More than 40 million people use ChatGPT every day to ask health questions, according to an OpenAI report published in January 2026. [1] The questions cover insurance, symptoms, and medicines. The answers shape what patients bring to an appointment and what clinicians check before they prescribe.
A typical pharmaceutical executive briefing covers regulatory actions, trial readouts, competitor moves, and sales data. We think it now has a gap: it carries no line on what AI systems say about the company’s own drugs.
This article contains no new AI test results. It combines published studies, FDA records, and our own calculations from those sources. It also proposes a four-type classification for AI findings and a one-minute briefing block that fits inside a ten-minute read. We label every calculation and every original framework as ours.
The Short Answer: What Is Missing From a 10-Minute Executive Briefing?
The missing input is a dated, repeatable record of what AI systems say about your drugs, compared against the current prescribing information. A one-minute block of about 238 words can hold it. The inputs already in the briefing do not cover it, because an AI answer is not a news item, a label, or a sales figure. It is a generated output that changes with the model, the prompt, and the day.
The Eight Findings That Matter
A 10-minute read holds about 2,380 words at the 238 words-per-minute silent reading rate for adult non-fiction readers (our calculation from Brysbaert, 2019). A one-minute AI block therefore holds about 238 words. [2]
In a KFF poll of 1,343 U.S. adults fielded February 24 to March 2, 2026, 32% said they had used AI chatbots for health information in the past year. [3]
In a study of Bing Copilot answers about the 50 most prescribed U.S. outpatient drugs, 126 of 484 answers (26.0%) had accuracy below 100%. [4]
In a BMJ Open audit of 250 prompts sent to five chatbots, 50% of responses were problematic and 20% were highly problematic. [5]
Both audits were published well after they were run. The gap was at least 21 months for the Copilot study and at least 11 months for the BMJ Open audit (our calculation). [4][5]
On February 28, 2025, FDA said it was removing language on adverse cardiovascular outcomes from the Boxed Warning on all testosterone products and adding TRAVERSE trial results to the labels. [6] An answer built from older text would differ from the current label.
In the 5W Pharma/Rx AI Visibility Index for Q2 2026, Eli Lilly and Novo Nordisk together held about 24.0% of estimated citation share (our sum of the published figures). [7]
In the AMA’s 2026 survey of nearly 1,700 physicians, 81% reported using AI professionally. [8]
How Much Can a 10-Minute Executive Briefing Hold?
A ten-minute briefing holds roughly 2,400 words of text if it has no tables. Most briefings have tables, so the real text budget is smaller. The AI block must be short enough to survive that budget.
How Many Words Is a 10-Minute Read?
A meta-analysis of reading-rate studies found that adults read non-fiction silently at 238 words per minute and fiction at 260 words per minute, in English. [2] We use those two figures to size the briefing. The table shows our calculations and the assumptions behind them.
Scenario
Assumption
Calculation
Words
All text, non-fiction rate
10 minutes at 238 wpm [2]
10 x 238
2,380
All text, fiction rate
10 minutes at 260 wpm [2]
10 x 260
2,600
Text with tables and charts
30% of the time goes to tables and charts (our assumption)
7 x 238
1,666
One-minute AI block
1 minute at 238 wpm [2]
1 x 238
238
Source: Brysbaert (2019) for the reading rates [2]. All other figures are our calculations.
A 238-word block is 10% of the ten-minute budget in the all-text case, and about 14% of the text budget once tables take 30% of the time. That is enough room for five or six short lines. It is not enough room for a dashboard export. The block must therefore report findings, not raw answers.
What Goes Into the Briefing Today?
We group the usual inputs into four layers. This grouping is our own working model, not an industry standard.
The evidence layer: FDA labels, safety communications, approval actions, ClinicalTrials.gov records, and journal publications.
The market layer: prescription data, share, pricing, and access decisions.
The competitor layer: filings, deals, pipeline moves, and litigation.
The narrative layer: news coverage, social posts, and congress reports.
The AI answer layer sits next to the narrative layer but behaves differently. We explain how below.
How Does the AI Layer Differ From the News Layer?
A news article has an author, a date, and a fixed text. An AI answer has none of these. The same question can return different text on different days or from different models. A briefing that reports “the AI answer” as a single item hides that variation. It must report a pattern across repeated runs, dated and attributed to a named platform.
How Does the AI Layer Differ From the Evidence Layer?
The evidence layer is the standard. The AI layer is a measured output compared against that standard. An AI answer is not evidence because it sounds fluent. The cited source must support the claim. In the briefing, the label is the reference and the AI answer is the specimen.
How Many Patients and Clinicians Ask AI About Drugs?
Large numbers do, though no source we reviewed gives a drug-specific count. The figures below measure health use in general. We cannot say from them how many people ask about a particular brand.
How Many People Use ChatGPT for Health Questions?
