More than one in four of OpenAI’s 800 million regular ChatGPT users send a health prompt each week, according to a January 2026 OpenAI report as relayed by Becker’s Hospital Review [1]. Multiplying those two figures gives at least 200 million people a week. That product is our arithmetic, not an OpenAI statistic. Each of those prompts is a question a real person typed, unprompted by any drug company, about a symptom, a dose, a cost, or a medicine they already take.
Pharmaceutical teams already know how to read three kinds of real-world signal: spontaneous adverse event reports, claims and records data, and (more recently) search logs. A fourth kind now exists. It is the stream of natural-language questions that people put to large language models. This article explains what that stream is, what the published evidence says about its size and about the quality of the answers people get back, and how Medical Information, Pharmacovigilance, Medical Affairs, Regulatory and brand teams can use it.
It also covers a data-supply question. DrugChatter maintains a database of real questions that users ask LLMs about drugs. DrugChatter’s position is that it is the only database of its kind for pharmaceutical questions. We state that as DrugChatter’s position, not as a verified market fact, and the section on data sources explains what we could and could not confirm.
The Short Answer
An AI drug question is a real-world intelligence signal because a person with an actual need wrote it, in their own words, without a drug company prompting them. The signal has value before anyone judges the answer. The question shows what people want to know, which drugs and combinations they worry about, and where they cannot get an answer elsewhere.
Scale: one in four regular ChatGPT users submits a health prompt each week, and more than 5% of all ChatGPT messages relate to health [1].
Reach: 32% of US adults used AI chatbots for health information in the past year, per KFF’s February to March 2026 poll [3].
Answer quality is uneven: in three published evaluations, between 26% and 74% of answers fell short of the reference standard, depending on the study design [5][6][7].
Existing safety data is thin: a systematic review put the median under-reporting of adverse drug reactions at 94% [9].
Question data is already traded: Profound sells prompt volumes built from consumer panels, but across all topics, not drugs [11].
What Counts as Real-World Drug Intelligence Today?
Real-world drug intelligence is information about how people use, understand and experience a medicine outside a controlled trial. Before AI chat, the main sources were spontaneous reports, claims and records, and search behavior. Each has a known blind spot.
Why Do Spontaneous Adverse Event Reports Capture So Little?
Spontaneous reporting systems capture only a small fraction of events. Hazell and Shakir reviewed 37 studies, produced 43 numerical estimates, and reported a median under-reporting rate of 94% (interquartile range 82% to 98%) [9].
Analysis: a 94% median means about 6 of every 100 events reach the system, or roughly 1 in 17. At the interquartile bounds, the reported share runs from 2% to 18%. This is our calculation from the published median, not a figure the authors state. It shows why safety teams look for other signals.
What Did Search Logs Show About Drug Interactions?
Search logs showed a drug interaction signal before the interaction was widely reported. White, Tatonetti, Shah, Altman and Horvitz published the study in the Journal of the American Medical Informatics Association in 2013 [8].
The study design
The authors analyzed logs from consenting users who had installed a browser add-on. The logs covered Google, Bing and Yahoo searches over all of 2010. The dataset held 82 million drug, symptom and condition queries from 6 million searchers [8].
The pravastatin and paroxetine result
Users who searched for both pravastatin and paroxetine were more likely to search for hyperglycemia terms. About 10% of the pair group did so, against about 5% of paroxetine-only searchers, about 4% of pravastatin-only searchers, and about 0.3% of all users [8]. The authors note that the interaction was reported in 2011, after the 2010 logs were collected [8].
What the study did not prove
The authors tested 31 known interacting pairs and 31 unrelated pairs. The area under the curve was 0.8189 with the full term list and 0.7429 with a focused list. The authors also note that a single pair could reflect an unmodeled mechanism or chance [8]. The method found candidates for review. It did not confirm causes.
The study matters here for one reason. It shows that unprompted consumer text can carry pharmacovigilance-relevant structure, and that researchers can test it against known interactions. An AI question stream is a richer version of the same kind of data.
