Field Note 08
Same Medicine, Same Model, Same Day. A Patient’s Question and a Doctor’s Got Different Answers.
To a question worded as a Dubai patient’s, Claude Haiku 4.5 said in all three repetitions that tirzepatide is prescribed in Dubai “under approved indications for weight management”, and in two that Mounjaro is licensed in the UAE for adults with obesity or overweight. To one worded as a UAE physician’s, it said in all three that weight management is off-label. Without the UAE-approved label, the study documents the disagreement but cannot say which answer is right.
The most revealing result is not that four AI models disagreed. It is that one model gave different accounts of the same medicine’s approved use to a question worded as a patient’s and to one worded as a physician’s.
On 22 October 2026, AlphaCitation put thirteen fixed questions, three times each, to four AI models with web search on: GPT-4o mini, Claude Haiku 4.5, Gemini 2.5 Flash and Sonar. The 156 answers concern two Eli Lilly medicines in the United Arab Emirates: Foundayo, a once-daily pill for chronic weight management approved in April 2026, and Mounjaro, a weekly injection first registered there in 2022. Some questions were written in the voices of patients, a physician and a buyer; all were sealed and hashed before the first call, and no patient or physician took part.
Foundayo is the clear case. The Emirates Drug Establishment (EDE), the UAE’s medicines regulator, announced its approval on 3 April 2026 and called the UAE “the second country in the world to register the medication” [2]. In 95 of 96 answers about its status, the models said it was approved, registered or could be prescribed in the UAE, and none said it was not approved. In 89 of the 95 the model states approval or registration in the UAE outright. The other 6, Claude’s and Sonar’s answers to the Dubai patient, say that a doctor there can prescribe it; the rule fixed before collection counts that as a status statement, as it counts “not legally prescribable” as a false one. The 96th answer said only that Foundayo was on sale. Availability is not registration, so it is not counted. The result held whether the question named the brand, the molecule or only “Eli Lilly’s once-daily GLP-1 pill”.
Mounjaro exposes a different problem. All 36 answers gave it a correct status under the same rule: 28 stated its approval or registration in the UAE, and 8, all to the Dubai patient’s question, said only that a doctor there can prescribe it or that it is available there by prescription. What they did not agree on is what it is approved to treat. Of the 36, 15 counted weight management among its approved uses, with type 2 diabetes. 6, all to the physician’s question, gave type 2 diabetes as the only approved use and called weight management off-label, 5 of them for the Mounjaro brand specifically and Sonar’s 3 unless the UAE product information lists it. 4 gave type 2 diabetes as the main use without saying whether weight management is approved. 11 made no UAE approved-use statement.
The sharpest disagreement is inside one model. To the question worded as a Dubai patient’s, whether a doctor can prescribe Mounjaro for weight loss, Claude Haiku 4.5 said in all three of its answers that tirzepatide is “widely prescribed off-label and under approved indications for weight management” in Dubai, and in two of them that Mounjaro is “licensed in the UAE for adults with a BMI ≥30, or ≥27 with a weight-related condition”. To the question worded as a UAE physician’s, in all three of its answers, it said Mounjaro “holds MoHAP (Ministry of Health and Prevention) registration in the UAE for type 2 diabetes” and that weight management is off-label. In two of them it was explicit: Mounjaro is “not formally approved in the UAE for weight management under the Mounjaro brand”. Same model, same day, same medicine.
Which answer is right is a question for the label, and the label is a record this study does not hold. The directory MOHAP publishes shows that Mounjaro was registered, first in 2022, and files it under a category, “ENDOCRINE SYSTEM-DIABETES”, not an approved use [3]. So the study scores neither reading. That limit does not erase the finding: the same model, on the same day, gave two accounts of a regulatory fact on which patients and prescribers both rely.
It matters because the point on which the answers split decides whether a prescription is on-label or off-label.
Download this Field Note as a PDF.