More than 40 million people worldwide use ChatGPT daily for health information, according to an OpenAI report covered by Axios. [1] More than 5% of all ChatGPT messages globally are about health care. [1] The report counted 1.6 to 1.9 million health insurance questions per week. [1] Seven in 10 health care conversations took place outside normal clinic hours. [1]
The Axios summary gave no medication-specific figure. The outside-clinic-hours figure still matters to a briefing reader. It means many of these answers arrive when no pharmacist, nurse, or medical information line is open.
What Share of U.S. Adults Use AI Chatbots for Health Information?
About one in three. KFF reported that 32% of U.S. adults used AI chatbots for health information in the past year. [3] The poll surveyed 1,343 adults from February 24 to March 2, 2026, with a margin of error of plus or minus 3 percentage points. [3] KFF also reported that 29% of adults sought physical health information from AI and 16% sought mental health information. [3] Thirteen percent had uploaded personal medical information, such as test results, to a chatbot. [3] Seventy-seven percent were concerned about the privacy of medical information shared with AI tools. [3]
KFF tracking poll (n = 1,343; fielded February 24 to March 2, 2026): 32% of U.S. adults say they used AI chatbots for health information in the past year. [3]
How Many Physicians Use AI at Work?
Eighty-one percent of physicians reported using AI professionally in the AMA’s 2026 Physician Survey on Augmented Intelligence, which drew nearly 1,700 respondents. [8] The AMA described that rate as more than double the 2023 level. [8]
The use that matters most to a brand team is summarizing medical research and standards of care. Thirty-nine percent of physicians reported that use in 2026, up 33 percentage points from 2023. [8] That implies a 2023 figure of about 6% (our subtraction). Physicians are now asking AI to summarize the evidence that your medical affairs team published or defended.
What Happens to Clicks When Google Shows an AI Summary?
Users click less. Pew Research Center analyzed 68,879 Google searches from 900 U.S. adults collected April 7 to 17, 2025. Users clicked a traditional search result in 8% of visits that showed an AI summary and in 15% of visits without one. [9] That is a relative drop of about 47% (our calculation: 8 divided by 15 is 0.53). Users clicked a link inside the AI summary in 1% of visits. [9] Browsing sessions ended on 26% of pages with a summary, against 16% of pages without one. [9]
Pew studied all search topics, not drug queries. We cite it for the mechanism. If the AI summary is the last page a reader sees, the label and the brand website are less likely to be read next.
Did Google Change AI Overviews After Health Complaints?
Yes, for some queries. A Guardian investigation published in January 2026 reported misleading health information in Google AI Overviews. Its examples included liver function test ranges and dietary advice for pancreatic cancer. [10] ETV Bharat reported that Google disabled AI Overviews for queries the investigation flagged. [10] AI Overviews still appeared for queries such as “normal Vitamin D3 level” and “normal range of Blood Pressure” on January 13, 2026. [10] Google said most AI Overviews are factual and that its clinicians found many flagged answers were supported by high-quality sources. [10]
We read this as a source-conflict example, not a verdict. The Guardian and Google disagree about the accuracy of the same outputs. A briefing cannot settle that disagreement from the outside. It can record the answer, the date, and the label text side by side.
What Do Studies Say About AI Drug-Information Accuracy?
The studies disagree on size but agree on direction. Every audit we reviewed found a meaningful share of answers that were inaccurate, incomplete, or potentially harmful. None found that fluent answers were reliably correct. The table summarizes four studies.
Study
Platform
Sample
Key finding
Source
Andrikyan et al., BMJ Quality & Safety, vol. 34(2)
Bing Copilot (balanced mode)
10 questions x 50 most prescribed U.S. outpatient drugs; 500 answers
26.0% (126 of 484) of answers below 100% accuracy; 3.3% (16 of 484) at 0% accuracy
[4]
ASHP 2023 Midyear posters 8-021 and 8-023
ChatGPT
39 questions (LIU); 30 drugs (Iwate)
10 of 39 satisfactory; 26 of 30 inaccurate against Lexicomp
[11]
Tiller et al., BMJ Open, 2026
Gemini, DeepSeek, Meta AI, ChatGPT, Grok
10 prompts x 5 topics x 5 chatbots; 250 prompts
50% problematic; 20% highly problematic
[5]
Bean et al., Nature Medicine, 2026
GPT-4o, Llama 3, Command R+
1,298 UK adults, 10 scenarios
Users with an LLM identified relevant conditions in fewer than 35% of cases
[12]
What Did the Bing Copilot Study of the 50 Most Prescribed Drugs Find?
What the study measured
A German research team asked Bing Copilot 10 patient questions about each of the 50 most prescribed U.S. outpatient drugs. That produced 500 answers. The team generated them in April 2023, using the “balanced” mode and a New York VPN location. [4]
Of 484 rated answers, 126 (26.0%) scored below 100% accuracy and 16 (3.3%) scored 0%. [4] Of 495 rated answers, 11 (2.2%) scored 0% for completeness. [4] The mean Flesch Reading Ease score was 37.2, which the authors place in the “difficult” band. [4]
Seven experts rated a subset of 20 answers for potential harm. They judged 66% potentially harmful and 22% able to lead to death or severe harm. [4] The authors wrote that AI-powered chatbots can provide overall complete and accurate patient drug information, yet experts deemed a considerable number of answers incorrect or potentially harmful. [4]
What the study does not show
The 66% and 22% figures come from 20 answers. The authors selected that subset for low accuracy, low completeness, or safety risk. [4] The figures do not describe all 500 answers. The study tested one product in one month. Any briefing that quotes the 66% figure as a general error rate misreads it.