How Many People Ask AI About Health and Drugs?
Two independent sources give the scale: OpenAI’s own usage analysis and KFF’s national polling. They measure different things, so we report them separately.
What Did OpenAI Report About Health Questions in January 2026?
OpenAI published “AI as a Healthcare Ally: How Americans Are Navigating the System With ChatGPT” in January 2026. We could read only trade-press coverage of the report, so the figures below are attributed to those articles [1][2].
Volume
Becker’s reports that more than 5% of all ChatGPT messages globally relate to healthcare, and that one in four of more than 800 million regular users submits a health prompt each week [1]. Becker’s also reports that more than 40 million Americans use ChatGPT daily for health questions [1]. Other outlets describe the same 40 million figure as global, so readers who quote it should check the original report.
Timing and topic
Seven in ten health conversations happen outside normal clinic hours [1][2]. US users send nearly 1.6 million to 1.9 million messages a week about health insurance, including plan comparisons, claims and cost sharing [1]. Managed Healthcare Executive rounds the insurance figure to nearly two million [2].
Limits of the report
Managed Healthcare Executive states that the report combines anonymized message data with survey research, that the survey data is self-reported, and that the report covers only ChatGPT [2]. The report is a company’s description of its own platform. It says nothing about the share of health messages that name a specific drug.
What Did KFF Find About AI Chatbot Use for Health Information?
“32% of adults used AI chatbots for health information in the past year” (KFF poll of 1,343 US adults, fieldwork February 24 to March 2, 2026, margin of error plus or minus 3 percentage points) [3]
Use, reasons and follow-up
KFF reports that 29% of adults sought physical health information from AI and 16% sought mental health information [3]. Among users, 65% wanted quick information, 41% wanted to research before deciding whether to see a provider, and 36% preferred to look things up privately [3]. Of those who asked about physical health, 42% did not follow up with a doctor or other professional [3]. KFF also reports that 41% of AI health users uploaded personal medical information, which is 13% of all adults [3].
Why the two KFF figures do not compare
KFF’s 2024 poll found that 17% of adults used AI chatbots at least once a month for health information and advice, with 25% among adults aged 18 to 29 [4]. The 2026 poll measures use in the past year [3]. A monthly-use rate and a past-year rate are different measures, so the two numbers do not show a growth rate. Treat both as separate snapshots.
The 2024 poll also measured trust. Twenty-nine percent of adults said they trusted chatbots a great deal or a fair amount for reliable health information [4]. That gap, wide use alongside modest trust, is typical of a verification need. People ask, and then some of them check.
Why Is an AI Question Different From a Search Query?
An AI question is usually longer, more specific and more personal than a search query. This section is interpretation, not a measured result, and we label it that way.
A question carries context that a keyword does not
A search query such as a drug name plus a symptom term compresses the need. A chat prompt can state the dose, the other medicines, the timing, the reason for taking the drug, and the worry. KFF found that 41% of AI health users have uploaded personal medical information, which means some prompts contain more patient detail than a query ever did [3].
Why Does the Phrasing of a Drug Question Matter?
The phrasing shows the reader’s frame. “Is it safe to take X with Y” and “I started X and now I feel Y” are two different needs. The first is an interaction question for Medical Information. The second may describe an experience that a safety team would want to see. Keyword counts erase that difference. A question dataset keeps it.
Which Part Is the Signal: The Question or the Answer?
The question is the signal about people. The answer is the signal about the AI system. Teams that monitor only answers learn what a model says. Teams that read questions also learn what people want a model to say. Both matter, and they need different handling.
What Do AI Systems Say Back? Evidence on Answer Accuracy
Published evaluations of AI drug answers exist, and they differ in design, model and date. None describes the current state of any system. We summarize three and then compare them.
What Did the Andrikyan Study Find About Bing Copilot and Drug Information?
Andrikyan and colleagues published a cross-sectional study in BMJ Quality and Safety (2025;34:100-109), first online on 1 October 2024 [5].