What the four models said
The two halves of the finding sit side by side. On status the four models agree with the record and with each other. On approved use they divide, and the division runs through the models as well as between them.
| Model | Said approved, registered or prescribable | Said not approved | Hedged or no status statement |
|---|---|---|---|
| GPT-4o mini | 24 of 24 | 0 | 0 |
| Claude Haiku 4.5 | 23 of 24 | 0 | 1 |
| Gemini 2.5 Flash | 24 of 24 | 0 | 0 |
| Sonar | 24 of 24 | 0 | 0 |
| Model | Type 2 diabetes and weight management | Type 2 diabetes only, weight management off-label | Type 2 diabetes as the main use | No UAE approved-use statement |
|---|---|---|---|---|
| GPT-4o mini | 6 | 0 | 0 | 3 |
| Claude Haiku 4.5 | 6 | 3 | 0 | 0 |
| Gemini 2.5 Flash | 3 | 0 | 1 | 5 |
| Sonar | 0 | 3 | 3 | 3 |
| All four | 15 | 6 | 4 | 11 |
Three models stated Foundayo’s status correctly in all 24 of their answers and Claude in 23, 95 of 96 in all (Wilson 95% interval 94.3% to 99.8%, computed as if the 96 answers were independent; three answers to one question from one model at temperature 0 are not, and 31 of the 32 question-and-model pairs were correct in all three rounds). None was false or hedged. The result did not bend to the wording. The only cuts short of a full score are the two that contain Claude’s one answer without a status statement: the brand name alone drew 35 correct answers of 36, and the buyer choosing between Foundayo and Mounjaro 23 of 24. Every other cut was complete: the molecule’s name 24 of 24, both names together 24 of 24, the bare description 12 of 12, patients in Dubai and Abu Dhabi 24 of 24, the UAE physician 12 of 12 and a direct question 36 of 36. Asked which daily GLP-1 pills are approved in the UAE, and which new weight-loss medicines it approved in 2026, every model named Foundayo in all six of its answers, each time with a correct status. No answer gave the pill the wrong form, and all 80 answers that named a UAE use for it said weight management, the use the EDE announced.
The two buyer questions repeat, word for word, two that our Field Note 07 study put to GPT-4o mini eight times on 13 to 16 October [1]. Then, 7 of 8 answers were correct and 1 hedged. This time, 6 of 6 were correct.
Mounjaro was the control, and it held: each model gave its UAE status correctly in 9 of 9 answers. None of the four was simply unable to state a UAE status. That is what makes the approved-use split informative. The models gave Mounjaro’s status correctly. They disagreed about what it is for.
| Model | Is Mounjaro approved in the UAE? | A Dubai patient asks | A UAE physician asks |
|---|---|---|---|
| GPT-4o mini | Both uses, 3 of 3 | No UAE statement, 3 of 3 | Both uses, 3 of 3 |
| Claude Haiku 4.5 | Both uses, 3 of 3 | Both uses, 3 of 3 | Type 2 diabetes only, weight management off-label, 3 of 3 |
| Gemini 2.5 Flash | No statement, 3 of 3 | Type 2 diabetes as the main use 1, no statement 2 | Both uses, 3 of 3 |
| Sonar | Type 2 diabetes as the main use, 3 of 3 | No UAE statement, 3 of 3 | Type 2 diabetes only, weight management off-label, 3 of 3 |
Read down the columns and the disagreement is between models: to the same questions, GPT-4o mini gives both uses and Sonar gives type 2 diabetes. Read across Claude’s row and it is inside one model, which is the sharper finding. Gemini also gave both kinds of answer: type 2 diabetes as the main use in one of its three answers to the patient’s question, and both uses in all three answers to the physician’s. GPT-4o mini’s patient answers said Mounjaro is “not specifically approved for weight loss”, in sentences that do not name the UAE; one of them names the United States.
One answer went further and ruled Mounjaro out. In one of its six buyer answers, GPT-4o mini said “Mounjaro (tirzepatide) is not currently available in the UAE market”, with no source on the sentence, and concluded that “Foundayo is the viable option”. The question, in the instrument’s buyer frame, asked the model to favour brands available in the UAE. The directory gives Mounjaro a first registration date of 18 May 2022 [3]. Availability is not something a register records, so the study does not call the sentence false. It reports it because a recommendation was made on an availability claim with no source. In the other two rounds, the same question drew answers saying Mounjaro is available in the UAE: “Available in select pharmacies” in one, and “Also approved in the UAE” and “available through prescription at licensed pharmacies” in the other.
What the records establish, and what they cannot
In every one of the 96 Foundayo status answers, the model searched. Of the 628 pages cited in those answers, 238 were news or trade-press pages, 144 UAE clinic, pharmacy or telehealth pages and 64 UAE regulator pages. A UAE regulator page (the EDE, MOHAP, Abu Dhabi’s Department of Health or the Dubai Health Authority) was among the sources of 21 of 24 Sonar status answers, 17 of 24 Gemini answers (grounding sources, classed by the domain Google gives as each source’s title) and 9 of 24 GPT-4o mini answers. Claude cited no UAE regulator page in any of its 39 answers; it was still right on status in 23 of 24, and its status answers cited news, clinic, pharmacy and other pages. Seventeen of GPT-4o mini’s 39 answers followed a search query carrying the year “2023”, such as “orforglipron approval status UAE 2023”; all 17 were coded correct. The record shows what each answer cited, not why it was right.