How Did ChatGPT Perform on Pharmacist Drug-Information Questions in the ASHP 2023 Studies?
What the studies measured
Two posters at the ASHP 2023 Midyear Clinical Meeting (December 3 to 7, Anaheim) tested ChatGPT. [11] Long Island University researchers collected 45 questions; the satisfactory-response figures refer to 39. ChatGPT answered 10 of the 39 satisfactorily. [11] Of the 29 unsatisfactory answers, 11 gave no direct answer and offered general background instead. [11] ChatGPT supplied references for 8 of the 28 questions that received a direct response. None of the listed references existed. [11]
Iwate Medical University researchers tested 30 FDA-approved drugs for common side effects against Lexicomp. They rated 26 answers inaccurate, 2 partly accurate, and 2 accurate. [11]
What the studies do not show
These were conference posters about an early ChatGPT version. They show that fabricated references appeared in 2023. They do not show the rate for current models. We cite them because they define the failure type that the later BMJ Open audit also reported: reference lists that do not hold up.
What Did the 2026 BMJ Open Audit of Five Chatbots Find?
What the audit measured
Tiller and colleagues tested Gemini, DeepSeek, Meta AI, ChatGPT, and Grok in February 2025. Each chatbot received 10 prompts in each of five areas: cancer, vaccines, stem cells, nutrition, and athletic performance. That made 250 prompts. [5] Fifty percent of responses were problematic: 30% somewhat problematic and 20% highly problematic. [5]
Open-ended prompts produced more highly problematic responses than closed prompts. [5] Grok produced 29 highly problematic responses out of 50 (58%). [5] Overall quality did not differ significantly among the five chatbots. [5] Answers were typically stated with certainty and carried few caveats. Only 2 of the 250 prompts were refused, both by Meta AI. [5]
Reference quality was poor. The average reference completeness score was 40%, and no chatbot produced a fully accurate reference list. [5] All responses were rated “difficult” to read, at about college-graduate level. [5]
What the audit does not show
The prompts were designed to test for misinformation, and the topics were not drug-label questions. The authors note that many real-world queries are not adversarial, so the design may overstate the rate of problematic content. [5] We use the audit for two points: confident tone with few caveats, and citation failure.
Does a Correct AI Answer Lead to a Correct Patient Decision?
What the trial measured
Bean and colleagues randomly assigned 1,298 UK adults to use GPT-4o, Llama 3, Command R+, or their usual resources, mostly internet search. Participants worked through ten physician-designed medical scenarios. [12] The models alone identified one or more relevant conditions in over 90% of cases and gave a correct disposition in roughly half. [12] Participants using the models identified relevant conditions in fewer than 35% of cases, significantly worse than the control group. [12] Disposition accuracy did not differ between model users and controls. [12]
What the trial does not show
The scenarios were general medical cases, not questions about marketed drugs. The trial shows a gap between what a model can say and what a lay user takes from the exchange. It means a clean answer in a monitoring log is necessary but not sufficient evidence that patients understood the drug correctly.
Why Do Published AI Audits Age Faster Than the Answers They Test?
Because journals publish months after testing, and models change in between. An audit tells you how a named model behaved on its test date. It does not tell you how a model behaves on the date your executive reads the briefing.
How Long Passes Between the Test Date and Publication?
We calculated minimum gaps for two of the studies. We use the earliest possible publication month within the stated year, so the real gaps are equal or longer.
Study
Test date
Publication
Minimum gap (our calculation)
Source
Andrikyan et al. (Bing Copilot)
April 2023
BMJ Quality & Safety, vol. 34, issue 2 (2025)
At least 21 months (April 2023 to January 2025)
[4]
Tiller et al. (five chatbots)
February 2025
BMJ Open (2026)
At least 11 months (February 2025 to January 2026)
[5]
ASHP 2023 posters (ChatGPT)
2023
Presented December 3 to 7, 2023
Test dates not stated in the source
[11]
The authors of the BMJ Open audit say that commercial AI changes rapidly. [5] We agree with that limit and draw one conclusion. A brand team cannot wait for peer review to learn what current models say. It needs its own dated records.
Can a Model’s Benchmark Score Predict Its Accuracy on Your Drug?