Design and accuracy results
The team asked Bing Copilot 10 patient questions about each of 50 drugs, the most frequently prescribed in the US outpatient market in 2020. That produced 500 answers, generated in April 2023 [5]. Mean accuracy against the reference database was 88.7%, and mean completeness was 76.7% [5]. Of 484 assessable answers, 126 (26.0%) had accuracy below 100%, and 16 (3.3%) had accuracy of 0% [5].
Harm assessment
In a 20-answer expert subset, 66% were judged potentially harmful and 22% potentially leading to death or severe harm [5]. The subset was chosen for low accuracy, low completeness or safety risk, so the percentages do not describe all 500 answers [5].
What Did the Long Island University Study Find About ChatGPT and Drug Questions?
Sara Grossman, PharmD, of Long Island University led a study presented in 2023. Forty-five drug questions were posed to pharmacists, and a lack of literature narrowed the set to 39 for ChatGPT [7]. Ten of the 39 responses (25.6%, our calculation) were satisfactory. The other 29 included 11 non-answers, 10 false responses and 12 incomplete responses, with some in more than one category [7]. Only eight responses supplied the references the researchers asked for, and all eight included references that did not exist [7].
What Did the 2026 BMJ Open Audit Find?
A BMJ Open audit (2026;16(4):e112695) tested five chatbots on 50 prompts in five topic areas: cancer, vaccines, stem cells, nutrition and athletic performance [6]. Prompts were entered on 1 February 2025. The chatbots were Gemini 2.0, DeepSeek V3, Meta AI (Llama 3.3), ChatGPT 3.5 and Grok 2 [6].
Overall result
Across 250 responses, 49.6% were problematic: 30% somewhat problematic and 19.6% highly problematic [6]. The differences among chatbots were not statistically significant overall (p=0.566) [6].
Citation result
The chatbots returned 1,013 of 1,250 requested references, about 81%. Median reference completeness was 40%, and no chatbot produced a fully complete and accurate reference list for any prompt [6].
How Do the Three Evaluations Compare?
Study
System tested
Test date
Questions and answers
Headline result
Source
Andrikyan et al., BMJ Qual Saf
Bing Copilot
April 2023
10 questions x 50 drugs = 500 answers
126 of 484 assessable answers (26.0%) below 100% accuracy
Analysis: the three headline percentages measure different things. One counts answers below perfect accuracy against a reference database. One counts answers rated satisfactory by pharmacists. One counts answers rated problematic across misinformation-prone topics that are not all drug questions. Do not average them. Read each against its own design.
Every study tested a model that has since changed. The numbers describe the systems on the test dates and nothing more. They show that answer quality can vary and that answers need checking against current prescribing information. They do not show how any current model performs.
The Six Types of Drug Questions People Ask AI: An Original Taxonomy
Original framework: we propose six question types that a team can use to sort a drug question dataset. This is not an industry standard. We built it from the kinds of drug questions that appear in the evaluations above and in the KFF and OpenAI data, and a team should test it against its own data.
What Are the Six Types?
Type
What the person asks
Example form
Usual owner
1. Label lookup
Dose, timing, food, storage, listed side effects
“How should I take [drug]?”
Medical Information
2. Interaction and combination
Whether two products can be used together
“Can I take [drug A] with [drug B]?”
Medical Information, Pharmacovigilance
3. Personal experience
A symptom after starting or changing a drug
“I started [drug] and now I feel…”
Pharmacovigilance
4. Unapproved use
A use not in the approved indication
“Can [drug] help with [condition]?”
Medical Affairs, Regulatory
5. Access and cost
Coverage, price, switching, availability
“Will my plan cover [drug]?”
Market Access, Patient Support
6. Verification and comparison
Whether a claim is true, or how two products differ
“Is the generic the same as [brand]?”