The records behind the two medicines differ. For Foundayo there is a dated regulator statement of the use: the EDE’s announcement says “chronic weight management” [2], and every answer that named a UAE use matched it. For Mounjaro the directory records no use, and the closest UAE record we hold is the Dubai Health Authority document [6], which says tirzepatide is “registered for obesity pharmacotherapy in UAE”; it is not a label. Both readings cited UAE clinic and pharmacy pages: all 15 “both uses” answers and all 10 that gave type 2 diabetes alone or as the main use cited at least one. Two GPT-4o mini physician answers quoted the Dubai Health Authority line through glpguidebook.com. The physician answers from Sonar and Claude said weight management is off-label for the Mounjaro brand, and some said the obesity-labelled tirzepatide brand is a different product. Only the UAE-approved product information would decide between them.
The two readings make different answers wrong. If Mounjaro’s UAE label includes weight management, the 6 answers that call it off-label understate the label, all of them to the physician; the 4 that call type 2 diabetes the main use neither deny nor state it. If it does not, the 15 “both uses” answers overstate it: 6 to the direct question (GPT-4o mini and Claude), 3 to the Dubai patient (Claude) and 6 to the UAE physician (GPT-4o mini and Gemini). Either way, a reader asking the same model in two voices can be told two things.
The study shows four models on one day, as AlphaCitation calls them, answering 13 fixed questions three times. Repeated answers at temperature 0 are not independent: the honest unit is 8 status questions on each of 4 models, and 31 of those 32 were answered correctly in all three rounds, Claude’s answer to one buyer question in two of three. A correct status is not a correct answer in every respect. Two answers called Foundayo the first oral GLP-1 medicine, or the first once-daily one, with no qualifier; under AlphaCitation’s register conduct rule the study keeps no rows for other medicines, so it does not score these sentences. One GPT-4o mini answer, after saying the pill was approved, reported a 2025 Department of Health circular, without its date, as “emphasizing that the authentic product remains in clinical development”; the study lists that reading as disputed. Every answer searched; a model answering from memory alone was not tested.
Field Note 07 found one engine calling a registered medicine not approved or not available in 20 of 52 answers to one buyer’s question and its paraphrases [1]. This study found no false status, for a different medicine, from four engines. The studies differ in medicine, date and design, and together they say nothing about why.
The study does not establish which account of Mounjaro’s approved use is correct; that needs the current UAE-approved product information. What it establishes is a different problem: a model can state a medicine’s registration correctly and still give different accounts of what it is approved to treat. In Claude’s answers the account differed between the two questions. The patient’s and the physician’s questions ask different things, whether a doctor can prescribe it for weight loss and what it is approved for, so the study does not show that the voice of the question caused the difference.
For a regulated medicine, a plausible answer is not enough. The evidence behind each claim must be identifiable, and authoritative for the claim being made.
How we know this
| Question | Do four AI models state a 2026 UAE registration of a new oral medicine correctly; how do they state its status, form and approved use; does that change with the name used and the framing? |
|---|---|
| Plan | Frozen and hashed before the first call, 22 October 2026 03:52:18 UTC; plan sha256 e0a564b1…07ba, collection script sha256 0cb5d3ab…90be6, sealed wordings sha256 0702bf32…88a9e. Changes after the freeze are numbered deviations in the study’s result. |
| Models | gpt-4o-mini (served as gpt-4o-mini-2024-07-18), Responses API with web search; Claude Haiku 4.5 (claude-haiku-4-5-20251001) with web search, at most 5 searches; Gemini 2.5 Flash with Google Search grounding; Sonar (Perplexity chat completions). Temperature 0; no system instruction; each model chose whether to search. |
| Questions | 13 wordings: 3 direct (brand, molecule, description), 2 patient, 1 physician, 2 buyer’s choices against Mounjaro (byte-identical to two Field Note 07 wordings), 2 open questions not naming the medicine, 3 Mounjaro controls. Each is identified by its hash; the texts are withheld, as Field Note 07’s were. |
| Schedule | 22 October 2026, 03:52 to 04:57 UTC: 3 rounds, each asking every wording once on every model in a seeded order, at least 30 minutes apart. 156 calls; 156 answered. |