Not reliably. A 2025 Nature Medicine study by Alber and colleagues replaced 0.001% of training tokens in test models with medical misinformation. The resulting models were more likely to produce harmful content. In one 4-billion-parameter model, harmful content rose 4.8% at 0.001% poisoning. [13] The poisoned models scored about the same as clean models on five medical benchmarks. [13]
The study tested deliberate poisoning of training data, not accidental error in a commercial product. We draw a narrower point from it. Benchmark scores did not reveal the harm in that experiment, so a benchmark score is weak evidence about answers on your specific drug. Only a test on your drug against your label can speak to that. This is our interpretation, not a finding of the study.
What Happens When Model Versions Change Between Monitoring Periods?
The record changes, so the record must name the platform and the date. A finding that says “ChatGPT” without a model version and test date cannot be reproduced. The reproducibility fields appear in the methodology section below.
Where Can AI Answers Diverge From the Current Label? Two Documented Test Cases
We did not run AI tests for this article. We chose two cases because FDA published exact dates, which lets any reader check an answer against the record. Each case is a question to put to an AI system, not a result.
Testosterone: What Changed in the Boxed Warning on February 28, 2025?
FDA informed testosterone sponsors on February 28, 2025 that it was removing language about increased risk of adverse cardiovascular outcomes from the Boxed Warning for all testosterone products. [6] It also said it was adding TRAVERSE trial results to all testosterone labels and keeping the Limitation of Use language for age-related hypogonadism. [6] The FDA page states that TRAVERSE found no increase in the risk of adverse cardiovascular outcomes in men using testosterone for hypogonadism. [6]
The same notice added blood pressure information. FDA said it was adding product-specific blood pressure data for products with completed ambulatory blood pressure monitoring (ABPM) studies. It said it was adding a new blood pressure warning to products that lacked one. [6] The completed studies “confirmed an increase in blood pressure with use of all testosterone products, class-wide.” [6]
What should a briefing reader ask about testosterone?
Ask whether a current AI answer about testosterone safety still repeats the older cardiovascular Boxed Warning language. Ask whether it mentions the new blood pressure warning. The two changes point in different directions, so an answer can be wrong in either direction. A scheduled test would show which version each platform gives, and on what date.
Compounded Semaglutide and Tirzepatide: What Did the Shortage Resolutions Change?
Compounding status depends on dates, which makes it a hard question for any system that blends sources of different ages. FDA resolved the semaglutide shortage on February 21, 2025. [14] For tirzepatide, FDA made a first determination on October 2, 2024, re-evaluated, and issued a new resolution decision on December 19, 2024. [14] FDA then set enforcement-discretion deadlines for compounders. We calculated the windows from the published dates.
Product
Resolution decision
503A deadline
503B deadline
Window to 503A / 503B (our calculation)
Semaglutide
February 21, 2025
April 22, 2025
May 22, 2025
60 days / 90 days
Tirzepatide
December 19, 2024 (re-evaluated)
February 18, 2025
March 19, 2025
61 days / 90 days
Source: FDA [14]. FDA reported that a court denied the preliminary injunction on April 24, 2025 for semaglutide and on March 5, 2025 for tirzepatide, after which it said the 503A periods had ended. [14] Window lengths are our calculations.
Why is a status question harder than a dose question?
A dose stays fixed until the label changes. A status changes on named dates and followed court rulings. An answer can mix pre-resolution and post-resolution text and still read smoothly. We offer that as a hypothesis for testing. We have no data from this article showing that any platform does it.
Which Companies Do AI Engines Mention Most? What the 5W Index Shows
In the 5W Pharma/Rx AI Visibility Index, Eli Lilly held the highest estimated citation share at 12.5%, followed by Novo Nordisk at 11.5%. [7] The index measures visibility, not accuracy. It tells you who gets named. It does not tell you whether the answer is right.
What Did the 5W Index Measure?
5W tested ChatGPT, Claude, Perplexity, Gemini, and Google AI Overviews. It used more than 60 common patient and consumer prompts grouped into six sub-categories. It ran each prompt five times per engine in clean sessions during Q2 2026, and it ranked 25 pharmaceutical companies by estimated AI citation share. [7] 5W says the percentages are estimates based on mention frequency and are directional, not precise market measurements. [7]
By our arithmetic, that design produces at least 1,500 responses: 60 prompts x 5 runs x 5 engines. The index page also gave two publication dates, May 2026 and June 28, 2026, so we cite it only as a Q2 2026 measure. [7]
What Do the 5W Numbers Show When You Add Them Up?
Rank
Company
Estimated citation share
Cumulative share (our calculation)
1
Eli Lilly
12.5%
12.5%
2
Novo Nordisk
11.5%
24.0%
3
Pfizer
8.5%
32.5%
4
Johnson & Johnson
7.0%
39.5%
5
Merck
6.0%
45.5%
6
AbbVie
5.0%
50.5%
7
AstraZeneca
4.0%
54.5%
8
Moderna
3.5%
58.0%
9
Amgen
3.0%
61.0%
10
Novartis
2.6%
63.6%
11
Bristol Myers Squibb
2.3%
65.9%
12
GSK
2.0%
67.9%
13
Sanofi
1.8%
69.7%
14
Gilead Sciences
1.5%
71.2%
15
Roche
1.3%
72.5%
Source: 5W Pharma/Rx AI Visibility Index 2026 [7]. The index lists ranks 16 to 25 without percentages. Cumulative shares are our calculations.