Brand, Competitive Intelligence
OpenAI’s report supports one slice of type 5 directly: US users send nearly 1.6 million to 1.9 million insurance messages a week [1]. The report does not break out the other five types for drugs.
Which Function Should Read Which Question Type?
Types 1 and 2 are Medical Information questions. The team already answers them one at a time. A question dataset shows which ones recur. Type 3 is the closest to safety. Whether any particular message meets a company’s case-intake criteria is a decision for its safety team and counsel, and this article does not make that decision. Type 4 intersects with FDA policy on unsolicited requests, discussed below. Types 5 and 6 belong to commercial teams.
Where Do AI Questions Sit Among Drug Intelligence Sources?
An AI question dataset adds a different angle. It does not replace the other sources. The table is analysis. Where a cell rests on a cited study, the source column says so.
How Are AI Questions Different From Spontaneous Reports?
A spontaneous report follows an event and needs someone to decide to file it. A question can come before any event and needs no decision to report. The two answer different questions. Reports say what happened to some people. Questions say what many people worry about or want to do.
How Are AI Questions Different From Search Logs?
The White study used 82 million queries from 6 million searchers and found a statistical association that held up against known interactions [8]. A question database has fewer rows per person but more text per row. The same design, comparing co-mentioned drugs and symptoms against a background rate, can in principle run on question text. Nobody has published that test on AI questions that we could find, so we call it an open research question.
Why Does Real Phrasing Matter More Than Test Prompts?
Output-based AI visibility indices run fixed prompt batteries written by analysts. That method measures how models answer the questions an analyst chose. It cannot show which questions people ask. The FDA’s draft guidance on unsolicited requests rests on a similar distinction: a request prompted by the manufacturer is treated differently from one that comes from an independent person [10]. A question that no company wrote is a different kind of evidence from a question a company wrote for testing.
Who Has Real Questions? Data Sources for AI Drug Questions
What Is an AI Question Database?
An AI question database is a collection of natural-language questions that real users have put to large language models, stored with enough metadata (date, platform, topic) to count and classify them. It differs from a set of prompts that an analyst wrote to test a model.
What Does Profound Sell, and Is It Pharma-Specific?
Profound sells Prompt Volumes, which it describes as drawn from licensed conversations from double opt-in consumer panels of AI assistant users [11]. The page cites “hundreds of millions of prompts per month from millions of active users,” a figure of “1.5 billion+ real AI conversations,” and coverage of ChatGPT, Gemini, Claude and Perplexity [11]. The page says the data is both direct user data and statistically modeled to correct for demographic and geographic bias [11]. The product covers all topics. We found nothing on the page that describes a pharma-specific product or a comparison of answers against prescribing information [11].
What Does DrugChatter’s Question Database Add?
DrugChatter maintains a database of real questions that users ask LLMs about drugs. DrugChatter says this is the only database of actual user questions to LLMs built for pharmaceutical use. That claim is DrugChatter’s, and the rest of this section describes what the database is for rather than rating the claim.
What the database contains, according to DrugChatter
The database holds the questions people ask about drugs, in the words they used. DrugChatter’s AI monitoring product, chatmon, takes questions from this kind of library and re-runs them across LLM vendors on a schedule. It then compares the answers with the drug’s FDA-approved labeling. We have not independently audited the size or collection method of the database, and this article does not report any count from it.
Why a drug-specific question set matters
A general prompt-volume product answers the question “how often is this prompt asked.” A drug-specific set can answer different questions: which interaction questions recur for a given active ingredient, which unapproved uses people raise, and which label statements models restate wrongly. Those questions belong to Medical Information, Pharmacovigilance, Medical Affairs and Regulatory teams, and a general marketing tool is not built to answer them.
How question data pairs with label checks
A question without a check is a count. A question run against several models and compared with the current prescribing information becomes a measurement. DrugChatter’s weekly briefings combine both: they compare answers across LLM vendors and over time, and they split the findings into views for different functions such as executive, pharmacovigilance, medical affairs, regulatory affairs and medical information teams.