| Coding | Each answer coded for its subject: correct, false, availability-negative, hedged or not addressed, by rules fixed before collection, with flags for stale claims, form, approved use and other wrong regulatory facts; flags added after collection are listed in the study’s result. Every answer read in full. A keyword pass flagged 21 answers for a second reading; none changed on that ground. After an adversarial audit, one answer was recoded from correct to no status statement and the Mounjaro approved-use flags were split, both recorded as deviations. |
| Truth | Registration: the registered medical product directory that MOHAP publishes [3]. MOHAP has told AlphaCitation in writing that registration has moved to the EDE and that the directory is not the most up-to-date source of registration data [7]: a listing shows that a product was registered, but the directory’s currency is not assured. For Foundayo, the EDE’s own approval announcement [2] corroborates the registration; for Mounjaro, no EDE record was read. The engines’ statements were coded against the directory’s dates, which no record we hold contradicts. Approved use: no UAE label in hand; Foundayo’s stated uses compared descriptively with the EDE’s announced use; the Dubai Health Authority document [6] is reported for Mounjaro and decides nothing. The FDA [4] and the EMA [5] are reference benchmarks only and decide nothing about the UAE. Availability: no dated supply record in hand, so no availability statement is called false. |
| Gemini | Reported as counts only: this note and its files carry no Gemini answer text, no grounding-redirect link and no source address. |
| Pre-specified | Coding classes, the eight-question status set, the control, the name and framing comparisons, the reading thresholds: a model is current if none of its 24 status answers is false and at least 18 are correct. All four were. A statement that the medicine can be prescribed in the UAE counts as a status statement; 6 of the 95 correct Foundayo answers rest on it alone, and for Mounjaro, 8 of the 36. |
| Added after collection | Mounjaro’s approved use by model and question (Exhibits 2 and 3); the Dubai Health Authority document; the one availability sentence about Mounjaro; the “first” claims; the disputed Department of Health sentence; the “2023” search queries; regulator citations by model. |
| Reproduce | Every count in this note is printed by the study’s tally script from the per-answer codes file, except the register rows and dates, which come from the register record; Field Note 07’s 7 of 8, from that study’s codes; and the two “first” answers and the one disputed Department of Health sentence, which rest on coder notes the released file withholds. The Dubai Health Authority line quoted through glpguidebook.com, and the counts of quoted wording in Claude’s and GPT-4o mini’s Mounjaro answers (in all three, in two of three, the one naming the United States), are readings of answer text, not counts from the codes file. The released files are the codes file, which holds for each answer the code and flags, the response id tail, a text hash, the search count and the classes of cited pages, and no question wording, answer text, coder notes or host names; the tally script; and the freeze record. |
The records
- MOHAP-published directory. The official “Download all” export of the registered medical product directory. Read for coding on 21 October 2026 at 22:48:36 UTC, and re-read through AlphaCitation’s register-of-record connector on 22 October 2026 at 05:47:33 UTC for the two medicines only (17,025 rows; newest registration 22 October 2026), with the same rows and dates. Foundayo: six tablet strengths (0.8, 2.5, 5.5, 9, 14.5 and 17.2 mg), prescription only, first registration 02/04/2026. Mounjaro: six single-dose prefilled pens first registered 18/05/2022 and six KwikPen multi-dose pens first registered 18/01/2024. The export records no registration number or status; its “Body System” column is a category and its “Therapeutic Group” column is empty for all 18 rows.
- MOHAP Statistics and Research Centre, written reply to AlphaCitation, 22 October 2026 [7]: responsibility for registering and regulating pharmaceutical and medical products has transferred to the EDE; the directory on the MOHAP portal is not considered the most up-to-date source of registration data; registration information should be confirmed with the EDE.
- EDE announcement, the EDE’s own record of the approval, published 3 April 2026, read once on 22 October 2026 from a path its robots.txt allows (sha256 40eb8759…dc8c; copy kept privately): approval of Foundayo “for chronic weight management”; the UAE “the second country in the world to register the medication”; availability “expected” from May 2026.
- Dubai Health Authority, 3 September 2024, page 9: tirzepatide (Mounjaro) a “GIP/GLP-1 agonist approved by FDA and registered for obesity pharmacotherapy in UAE”. Read for Field Note 07 on 19 October 2026 and again on 22 October.
- FDA, 1 April 2026: Foundayo, a once-daily oral tablet, approved “for use in combination with a reduced-calorie diet and increased physical activity” to reduce excess body weight and maintain weight reduction long term in adults with obesity, or with overweight and at least one weight-related comorbid condition.
