An AI Engine Said a Medicine on the UAE Register Was Not Approved There. A Mention Count Would Have Scored It Present.

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

Yet OpenAI’s gpt-4o-mini was asked one fixed head-to-head question 13 times through OpenAI’s API, in the voice of a buyer in the UAE. In six answers it said the medicine was not approved there, and it recommended the competitor.

In a head-to-head, both medicines are named by definition. A visibility measure scores both as present.

All six concluded for the competitor after a statement about approval, not about presence, and the regulator’s register contradicts that statement.

In AlphaCitation’s published records, AI engines queried through their APIs stated access facts directly:

• whether a medicine is approved;
• whether it is available;
• which medicine to choose for a patient in a market.

Of the five records drawn on here, Field Note 06 (OpenAI’s gpt-4o-mini through the API, not the ChatGPT app) is the one in which such a statement contradicts an authoritative record. The other four measure something else:

• where a newer treatment trigger appears;
• whether pathway answers read the same in two markets;
• whether congress readouts could be verified in the answers;
• whether a chosen molecule is credited to its brand.

Field Note 01 says of its own reading: “Nothing here suggests the engines were wrong.”

Access teams can read the approval statements per market and per engine against the regulator’s register, and the availability statements against a supply record where one exists. AlphaCitation’s records describe what the engines said when asked. None of the five records observes what patients, physicians or payers did with the answers.

Approval is stated, and sometimes stated wrongly

In Field Note 06, a fixed head-to-head question was asked 13 times through OpenAI’s API between 22 August and 25 September 2026, in the voice of a buyer in the UAE. The model was gpt-4o-mini, not the consumer ChatGPT app.

Its answers ruled out a weight-management medicine registered in the UAE in 10 of those runs, and all ten recommended the competitor. Six said the medicine was not approved, which the regulator’s register contradicts. Four said it was not available, which the register cannot test, because it is a record of registration rather than supply.

Three of the ten reasoned from a knowledge cutoff of October 2023. The medicine was already registered by that date. On the same days, answering other questions, the same engine said the medicine was approved or on sale there.

The record also holds approval statements from two other engines that the register contradicts: five answers said that a tablet registered in the UAE in 2026 was still awaiting approval, four from Gemini and one from Claude. Perplexity met the out-of-date claim about the injectable medicine in its sources seven times, and each time reported it as one source’s view without adopting it.

Across the record’s 3,104 answers from four engines:

• the regulator’s website appeared among the sources of 127 answers;
• its live drug register appeared among the sources of none.

A trigger approved in Saudi Arabia appeared with the product name, not with the clinical question

In Field Note 01, Saudi Arabia’s regulator approved a breast-cancer therapy 112 days before the FDA granted it accelerated approval. The Saudi approval covers an ESR1 mutation that emerges during first-line hormonal therapy: a trigger during first-line therapy, ahead of progression.

Four engines answered in a Saudi Arabia market frame, in a single pass from 31 August to 2 September 2026, before the US approval:

• The earlier trigger appeared in all four answers to a question naming the product.
• It appeared in none of the four answers to the unbranded clinical question about when to test.

These were different questions, so the contrast is descriptive, not controlled, and the note does not establish why.

Nor does the note find an error. Testing at progression is well-grounded, and the FDA states that it is not yet confirmed whether intervening earlier delivers a clinically meaningful benefit; it has required confirmatory studies. What the reading shows is where a fact approved by the Saudi regulator, and not yet by the FDA, sat in the answers: alongside the brand name and not alongside the clinical question, in that frame, on those dates.

Who sets the standard read the same in two Gulf markets. Where testing and monitoring happen did not.

Field Note 02 put one pathway battery, unchanged, to the same four engines twice: once in a Saudi Arabia frame and once in a UAE frame. The runs were two days apart. It was a single pass, in oncology.

Who sets the standard and who orders the test read the same in both markets. In each market, three of four engines named a national standard-setting body, and three of four named the ordering clinician.

Every reading about where testing and monitoring happen differed between the two markets. Three of the ten market readings produced no count. On each place reading, one market resolved and the other produced no count: Saudi Arabia resolved to 0 of 4 on two readings where the UAE produced none, and the UAE resolved to 3 of 4 on the third, where Saudi Arabia produced none.

