One in three US adults used an AI chatbot for health information in the past year, according to a KFF poll fielded from February 24 to March 2, 2026. [1] OpenAI says more than 40 million people use ChatGPT for health questions every day. [4] In a randomized trial of 1,298 people, the same kind of chatbot that scored 94.9% on its own helped users name the right condition in fewer than 34.5% of cases. [3] Those three facts define the question: the channel is large, it is growing, and it does not yet perform the way its volume implies.
The Short Answer: Is AI Becoming the First-Line Medical Information Channel?
For a large and growing share of people, yes. AI chatbots are now a common first place to ask a health question, especially outside clinic hours and before a decision to see a provider. [1][4] The evidence does not support the stronger claim that AI replaces clinicians or label-based sources. It also does not show that the answers are reliable enough to carry the weight of a first-line channel.
This article separates what is documented from what is forecast. Documented facts come with a numbered source. Calculations are labeled as analysis. The scenarios in the later sections are interpretation, not prediction.
Usage is measured by surveys and by company-reported data. Both have limits, which we describe.
Accuracy is measured by a small number of studies, most of them testing models that have since been replaced.
Regulation of consumer chatbots is thin. The January 2026 FDA and EMA actions address other products. [12][13]
For pharma teams, the practical change is that questions about their drugs are now being answered by systems they do not control.
What the Data Shows: How Many People Ask AI About Health?
How many US adults use AI chatbots for health information?
About 32% of US adults used an AI chatbot for health information in the past year. KFF reported this figure on March 25, 2026, from a nationally representative poll of 1,343 adults with a margin of sampling error of plus or minus 3 percentage points. [1]
The poll gave more detail on what people ask about and why. [1]
29% of adults used AI for physical health information.
16% of adults used AI for mental health information.
Among users, 65% wanted quick or immediate information or support.
Among users, 41% looked up information before deciding whether to see a provider.
Among users, 36% felt more comfortable researching privately.
How many health questions does ChatGPT receive?
OpenAI reports that more than 40 million people use ChatGPT each day for health information, and that over 5% of all ChatGPT messages globally concern health care. Axios published these figures on January 5, 2026, from an OpenAI report. [4] OpenAI also stated at the launch of ChatGPT Health that over 230 million people globally ask health-related questions on the platform each week. [5]
“Over 5% of all ChatGPT messages globally concern health care, and 40 million people use ChatGPT each day for health information.” (OpenAI report, as reported by Axios, January 5, 2026) [4]
These are company-reported figures. The Axios article says the data came from anonymized ChatGPT interactions and a user survey, and it does not give the full method. [4] Treat them as strong evidence of scale, not as audited measurements.
Why do people ask AI before they ask a clinician?
The documented reasons are time, cost, access, and privacy. [1][4] The first two are the easiest to measure.
Timing: seven in ten conversations happen outside clinic hours
OpenAI reports that seven in ten health conversations on ChatGPT happen outside normal clinic hours. [4] A chatbot answers at 2 a.m. A medical information line usually does not.
Cost and access: one in five users cites affordability
Among AI health users, 19% said a major reason was that they could not afford to see a health professional. Another 18% said they had no regular doctor or could not get an appointment. [1] Among users under 30, 29% cited cost and 38% cited access. [1]
Why the KFF 2024 and 2026 figures cannot be compared directly
KFF measured 17% of adults using AI chatbots at least once a month for health information and advice in a poll fielded June 3 to 24, 2024. [2] The 2026 figure of 32% measures use in the past year. [1] The two questions use different time windows. A reader who subtracts one from the other will overstate or understate growth.
What the pair does show is direction. Monthly use was 25% among adults aged 18 to 29 in 2024. [2] The 2026 poll found that the share of adults using AI for health information was about equal to the share using social media for health information. [1]
What Counts as a First-Line Medical Information Channel? A Definition and a Test
Definition: first-line medical information channel
A first-line medical information channel is the source a person consults first when a health or drug question arises, before a clinician, pharmacist, label, or official website. This is our working definition for this article. It is not an established regulatory term.
A three-part test for first-line status
We use three criteria. A channel is first-line when it is (1) consulted before other sources, (2) consulted at the moment of the question, and (3) able to change what the person does next. This framework is original to this article.