Concentration in the top two, five, and fifteen
The top two companies hold 24.0% of estimated share. The top five hold 45.5%. The top fifteen hold 72.5%. These are sums of the published figures. 5W does not say the 25 shares total 100%, so we do not claim the ranks below 15 split the remainder.
The ratio between first and fifth place is 12.5 divided by 6.0, or about 2.1 (our calculation). Lilly appears about twice as often as Merck in this sample of prompts.
Does Direct-to-Consumer Ad Spend Predict AI Mentions?
5W’s founder argues that it does not. Ronn Torossian said in a release accompanying the report: “Pharma is the largest direct-to-consumer ad category in America. The AI engines are not impressed.” [15] Fierce Pharma, reporting on the index, cited iSpot.TV data on television spending and said Eli Lilly spends about $20 million per month on Zepbound. [15]
5W is a public relations firm, and its method is a prompt battery it chose. We treat the index as one dated measure of visibility and not as proof of cause. For a briefing, the useful question is narrower: does your brand appear in the answers a patient would trigger, and is the text beside it correct?
What Have FDA and EMA Said About AI Use?
FDA and EMA have said that AI used to generate evidence needs human oversight, a defined context of use, and lifecycle management. Neither agency has treated an AI-generated answer as evidence about a marketed drug. The two documents below show the direction of regulatory thinking.
What Do the FDA-EMA Guiding Principles Cover?
In January 2026, FDA’s CDER and CBER worked with EMA to publish 10 guiding principles for good AI practice in drug development. [16] The principles are:
Human-centric by design
Risk-based approach
Adherence to standards
Clear context of use
Multidisciplinary expertise
Data governance and documentation
Model design and development practices
Risk-based performance assessment
Life cycle management
Clear, essential information
Source: FDA [16].
RAPS reported that the document describes principles to inform the use of AI for generating evidence across all phases of the drug product life cycle. [16] Our reading is that the principles address AI that sponsors and developers use. They do not address what public chatbots say about marketed drugs. A briefing team cannot borrow regulatory comfort from them for AI answers it does not control.
What Does the Purolea Warning Letter Say About Over-Reliance on AI?
FDA issued a warning letter to Purolea Cosmetics Lab dated April 2, 2026 (reference 320-26-58), after an inspection on October 28 to 30, 2025. [17] The firm told investigators it used AI agents to help meet FDA requirements, including drug product specifications, procedures, and master production or control records. [17] FDA wrote that “Overreliance on artificial intelligence for your drug manufacturing operations was also documented.” [17]
The firm had not performed process validation before distributing products. It said it was unaware of the requirement because its AI agent never mentioned it. FDA cited 21 CFR 211.22(c) for failing to review AI-generated documents for accuracy and CGMP compliance, and 21 CFR 211.100 for process validation. [17] The letter says AI-generated documents “must be reviewed and cleared by an authorized human representative.” [17]
Why does a manufacturing letter matter to a briefing reader?
The letter shows a regulator holding a company responsible for an AI system’s omission. The AI never mentioned the validation requirement, and FDA placed the responsibility on the firm. Our interpretation: if an AI answer about your drug omits a warning, no regulator will treat the answer as authoritative. The omission is still a gap that someone at your company should know about.
A Four-Type Classification for AI Findings in an Executive Briefing
We propose four types: Drift, Gap, Phantom, and Share. This classification is our original framework, not an industry or regulatory standard. It sorts AI findings by the question an executive must ask and by the function that owns the answer. The table maps each type to published evidence of that failure pattern.
Type
Definition
Briefing question
Published evidence of the pattern
Likely owner
Drift
The answer contradicts the current label or regulatory status
The answer omits a warning, contraindication, dose limit, or caveat that the label contains
What did the answer leave out?
Zero-completeness and zero-accuracy answers [4]; few caveats in most answers [5]
Medical Information, Pharmacovigilance
Phantom
The answer makes an unsupported claim or cites a source that does not exist or does not support it
Can the cited source be found, and does it say that?
None of the references existed [11]; no fully accurate reference list [5]
Medical Affairs, Legal
Share
The drug, brand, or company is present, absent, or ranked against competitors
Are we named, and next to whom?
Citation-share estimates [7]
Brand, Competitive Intelligence
Type 1: Drift
Drift is time-based error. The answer may have been right once. A label change, a safety communication, or a status ruling made it wrong. Drift is the type most suited to a weekly briefing because it follows dated events that the evidence layer already tracks.
Type 2: Gap
Gap is omission. The statements in the answer may be true, yet the answer leaves out the warning that matters. Gap is the hardest type to catch by reading an answer alone. The reviewer must hold the label next to the answer and check each section.
Type 3: Phantom
Phantom is unsupported content. It includes a made-up reference, a real reference that does not say what the answer claims, and a claim with no source at all. The ASHP posters and the BMJ Open audit both reported reference failures. [5][11] The test is simple: open the cited source.