Is DrugChatter’s Database the Only One of Its Kind?
DrugChatter says yes. We could not verify that. We confirmed that at least one other vendor sells real-user prompt data from panels across all topics [11]. We did not find a published pharma-specific equivalent in the sources we checked, but absence from our search is not proof that none exists. A reader who needs to cite the claim should attribute it to DrugChatter.
What Does FDA Policy Say About Unsolicited Questions?
How Does FDA Define an Unsolicited Request?
FDA’s December 2011 draft guidance, “Responding to Unsolicited Requests for Off-Label Information About Prescription Drugs and Medical Devices,” states: “Unsolicited requests are those initiated by persons or entities that are completely independent of the relevant firm.” It also states that “requests that are prompted in any way by a manufacturer or its representatives are not unsolicited requests” [10]. The FDA page lists the document as a draft that contains non-binding recommendations [10].
How Should a Firm Answer a Public Question?
For public forums and social media, the guidance says a firm should respond only when the question pertains specifically to its own named product. The public response should be limited to contact information and should not include off-label information. The representative should disclose affiliation and direct the requester to medical or scientific personnel [10]. For private requests, the draft says information should go only to the individual requester, come from personnel independent of sales and marketing, and include labeling and safety information [10].
Does an AI Chat Question Fit This Framework?
Analysis: the guidance addresses requests made to a firm or in a forum where a firm could reply. A question typed into a chatbot is neither. The guidance does not mention chatbots. This article does not claim that FDA treats a chatbot question as an unsolicited request. We make a narrower point: the guidance shows that regulators already separate independent questions from company-prompted ones, and the same separation is useful when a team labels its own data.
What is not settled
We found no FDA guidance that addresses how a company should handle a dataset of third-party questions to AI systems, or what duties follow from reading one. Legal and Regulatory teams should decide that before a dataset is circulated.
Scenario Analysis: What a Question Stream Could Show
This scenario is hypothetical. It uses a real drug pair from the published literature to show a method, and it reports no data.
Scenario: Rising Questions About Two Drugs Taken Together
Suppose a team’s question dataset shows that questions pairing pravastatin and paroxetine, the pair from the 2013 search-log study [8], rose over two months. The team would take four steps. It would classify the questions by type (the interaction type, and any experience-type questions that mention high blood sugar). It would check what the current prescribing information for each product says. It would run the same questions across several models and record whether the answers match the labels. It would pass the experience-type questions to the safety team under its own intake rules.
The question volume alone would prove nothing about harm. The White study itself warns that one pair can reflect chance [8]. The value of the dataset is that it tells the team where to look first.
Scenario: Models Disagree With the Label
Suppose three models answer a dosing question and one states an interval that does not match the label. The Andrikyan study found 3.3% of assessable answers with accuracy of 0% against the reference [5], so a mismatch is not unusual in published testing. The Medical Information team would document the discrepancy, record the model, the date and the prompt, and decide whether to publish a clarification on its own site.
What This Means for Each Function
What This Means for Medical Information
Medical Information teams answer incoming questions one at a time. A question dataset shows which questions recur before they reach the team. The first use is a content gap list: questions that people ask AI and that the team’s public resources do not answer.
What This Means for Pharmacovigilance
Pharmacovigilance teams work with a system that, by the published median, misses most events [9]. Question text is not a case report and should not be treated as one by default. It is a source for hypothesis generation, in the way the search-log study used consumer text [8]. Teams should set a written rule for when a question dataset is reviewed and by whom.
What This Means for Medical Affairs
Medical Affairs sees questions from clinicians at congresses and through field medical teams. AI questions reach a different population: 32% of US adults used AI chatbots for health information in the past year [3]. Type 4 questions, about unapproved uses, show where patient interest runs ahead of the label. Handle them within the firm’s off-label policy and the FDA framework above [10].
What This Means for Regulatory Affairs
Regulatory teams own the reference: the current approved labeling. Every comparison of an AI answer against the label requires an exact label version and date. If the label changed between two test dates, an answer that was right in one month can be wrong in the next. Record the version with every comparison.
