The LabelTrue Index, round 1: GLP-1 medicines
AI answers about GLP-1 medicines, checked against the FDA label
Patients and clinicians now ask AI tools about their medicines. We asked five of them the same 54 questions about GLP-1 medicines for diabetes and weight, each in a fresh chat, and checked every answer that named a GLP-1 brand against that brand's current US label.
How five AI tools answer questions about GLP-1 medicines, on five measures
Ranked on agreement with the label, highest first. Collected 15 November 2026, US connection. Round 1 of a monthly index. Last updated 20 November 2026.
| Rank | AI tool | Agrees with the labelShare of answers that named a GLP-1 brand with no statement the label contradicts | Safety completenessDoes the answer give the label's key warnings for the question asked? Scored 0, 1 or 2 | Fair balanceAre benefits and risks given together? Scored 0, 1 or 2 | Off-label statementsShare of answers that recommend or present as approved a use, age or dose the label does not include | Outdated or inventedShare of answers with an old label fact, an approval that differs from the label, or a made-up side effect or study |
|---|---|---|---|---|---|---|
| 1 | ChatGPT | 98.1%53 of 54 answers | 1.3of 2, across 45 answers | 1.4of 2, across 25 answers | 5.6%3 of 54 answers | 5.6%3 of 54 answers |
| 2 | Microsoft Copilot | 94.4%51 of 54 answers | 1.1of 2, across 43 answers | 0.7of 2, across 17 answers | 3.7%2 of 54 answers | 3.7%2 of 54 answers |
| 3 | Google AI Overviews | 88.7%47 of 53 answers | 1.0of 2, across 43 answers | 1.2of 2, across 13 answers | 1.9%1 of 54 answers | 7.4%4 of 54 answers |
| 4 | Claude | 82.7%43 of 52 answers | 1.2of 2, across 42 answers | 1.5of 2, across 17 answers | 1.9%1 of 54 answers | 11.1%6 of 54 answers |
| 5 | Gemini | 81.5%44 of 54 answers | 1.3of 2, across 43 answers | 1.5of 2, across 12 answers | 7.4%4 of 54 answers | 11.1%6 of 54 answers |
Agreement with the label is out of the answers from that tool that named at least one GLP-1 brand (267 of 270). The other four measures are out of all 54 answers per tool, because they also apply to answers about GLP-1 medicines in general. The sample size sits next to every figure.
Higher is better for agreement, safety completeness and fair balance. Lower is better for off-label and outdated or invented.
The tools are ranked on agreement only. Safety completeness and fair balance do not give a clear order: ChatGPT, Gemini and Claude score within 0.1 of each other on safety (1.2 to 1.3), and fair balance rests on 12 to 25 answers per tool.
The measures are separate. An answer can agree with the label and still be flagged, for example when it presents a medicine for a use its label does not include without contradicting any sentence of the label. 14 answers are in that position.
Safety completeness is scored only where safety is relevant to the question (not for cost, coverage or approval-only questions), and it counts what an answer leaves out. Fair balance is scored only where the answer claims a benefit, so its sample is smaller. Each answer is scored once, across every GLP-1 medicine it discusses.
Every answer was scored on all five measures by two separate scoring passes under rules written before scoring, and every flag was confirmed against the label before it counts.
Patient and clinician questions are shown separately below.
One round, small samples
This is a single round: 54 questions per tool, asked once each. A difference of one or two answers can change the order of tools that sit close together. We will repeat the round each month and show how each tool moves. Answers from these tools change over time, so this index describes what they said on 15 November 2026.
Patient questions and clinician questions
38 of the 54 questions were written the way a patient or parent would ask them, and 16 the way a nurse practitioner, physician assistant, specialist or pharmacist would. Answers that differ from the label, out of the answers that named a GLP-1 brand:
| AI tool | Patient questions | Clinician questions |
|---|---|---|
| ChatGPT | 1 of 38 (2.6%) | 0 of 16 (0.0%) |
| Microsoft Copilot | 1 of 38 (2.6%) | 2 of 16 (12.5%) |
| Google AI Overviews | 5 of 37 (13.5%) | 1 of 16 (6.3%) |
| Claude | 9 of 37 (24.3%) | 0 of 15 (0.0%) |
| Gemini | 7 of 38 (18.4%) | 3 of 16 (18.8%) |
| All five | 23 of 188 (12.2%) | 6 of 79 (7.6%) |
What kinds of differences we found
We report by AI tool only, never by brand. The differences fell into a few kinds.
-
Dosing
Dosing after missed doses
The label says what to do when doses are missed, including when to restart at a lower dose. Some answers gave a different rule.
-
Dosing
Maximum dose
Some answers gave a maximum dose lower than the one on the current label, or described a labeled dose as still being studied.
-
Safety
A warning that has since been removed from the label
Some answers still described a warning that the current label no longer carries.
-
Approval
Who the medicine is approved for
Approved uses, age groups and patient groups that differ from the label: a use described as approved when it is not, an approved use described too narrowly, an age or weight limit the label does not set, or a date for when an approval was added that differs from the label.
-
Use
How to take it
Instructions for taking a tablet that differ from the label's instructions.
-
Safety
Who should not take it, and other safety wording
A condition given as a reason not to take the medicine when the label does not list it, breastfeeding advice that differs from the label, and the animal study findings behind a boxed warning described differently from the label.
-
Studies
Study results
Results from the label's clinical studies described differently from how the label describes them.
Labels change. Several of the differences we found follow a recent label change that the answer had not caught up with.
How we checked
Every round uses the same steps.
- Questions
- 54 questions, written the way patients, parents, nurse practitioners and physician assistants, specialists and pharmacists would ask them. They cover approvals, side effects and safety, who should not take a medicine, dosing, comparisons and switching, cost and coverage, and children and teens.
- AI tools
- ChatGPT, Claude, Gemini, Microsoft Copilot and Google AI Overviews. Each answer was asked in a fresh chat (a fresh search for Google AI Overviews), once per question.
- When and where
- 15 November 2026, from a US connection. 270 answers in all.
- The label
- Each answer was compared with the current US label on DailyMed for every GLP-1 brand it named. An answer that named three brands was checked against all three labels: 599 answer and label pairs in this round.
- What counts
- For agreement with the label, a statement counts when the label contradicts it, and only differences confirmed after review count. That measure does not count what an answer left out, statements the label does not address, cost and coverage, or study results given with their source. Safety completeness does count what an answer leaves out, and off-label and outdated flags count only once confirmed against the label.
- Denominators
- Agreement with the label is out of each tool's answers that named at least one GLP-1 brand (267 of 270 answers). Safety completeness, fair balance, off-label and outdated or invented are out of all 270 answers, with safety and fair balance scored only where they apply.
- Next round
- This is round 1. We plan to repeat it monthly with the same questions and add other medicine areas.
See what AI says about your brand
The index shows the category. A LabelTrueAI brand report shows your medicine: what each AI tool tells patients and clinicians about it, side by side with your current label.
- Every answer that differs from your label, with the label section it differs from
- Questions written from patient, caregiver and clinician viewpoints, in the words they use
- The pages each answer cites, so your team can see where an answer comes from
- Repeat checks, so your team can see when answers change