The record is a descriptive contrast between two market frames, not a controlled comparison. Field Note 02 offers it as a question to ask in each market, not as a number to carry from one to another.

Field Note 01, reading the Saudi half, sets the limit on what such a reading means: “The absence of a named route is missing information, not a missing service.”

The study’s inspection procedure verified none of seven registered congress readouts in its own series’ answers by day 10 to 12

Research Paper 01 followed four developer-API model endpoints with web search through EASD 2026, from 22 August to 11 October 2026. Their question batteries covered three diabetes and obesity medicines in the UAE, Singapore and Saudi Arabia.

For none of the seven registered trial readouts did the study’s implemented inspection procedure verify a qualifying citation or result text in its own series’ answers by day 10 to 12.

In the registered arms, seven of 2,408 after-window answers qualified under the study’s either-route definition:

• six through cited sponsor releases or media;
• one by reporting a congress result without a qualifying source.

The paper states that this is “a finding about what the implemented procedure verified, not about whether the readouts reached the engines”. It also states that up to a third of new-URL citations could not be read. The study covered congress readouts. It did not study launches or label changes.

For an access team, the step that transfers is the paper’s own: ask readout by readout and source by source.

Choosing the molecule is not the same as crediting the brand

In Field Note 05, three engines were each asked for a treatment decision for a described patient in four settings. They were queried through their APIs on 27 September 2026, with no market named.

Perplexity’s API model (sonar) named the product in 26 of the 28 Alzheimer’s decisions in which it chose the medicine’s molecule (93%). It named the product in none of the 30 lung-cancer decisions in which it chose the molecule. In the same lung-cancer decision, all three engines chose the same molecule, and Gemini and OpenAI almost always named the product.

Brand credit can be lost even when the engine chooses the medicine’s molecule. The note makes no judgement about which medicine is clinically right. Because no market was named, it does not show whether attribution changes by market.

Four readings an access team can take that a mention count does not

A mention count does not separate these cases:

• In a head-to-head both medicines are named by definition, so a visibility measure scores both as present (Field Note 06).
• A brand that goes unnamed looks the same in a mention count whether its molecule was chosen or not (Field Note 05).

The four readings below take what a mention count leaves out:

• Status claims. Check whether an answer says a medicine is approved against the regulator’s register for that market. Whether it says the medicine is available is a separate question. A register of registration cannot settle it, and it needs a supply record.
• Market divergence. Ask the same question for two markets. Check whether each answer returns the local approval and the local route to care, read in each market rather than taken from one.
• Evidence uptake. Record when a new readout, label or approval is first detected in sampled answers. Count it in days from publication, and read it as an interval set by the sampling schedule.
• Attribution. Check whether an answer that chooses the medicine’s molecule also names the brand.

Read each per market and per engine. The five records differ on where the differences fall:

• In Field Notes 05 and 06 and in Paper 01, the engines answered differently from one another.
• In Field Note 01, on the three questions of its comparison, all four engines gave the same coded result: 0 of 4, 0 of 4 and 4 of 4. On its route question, three of four named a national standard-setting body.
• In Field Note 02, pathway answers read the same in two markets on who sets the standard, and differently on where care happens.

These five records do not measure whether status claims, evidence uptake or attribution differ by market.

For an asset facing an approval, take the reading before the event. Afterwards, as Field Note 01 puts it, “the prior state cannot be recovered”.

Limits

• These are AlphaCitation’s own investigations, through engine APIs rather than consumer applications, on the medicines and markets each one describes. Field Notes 05 and 06 do not name the medicines they measured, and Field Note 05 named no market to the engines.
• Field Notes 01 and 02 were single passes, and their comparisons are descriptive, not controlled.
• Field Note 06’s counts describe the record; they are not a test.
• The five records describe what the answers said. They do not show how often it happens across medicines, markets or engines, why the answers said it, or what anyone did with the answers.
• Each linked investigation states its own scope and limits.

AlphaCitation measures how AI engines, queried through their developer APIs, answer questions about a medicine’s approval, availability and place in treatment: the claims an answer turns on, the sources behind them, and whether approval statements agree with the regulator’s record.

How these readings are taken is set out on our methodology page. To find out which approval and availability statements the answers about your medicines carry, request a briefing.