Type 4: Share
Share is presence and position. It does not depend on accuracy. A brand can be absent from an answer where it belongs, or named after three competitors. Share is the type that brand and competitive intelligence teams already ask about, and the 5W index measures it. [7]
How Should Severity Be Scored?
We propose three levels, also our own scale.
Level 1: informational. The answer differs from the label in a way that does not affect safe use, such as an outdated brand-name list.
Level 2: label-inconsistent. The answer contradicts a label statement on indication, dose, or status, and the contradiction could change a decision.
Level 3: safety-relevant. The answer omits or reverses a Boxed Warning, contraindication, or serious interaction. Route it the same day to Medical Information and Pharmacovigilance.
How Do You Build a One-Minute AI Block for the Executive Briefing?
Record the test conditions, compare each answer to the current label, classify each discrepancy, and report patterns in about 238 words. The block summarizes findings. It does not reproduce answers.
Which Fields Make an AI Test Reproducible?
The fields below follow the problems we found in published studies. For example, Andrikyan et al. documented the mode and VPN location, and 5W documented the number of runs and clean sessions. [4][7] A researcher who reads your record should be able to repeat the test.
Field
Why it matters
Platform and model version
Models change; a finding without a version cannot be repeated
Exact prompt text
Open-ended and closed prompts gave different results in the BMJ Open audit [5]
Date and time of each run
Needed to match the answer to the label in force that day
Number of runs per prompt
5W ran each prompt five times to see variation [7]
Drug, indication, and geography
Labels and status differ by country
Web search on or off; citations requested or not
Changes the sources the system may draw on
Source set and label version used for checking
States which text the answer was compared to
Evaluation criteria and classification (Drift, Gap, Phantom, Share)
Makes classification consistent between reviewers
Severity level (1 to 3)
Controls routing and escalation
Limitations
States what the sample cannot show
What Does the 238-Word Block Look Like?
The template below uses bracketed placeholders. It contains no real data. Each line has a fixed job, so a reader can find it in seconds.
Coverage: [number of prompts] prompts, [number of platforms] platforms, [number of runs], dates [start to end]. Label version checked: [date].
Changes since last period: [number] new Level 2 or Level 3 findings; [number] resolved.
Drift: [drug] answers on [platform] state [claim]. Current label states [claim]. Last label change: [date].
Gap and Phantom: [number] answers omitted [section]; [number] cited sources could not be located.
Share: named in [percent] of runs for [query group]; competitors named first in [percent].
Action and owner: [function] to [action] by [date].
Line 1: Coverage
Coverage tells the reader how much to trust the rest. It states the sample and the label version. Without it, every later number is unanchored.
Lines 3 and 4: Drift, Gap, and Phantom
These lines carry the findings that need a decision. Each cites the label section and the answer date. Level 3 findings go to the top of the briefing, not into this block, because they need action the same day.
Line 6: Action and Owner
Every block ends with an owner. A finding with no owner becomes a recurring item that no one resolves. Our classification table suggests default owners.
How Should Repeated Runs and Duplicate Responses Be Handled?
Run each prompt several times and report the pattern, not a single answer. 5W used five runs per prompt per engine in clean sessions. [7] We recommend the same. Count a finding when it appears in a majority of runs. Report a finding that appears in a minority of runs as “intermittent.” Keep the full text of every run in the record so a reviewer can check the count. This is our recommendation, not a published standard.
What Does This Mean for Each Function?
Each function reads a different line of the block. The sections below say which line each function should own.
What This Means for Medical Affairs
Medical Affairs owns Phantom findings about clinical evidence. Physicians use AI to summarize research and standards of care. [8] If an answer misstates a pivotal trial or cites a paper that does not exist, Medical Affairs should know. It can then check whether its own published materials are clear and easy to find.
What This Means for Pharmacovigilance
Pharmacovigilance owns Level 3 Gap findings. An answer that omits a Boxed Warning is a communication risk, even though it is not an adverse event report. In our reading, a company’s reporting duties attach to adverse events it receives. An inaccurate chatbot answer is not an adverse event by itself. If a patient report links harm to the product and mentions an AI answer, normal case processing applies.
What This Means for Regulatory Affairs
Regulatory Affairs owns Drift. It already tracks label changes and approval actions. It is the function best placed to say which label text is current on which date. It should supply the label version for each test cycle.
What This Means for Brand Teams
Brand teams own Share. The Pew data show fewer clicks to source pages when an AI summary appears. [9] The 5W data show that engines name some companies far more often than others. [7] A brand team should know where its product appears, and whether the surrounding text matches approved claims.
What This Means for Competitive Intelligence
Competitive intelligence can use Share data to see which competitors are named beside your drug, and for which query groups. It should treat the data as a dated sample. The 5W figures are directional estimates by the publisher’s own description. [7]
What This Means for Legal and Corporate Communications
Legal needs the full record of any Level 3 finding, including the prompt, platform, and date. The Purolea letter shows that FDA expects a company to review what an AI system produces. [17] Corporate Communications needs to know which findings it can correct through its own published content and which need a platform feedback channel. We found no source that shows a reliable method for changing an AI system’s answer, so we do not claim one.