- EMA, 15 October 2026: the CHMP adopted a positive opinion recommending authorisation of Foundayo for weight management and for adults with insufficiently controlled type 2 diabetes; no European Commission decision was shown when read on 22 October. No answer claimed an EU approval.
Limitations
- One day, four API models and 13 fixed questions; not the consumer apps, not other models, not other markets.
- Three answers to one question from one model at temperature 0 are not independent draws, so neither are the per-model rates.
- No UAE label for either medicine was read, so no approved-use statement is scored. No dated supply record was read, so no availability statement is scored.
- Coding was done by one coder, an AI model (below), under the founder’s direction; no human second coder has reviewed the codes.
Disclosures
Relationship with Eli Lilly. AlphaCitation has no commercial relationship with Eli Lilly, the maker of Foundayo and Mounjaro. AlphaCitation may present these findings to Eli Lilly when offering its services. Eli Lilly cannot edit, approve or suppress the finding.
Funding. The author funded the study. No one paid for this study, and no company commissioned or funded it.
Competing interests. The author founded Akunudo LLC-FZ, which owns and operates AlphaCitation. AlphaCitation sells measurement of how AI engines represent medicines. The manufacturers of the medicines measured in this study are, or may become, commercial prospects of AlphaCitation.
- The author was employed by Novo Nordisk, which makes competing weight-management medicines, from 2013 to March 2025; the last role was Pricing and Access Director for its obesity and diabetes portfolio in Asia-Pacific.
- Novo Nordisk is a commercial prospect of AlphaCitation, not a client. In 2026, before this study, AlphaCitation sent Novo Nordisk staff analyses of one of Novo Nordisk’s medicines in a commercial conversation.
- The author holds no shares, share options, pension entitlement or deferred compensation from Novo Nordisk, and no shares or options in Eli Lilly, and has no employment or advisory role with either.
Relationship with a cited website. Three gpt-4o-mini answers cited glpguidebook.com, a site discussed in Field Note 07. Neither the author, AlphaCitation nor Akunudo LLC-FZ has any relationship with it.
Coding and drafting. Coding was done by Anthropic’s Claude, which is made by the company behind one of the four models measured. To limit that, each code on a non-Gemini answer keeps its deciding sentence in the study’s internal record, and a keyword pass flagged answers for re-reading. The note was drafted with AI assistance and checked against the study record in adversarial audits, also AI-assisted.
Scholarly correspondence and substantive commentary: analyst@alphacitation.com
References
- 1. AlphaCitation. Wegovy Was Approved. An AI Model’s Answer Changed With How the Question Was Framed. AlphaCitation Field Note 07. 22 October 2026.
- 2. Emirates Drug Establishment. Emirates Drug Establishment approves oral obesity treatment, marking shift away from injections. News, 3 April 2026. Read 22 October 2026; sha256
40eb875980ea4a9b95fade2028235970416c337884bcffad306f1bd1d00bdc8c. - 3. Ministry of Health and Prevention. Registered medical product directory, “Download all” export. MOHAP states that registration has transferred to the EDE and that this directory is not the most up-to-date registration source [7]. Read 22 October 2026 05:47:33 UTC (sha256
413465f41a43dc698a1970212b09233e3e6dd581226928cc7e2e4a7a0bc1a00d); coding-time read 21 October 2026 22:48:36 UTC (sha256d587f70c37cfbd81d955c60ef02ea98771fe747e664dfde5331b4ba1f00c7b9b). - 4. US Food and Drug Administration. FDA approves first new molecular entity under national priority voucher program. Press announcement, 1 April 2026.
- 5. European Medicines Agency. Foundayo (orforglipron): CHMP positive opinion, adopted 15 October 2026; page last updated 16 October 2026. Read 22 October 2026.
- 6. Dubai Health Authority, EJADA Program. Obesity and Metabolic Syndrome KPIs and Recommendations. DHA/DHIC/HEIPD/CG-12, Version 2. Issue and effective date 3 September 2024. Page 9. Hosted on isahd.ae; read 19 and 22 October 2026; sha256
b65bc414be003195c12fe027221a0192f4554f2f64a56253013b11c1b1209075. - 7. MOHAP Statistics and Research Centre. Written reply to AlphaCitation. 22 October 2026.
AlphaCitation Research is the research programme of AlphaCitation, a digital product of Akunudo LLC-FZ.
© 2026 AlphaCitation (Akunudo LLC-FZ). Akunudo LLC-FZ is the copyright holder and licensor of the original material in this publication, licensed under CC BY 4.0. Quoted third-party text, third-party data fields, API/provider-restricted material and third-party marks are excluded from that grant.