What This Means for Brand and Competitive Intelligence Teams
Type 6 questions show how people compare products and test claims. Brand teams can see which comparisons come up and whether models restate them accurately. They should keep this separate from promotion. A question dataset describes demand for information. It does not authorize any claim.
How to Analyze an AI Drug Question Dataset: A Reproducible Method
Any team, with any data source, can follow this method. It lists the fields that another researcher would need to reproduce the work.
What Should You Record for Every Question?
Platform and model version, and the date and time of each test.
The exact question text, and the drug and indication it concerns.
Geographic context, and whether web search was enabled.
Whether the question was real or analyst-written, and how the real questions were collected.
The label or guideline version used for comparison, with its date.
How duplicate questions and duplicate answers were handled.
How Should You Classify Discrepancies?
Original framework: we suggest three severity levels. Level 1 is an omission, where the answer leaves out a labeled fact. Level 2 is a contradiction, where the answer conflicts with the label. Level 3 is a safety-direction contradiction, where the answer conflicts with the label in a way that could lead a user to take more, take it with a contraindicated product, or skip a labeled warning. Two reviewers should classify each discrepancy, and the team should record how they resolved disagreements. The Andrikyan team used a different harm-assessment method with expert review [5], so the levels here do not map directly onto their results.
How Do You Handle Unsupported Claims and Fake References?
Check every reference the model gives. In the Long Island University study, all eight responses that supplied references included nonexistent ones [7]. In the BMJ Open audit, median reference completeness was 40% [6]. A citation in an AI answer is not evidence until a person has opened it.
What Are the Limits of This Method?
Sample bias
A question dataset reflects the people and platforms that produced it. Profound states that it models its panel data to correct for demographic and geographic bias [11]. Any database should say how it handles the same problem.
Questions are not events
A person who asks about a symptom may not have taken the drug. A question about an interaction may come from curiosity.
Models change
An answer tested in one month may differ the next. Scheduled repeat testing addresses this, and it is why a single snapshot study cannot stand in for monitoring.
Privacy
KFF found that 77% of adults are concerned about the privacy of medical information shared with AI tools [3]. Any question dataset needs a documented privacy basis, de-identification and legal review before use.
Open Questions and What Happens Next
Four questions remain open. None of the published studies we found tests whether AI question text can detect drug interactions the way search logs did [8]. OpenAI has not, in the coverage we read, published a breakdown of health questions by drug [1][2]. The published accuracy studies tested models that have since changed [5][6][7]. FDA’s guidance on unsolicited requests has been in draft status since 2011, and it does not address AI systems [10].
We expect these gaps to close through published research, not through vendor claims. A team that records its method in the form above will be able to compare its results with whatever appears.
Key Takeaways
One in four of more than 800 million regular ChatGPT users submits a health prompt each week, and more than 5% of all ChatGPT messages relate to health, per OpenAI’s January 2026 report as relayed by Becker’s [1].
32% of US adults used AI chatbots for health information in the past year, and 42% of those who asked about physical health did not follow up with a professional [3].
Published tests found a substantial share of AI drug answers falling short of reference standards: 26.0% of assessable answers below 100% accuracy in one study, and 10 of 39 satisfactory in another [5][7].
Adverse event reporting is a partial view: the median under-reporting estimate is 94% [9].
Consumer text can carry drug-safety structure. A 2013 search-log study found a pravastatin and paroxetine hyperglycemia signal with an AUC of 0.8189 across known pairs [8].
Real user questions are a separate evidence type from analyst-written test prompts. Profound sells general-topic panel data [11]. DrugChatter says its database of real user questions to LLMs is the only one of its kind built for drugs, and that is DrugChatter’s own position.
FAQ
What is real-world intelligence from AI questions?
It is information drawn from the natural-language questions that real people put to AI systems about medicines. The questions show what people want to know and what worries them, in their own words, before any company shapes the question. It supplements spontaneous reports and claims data. It does not replace them.