Where Does DrugChatter Fit?
DrugChatter is one way to produce the record this article describes. It is a method for collecting dated answers. It is not a source of evidence about drugs.
How Does DrugChatter AI Monitoring Produce Briefing Inputs?
DrugChatter AI Monitoring runs a library of natural-language questions about a brand across several AI vendors on a schedule. It stores each question, each answer, and the vendor that produced it. It then compares results week over week and across vendors. The same dataset can feed separate briefings for executive, marketing, pharmacovigilance, medical affairs, regulatory affairs, and medical information readers. You can read more at drugchatter.com.
We do not report DrugChatter test results in this article. Any monitoring tool, including ours, produces answers that still need review against the prescribing information by a qualified person. The FDA’s Purolea letter makes the same point about AI output in general. [17]
What Are the Limits of This Analysis?
This article contains no original AI test data. Every statistic comes from a cited source, or is our calculation from cited figures, and we label it as such.
The studies we cite tested different platforms, dates, and question types. They do not form a single dataset. We did not pool them.
The usage figures measure health use in general. We found no source that counts drug-specific questions to AI systems.
The Drift test cases (testosterone and GLP-1 compounding) are questions we propose. We did not test any AI system on them.
The Drift, Gap, Phantom, and Share classification and the three-level severity scale are our proposals. They have not been validated by an outside group.
Key Takeaways
A ten-minute briefing holds about 2,380 words at 238 words per minute, so a one-minute AI block holds about 238 words. [2]
OpenAI reports more than 40 million daily health users, and KFF found that 32% of U.S. adults used AI chatbots for health information in the past year. [1][3]
Published audits found inaccurate or problematic answers in 26.0% of Copilot answers about common drugs and 50% of responses in a five-chatbot audit. [4][5]
Published audits report test dates months before publication, so they cannot describe current model behavior. [4][5]
FDA’s testosterone labeling change of February 28, 2025 and the GLP-1 compounding deadlines give dated test cases for Drift. [6][14]
5W’s index puts Eli Lilly and Novo Nordisk at 24.0% of estimated citation share combined. This is a visibility measure and not an accuracy measure. [7]
FDA’s Purolea letter shows that a regulator holds the company responsible for reviewing AI output. [17]
Drift, Gap, Phantom, and Share are our proposed types for classifying AI findings, with a three-level severity scale.
FAQ
Can an executive cite an AI answer as evidence in a briefing?
No. Cite the label, the regulatory record, or the publication. The BMJ Open audit found that no chatbot produced a fully accurate reference list. [5] Use the AI answer only as a specimen to compare against the label.
How often should AI answers about a drug be re-tested?
We recommend weekly for active brands and immediately after any label change. FDA published exact dates for the testosterone and GLP-1 changes, and answers can move when those dates pass. [6][14] The 5W index measured one quarter, which is too coarse for label-change tracking. [7]
What is the difference between an AI visibility score and label alignment?
Visibility measures whether a drug or company is named. Label alignment measures whether the answer matches the prescribing information. The 5W index measures the first and states that its shares are directional estimates. [7] In our classification, visibility is Share and alignment covers Drift and Gap.
Does turning on web search fix outdated or wrong answers?
No source we reviewed isolates that effect. The Copilot study used a chatbot built into a search engine, and 26.0% of its rated answers still scored below 100% accuracy. [4] Record whether web search was on, and test both settings if the platform allows it.
Which function should own AI findings?
Assign by finding type. Regulatory Affairs and Medical Information own Drift. Medical Information and Pharmacovigilance own Gap. Medical Affairs and Legal own Phantom. Brand and Competitive Intelligence own Share. Send any Level 3 finding to Medical Information and Pharmacovigilance the same day.
Is an inaccurate chatbot answer a reportable adverse event?
In our reading, no. A reportable adverse event arises from a report of harm linked to a product. A wrong answer alone is a communication problem. If a patient or clinician reports harm and mentions the answer, process the report under your normal PV procedure and keep the AI record with the case file.
How big should a monitoring sample be?
Published studies give a range. The Copilot study used 500 answers. [4] The BMJ Open audit used 250 prompts. [5] The 5W design implies at least 1,500 responses (60 prompts x 5 runs x 5 engines, our calculation). [7] Size depends on how many drugs and question types you cover. Report the sample size in every briefing block.
Why does physician AI use matter for a brand team?
Physicians use AI at work: 81% in the AMA’s 2026 survey, and 39% for summarizing research and standards of care. [8] Test how AI systems summarize your pivotal trial and your drug’s place in guidelines, as well as patient-facing questions.
What does the Nature Medicine trial mean for monitoring programs?
