eye exams, and anyone with prior NAION should discuss the risk to the other eye before starting.¹⁰ FDA doesn’t mention baseline exams.⁴

I side with the neuro-ophthalmologists for most patients, with one adjustment. A crowded disc is common and NAION is rare, and I wouldn’t withhold a drug with cardiovascular outcome data over disc anatomy alone. But a history costs nothing. In a virtual practice like mine, nobody looks at the optic nerve before the first dose, so the screening happens in the questions we ask, and if the answers raise a flag, I’d want an eye exam before that dose.

Here is what I’d change in a prescribing workflow this week. Patients, your list follows it.

When a patient does develop NAION on one of these drugs, the official advice diverges.

I land between FDA and the societies. For someone taking a GLP-1 drug mainly for weight, with other treatments available, I’d lean toward stopping after a confirmed NAION. For someone with established cardiovascular disease whose drug is doing outcome-level work, I’d want neuro-ophthalmology and cardiology in the conversation before anyone stops anything.

One more data point belongs there, with a caveat. In an analysis presented at this fall’s AAO meeting, among patients who developed NAION in both eyes, 56.8% of second-eye events in GLP-1 users came within 6 months of the first, against 40.6% in nonusers.¹⁶ It’s a meeting abstract, not a peer-reviewed paper, and the GLP-1 users’ HbA1c ran 1.0 to 1.4 points higher, so worse diabetes may explain some or all of it.

The case for the warning is real. The direction holds across most large observational studies against active comparators, the trial pool points the same way if faintly, every product has reported cases, and regulators in Europe, the UK, Australia, and now the United States have reached versions of the same conclusion.^1-4

So is the case against causation. No animal mechanism has been found, and the randomized signal rests on 15 events. FDA’s own comparison against SGLT2 inhibitors wasn’t significant, and a large comparison against DPP-4 inhibitors found no significant difference.^4,9 People who get GLP-1 drugs carry the same risk factors as people who get NAION, from diabetes and sleep apnea to hypertension, and the diagnosis codes these studies depend on are imperfect.^4,12 A 2026 review of 7 serious adverse events attributed to these drugs put NAION among those with uncertain causality, alongside thyroid cancer and pancreatitis, and made the point that matters most: when an event is this rare, multiplying its risk still adds few cases across a broad population, though the same multiplier counts for more in people who start at high risk.¹⁷

My read. A small causal effect is plausible, and a warning is exactly the right response to a harm that is plausible, rare, and often permanent. NAION has no proven treatment. A side effect you can reverse can wait for certainty, but one you can’t should be disclosed while the evidence is still arguing with itself. If FDA had asked me, I’d have added it too.

The drugs didn’t change on November 9. What their labels must say did, and a risk that has earned a line on the U.S. label belongs in every consent conversation. That is a good thing.

For most patients, the right response is a better conversation rather than a different prescription: name the symptom, ask the 3 questions, and say what to do if it happens. For the small group who have already had NAION, the warning should change the decision, or at least who is in the room when it’s made.

And for anyone forwarding headlines: the study that launched this reported more than 4 times the risk, and FDA’s own analysis, 2 years later, put it at 1.2 to 1.4 times the risk seen with other diabetes drugs, on a condition that is still rare. Both facts are true, and only one tends to make the headline.

Clinicians: are you adding a line to your consent notes, or ordering baseline eye exams? Reply and tell me what you’re doing. Patients: what do you still want answered? Reply with that too.