The Medicine Was Approved. ChatGPT Ruled It Out Anyway.

A weight-management medicine had been active on the UAE regulator’s register for years.

Yet in 10 of 13 runs of one fixed head-to-head question between 22 August and 25 September 2026, asked in the voice of a buyer in the UAE, ChatGPT said that it was not approved or not available there, and recommended its competitor.

The stranger part: on those same days, answering other questions, the same engine said that the medicine was approved, or on sale in UAE pharmacies.

Across 3,104 answers from four AI engines, the regulator’s website appeared among the sources of 127. Its live drug register appeared in none.

The fact existed. The engine could retrieve it. When the question became a choice, the wrong version of the fact won.

“ChatGPT” here means the engine we measure: OpenAI’s gpt-4o-mini model, queried through OpenAI’s API on a fixed schedule. The consumer ChatGPT app runs other models with its own search and may answer differently. This note makes no claim about it.

The engine held the right fact and decided on the wrong one

ChatGPT answered the same head-to-head thirteen times. In six answers it said the medicine was not approved for use in the UAE; in four, that it was not available there. All ten recommended the competitor, in words such as “the available and legally prescribed option”. The error appeared in every one of the last seven cycles, from 9 to 25 September.

Three of the ten reasoned from a knowledge cutoff of October 2023. The medicine was already registered by then, so the two of them that said it was not approved were wrong even on their own stated terms; the third said it was not available, which the register cannot test.

Answering other questions in the same cycles, the same engine said the medicine “is approved for weight management in the UAE” and named pharmacies it said sell it. Its answers carried both versions, and in the six “not approved” head-to-heads the version that decided the answer was the one the register contradicts. Perplexity met the same out-of-date claim in its sources seven times; each time it reported the claim as one source’s view against others and did not adopt it.

May came and went. The answer stayed in April.

The regulator, the Emirates Drug Establishment, registered Foundayo (orforglipron), Eli Lilly’s once-daily oral GLP-1 tablet, on 2 April 2026. Its announcement the next day said the treatment “is expected to become available to eligible patients starting May 2026.” On the day it was published, that was a correct forecast.

Of 733 answers in the same record that named Foundayo, 52 still described its UAE availability as something to come. Twenty-five gave May 2026 as the expected date, in answers as late as 25 September, four months after May had passed. Gemini gave 35 of the 52. A tablet registered in the UAE this year fared worse: five answers, four from Gemini and one from Claude, said it was still awaiting approval.

The engines found the regulator. They missed the page that mattered.

The obvious explanation is that the engines lacked an authoritative source. They did not. Of the 3,104 answers, 127 drew on the regulator’s own website. Every regulator page visible among their sources was an approval announcement. None was the drug directory, the page that gives each medicine’s status on the day it is read.

An announcement records an event in the tense of its date. A register records a state and is kept current. The engines reached the authority and chose the kind of page that ages, so an answer carried the date of its source rather than the date of the question.

The brand was visible. The false fact still decided the answer.

In a head-to-head, both medicines are named by definition. A visibility measure scores both as present and misses the most important event in the answer: one medicine lost because the engine said that it was not approved (six answers, which the register contradicts) or not available (four, which the register cannot test). This is not a visibility problem. It is a decision error.

Its cost does not show up as a lower share of answers. A buyer told that a registered medicine is unapproved has been given a reason not to ask a doctor about it. A buyer told that a registered medicine is “expected soon” has been told to wait.

The record that settles it already exists, and no company writes it: the regulator’s register, dated entry by entry. Regulatory Affairs owns that record, and an AI engine queried through its public API contradicted it. Medical Information owns what follows when a false approval statement reaches the public. Brand and Commercial lost on status, not efficacy, to a competitor that won by default.

How we tested this

Visibility tells you that you were present. It does not tell you why you lost.

A brand can be named in every answer and still be ruled out by one false sentence about its approval. That is why AlphaCitation measures more than whether a brand appears. We trace the claims that decide an answer, the sources behind them, and whether material statements agree with the authoritative record.

How these readings are taken is set out on our methodology page. To find out which claims are deciding the answers about your medicines, request a briefing.