It means answer text is only half the picture. Models alone identified relevant conditions in over 90% of cases, but participants using them did so in fewer than 35%. [12] Monitoring the answer is necessary. Complement it with medical information inquiries and patient support data to see what people understood.
Can a company change what an AI system says about its drug?
No source we reviewed shows a reliable method. The 5W index tracked the source types engines drew on, including editorial, reference, news, and brand-owned content. [7] A company can keep its label, medical information pages, and published evidence accurate and easy to find, and it can document errors for platform feedback. The effect of those steps on any single answer is not established.
References
Axios. (2026, January 5). Exclusive: 40 million people turn to ChatGPT for health care. https://axios.com/2026/01/05/chatgpt-openai-health-insurance-aca
Brysbaert, M. (2019). How many words do we read per minute? A review and meta-analysis of reading rate. Journal of Memory and Language, 109, 104047. https://doi.org/10.1016/j.jml.2019.104047
KFF. (2026, March 25). Poll: 1 in 3 adults are turning to AI chatbots for health information, equaling the share who use social media for health. https://www.kff.org/health-information-trust/poll-1-in-3-adults-are-turning-to-ai-chatbots-for-health-information-equaling-the-share-who-use-social-media-for-health/
Andrikyan, W., et al. (2025). Artificial intelligence-powered chatbots in search engines: a cross-sectional study on the quality and risks of drug information for patients. BMJ Quality & Safety, 34(2), 100. https://qualitysafety.bmj.com/content/34/2/100
Tiller, N. B., et al. (2026). Generative artificial intelligence-driven chatbots and medical misinformation: an accuracy, referencing and readability audit. BMJ Open. https://doi.org/10.1136/bmjopen-2025-112695. Summary read: News-Medical. (2026, April 15). Popular chatbots often provide inaccurate and incomplete medical information. https://www.news-medical.net/news/20260415/Popular-chatbots-often-provide-inaccurate-and-incomplete-medical-information.aspx
U.S. Food and Drug Administration. (2025, February 28). FDA issues class-wide labeling changes for testosterone products. https://www.fda.gov/drugs/drug-alerts-and-statements/fda-issues-class-wide-labeling-changes-testosterone-products
5W Public Relations. (2026). Pharma/Rx AI Visibility Index 2026. https://www.5wpr.com/ai-visibility-index/pharma-rx-ai-visibility-index-2026/
American Medical Association. (2026, March 12). More than 80% of physicians use AI professionally: AMA survey. https://ama-assn.org/practice-management/digital-health/more-80-physicians-use-ai-professionally-ama-survey
Pew Research Center. (2025, July 22). Google users are less likely to click on links when an AI summary appears in the results. https://pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results
ETV Bharat. (2026, January 13). Google pulls AI Overviews for some health queries after accuracy concerns (reporting on a Guardian investigation published January 2, 2026). https://www.etvbharat.com/en/technology/google-pulls-ai-overviews-for-some-health-queries-after-accuracy-concerns-enn26011302897
Drug Topics. ChatGPT provides inaccurate responses to questions about drug information (reporting on ASHP 2023 Midyear Clinical Meeting posters 8-021 and 8-023, December 3 to 7, 2023). https://www.drugtopics.com/view/chatgpt-provides-inaccurate-responses-to-questions-about-drug-information
Bean, A. M., et al. (2026). Reliability of LLMs as medical assistants for the general public: a randomized preregistered study. Nature Medicine. https://doi.org/10.1038/s41591-025-04074-y. Summary read: News-Medical. (2026, February 10). https://www.news-medical.net/news/20260210/Large-language-models-excel-in-tests-yet-struggle-to-guide-real-patient-decisions.aspx
Alber, D. A., et al. (2025). Medical large language models are vulnerable to data-poisoning attacks. Nature Medicine, 31(2), 618-626. https://doi.org/10.1038/s41591-024-03445-1
U.S. Food and Drug Administration. (2025). FDA clarifies policies for compounders as national GLP-1 supply begins to stabilize. https://www.fda.gov/drugs/drug-alerts-and-statements/fda-clarifies-policies-compounders-national-glp-1-supply-begins-stabilize
Fierce Pharma. Eli Lilly, Novo Nordisk top AI citation share as new report questions DTC spend culture. https://www.fiercepharma.com/marketing/eli-lilly-novo-nordisk-top-ai-citation-share-new-report-questions-dtc-spend-culture
U.S. Food and Drug Administration. (2026, January). Guiding principles of good AI practice in drug development. https://www.fda.gov/about-fda/artificial-intelligence-drug-development/guiding-principles-good-ai-practice-drug-development. Context: RAPS. (2026, January 14). EMA, FDA issue joint AI guiding principles for drug development. https://www.raps.org/resource/ema-fda-issue-joint-ai-guiding-principles-for-drug.html
U.S. Food and Drug Administration. (2026, April 2). Purolea Cosmetics Lab, warning letter 320-26-58. https://www.fda.gov/inspections-compliance-enforcement-and-criminal-investigations/warning-letters/purolea-cosmetics-lab-722591-04022026