How is a real user question different from a test prompt in AI visibility monitoring?
An analyst writes a test prompt, so it shows how a model answers what the analyst chose to ask. A real user question comes from a person with a need, so it shows what people ask. FDA’s 2011 draft guidance draws a similar line between requests that are independent of a firm and requests the firm prompted [10].
Does a question typed into a chatbot count as an adverse event report?
This article does not establish that it does. A company’s safety team decides whether a piece of text meets its intake criteria, with legal input. A question describing a personal symptom after taking a drug is the type most likely to need that review, which is why the taxonomy above separates experience questions from the rest.
How many people ask AI about health?
KFF found that 32% of US adults used AI chatbots for health information in the past year, in a poll of 1,343 adults fielded in February and March 2026 [3]. OpenAI’s report, as relayed by Becker’s, says one in four of more than 800 million regular ChatGPT users submits a health prompt each week [1].
How accurate are AI answers to drug questions?
It depends on the model, the date and the question. A 2023 test of Bing Copilot on 500 answers found 26.0% of assessable answers below 100% accuracy against a reference database [5]. A Long Island University study found 10 of 39 ChatGPT responses satisfactory [7]. These systems have changed, so use the results as a reason to check, not as a current score.
Do AI systems cite real sources for medical claims?
Not reliably in the studies we found. In the BMJ Open audit, the chatbots returned 1,013 of 1,250 requested references, median completeness was 40%, and none produced a complete and accurate list for any prompt [6]. In the Long Island University study, all eight responses that supplied references included nonexistent ones [7].
Can question or search data detect drug interactions?
Search data did in one published study. White and colleagues analyzed 82 million queries from 6 million searchers and flagged the pravastatin and paroxetine pair for hyperglycemia, with an AUC of 0.8189 across 31 known interacting and 31 unrelated pairs [8]. The authors caution that a single pair can reflect chance. We found no published equivalent for AI question text.
Is DrugChatter’s question database the only one of its kind?
DrugChatter says it is the only database of actual user questions to LLMs built for pharmaceutical use. We could not verify that claim. We confirmed that Profound sells real-user prompt data from consumer panels across all topics [11], and we did not find a pharma-specific equivalent in our sources.
Which team should read AI question data first?
Medical Information is the natural starting point, because it already answers drug questions and can act on recurring ones. Pharmacovigilance should agree in writing how experience-type questions are handled. Other teams can then add their own views.
How should a team verify an AI answer about a drug?
Compare it with the current FDA-approved prescribing information, record the label version and date, and record the model, the prompt and the test date. Open every cited reference. Have a second reviewer classify any discrepancy, and send safety-relevant discrepancies to the right owner.
Andrikyan, W., Sametinger, S. M., Kosfeld, F., Jung-Poppe, L., Fromm, M. F., Maas, R., & Nicolaus, H. F. (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, 100-109. https://qualitysafety.bmj.com/content/34/2/100.full.pdf
Generative artificial intelligence-driven chatbots and medical misinformation: An accuracy, referencing and readability audit. (2026). BMJ Open, 16(4), e112695. https://bmjopen.bmj.com/content/16/4/e112695
White, R. W., Tatonetti, N. P., Shah, N. H., Altman, R. B., & Horvitz, E. (2013). Web-scale pharmacovigilance: Listening to signals from the crowd. Journal of the American Medical Informatics Association, 20(3), 404-408. https://pmc.ncbi.nlm.nih.gov/articles/PMC3628066
Hazell, L., & Shakir, S. A. W. (2006). Under-reporting of adverse drug reactions: A systematic review. Drug Safety, 29(5), 385-396. https://pubmed.ncbi.nlm.nih.gov/16689555/
U.S. Food and Drug Administration. (2011, December). Guidance for industry (draft): Responding to unsolicited requests for off-label information about prescription drugs and medical devices. https://www.fda.gov/media/82660/download