Where the evidence passes and where it falls short
Criterion
Evidence
Assessment
Source
Consulted before other sources
41% of AI health users looked up information before deciding whether to see a provider
Supported for a large minority
[1]
Consulted at the moment of the question
Seven in ten ChatGPT health conversations occur outside clinic hours
Supported (company-reported)
[4]
Able to change behavior
42% of physical-health users and 58% of mental-health users did not follow up with a provider
Supported, direction of effect not measured
[1]
Consulted before a label or pharmacist for drug questions specifically
No survey in this review measured drug-label or pharmacist bypass directly
Not established
Gap in the evidence
The last row matters for pharma. The surveys we reviewed measure health information in general. They do not isolate questions about specific drugs.
The Five Findings That Matter
One in three US adults (32%) used an AI chatbot for health information in the past year. [1]
OpenAI reports 40 million daily and over 230 million weekly users asking health questions. [4][5]
In a 1,298-person randomized trial, users of large language models identified relevant conditions in fewer than 34.5% of cases, no better than a control group. [3]
In a 2023 pharmacist-led test, ChatGPT gave a satisfactory answer to 10 of 39 drug questions (25.6%, our calculation). [8]
In a 2026 audit of five chatbots, 49.6% of responses on misinformation-prone topics were problematic. [9]
What the Clinical Evidence Shows: Do Chatbots Help the Public Make Medical Decisions?
The Oxford randomized trial in Nature Medicine
Bean and colleagues published the trial in Nature Medicine on February 9, 2026. It randomized 1,298 participants to use a large language model or a source of their choice. Each participant worked through one of 10 medical scenarios. The models tested were GPT-4o, Llama 3, and Command R+. [3]
What the study measured
The study measured two outcomes: whether participants identified the underlying condition, and whether they chose the correct disposition, meaning the right course of action. [3]
The models tested alone identified the condition in 94.9% of cases. [3]
The models tested alone chose the correct disposition in 56.3% of cases on average. [3]
Participants using the models identified relevant conditions in fewer than 34.5% of cases. [3]
Participants using the models chose the correct disposition in fewer than 44.2% of cases. [3]
The authors reported that the participant results were no better than the control group. [3]
The calculated gap between model and user performance
Analysis: the condition-identification gap is at least 60.4 percentage points (94.9% minus 34.5%). The disposition gap is at least 12.1 percentage points (56.3% minus 44.2%). Both are lower bounds, because the paper reports the user figures as “fewer than.” [3] In relative terms, users reached less than 36% of the model’s standalone condition-identification rate (34.5 divided by 94.9).
The Register, reporting on the study, noted that users often struggled to give the chatbots relevant information and that the models sometimes mixed good and bad advice. The researchers also said that benchmarks based on medical licensing exams do not capture human and model interaction. [3]
What the Oxford trial did not test
The trial did not test drug-specific questions, current frontier models, or models with web search and citations. It tested three models available when the study was designed. [3] The result describes the interaction between untrained users and chatbots in clinical scenarios. It does not describe the accuracy of a chatbot answering a narrow question such as a dosing limit.
What AI Gets Wrong About Drugs: The Accuracy Studies
How accurate was ChatGPT on 39 real drug information questions?
Not very. Long Island University pharmacists presented a study at the ASHP Midyear Clinical Meeting in December 2023. They researched 45 questions from the university’s drug information service, excluded 6 for lack of literature, and gave 39 to the free version of ChatGPT. Ten answers were satisfactory. [8]
Analysis: 10 of 39 is 25.6% satisfactory, and 29 of 39 (74.4%) were not. The study’s categories overlapped. Eleven answers did not directly address the question, 10 were inaccurate, and 12 were incomplete. [8]
The Paxlovid and verapamil interaction
Asked whether Paxlovid and verapamil interact, ChatGPT said no interactions had been reported. The lead author, Sara Grossman, PharmD, said the drugs can interact and that combined use may lower blood pressure excessively. [8]
Citation failures
When asked for references, ChatGPT supplied them for 8 of the 39 questions (20.5%, our calculation). Every one of those eight responses included at least one reference that did not exist. [8]
What did the BMJ Quality and Safety study find about search-engine chatbots?
Andrikyan and colleagues queried Bing Copilot in April 2023 with 10 patient questions about each of 50 frequently prescribed US outpatient drugs, which produced 500 answers. [7]
Mean answer accuracy was 88.7% and mean completeness was 76.7%. [7]
The mean Flesch Reading Ease score was 37.2, which indicates college-level text. [7]
Of the 20 answers that experts reviewed for patient harm, 66% (95% CI 50% to 85%) were judged as potentially leading to harm. [7]
Of those same 20 answers, 22% (95% CI 10% to 40%) were judged as possibly leading to death or severe harm. [7]
Analysis: 66% and 22% of 20 answers equal roughly 13 and 4 answers. The expert-reviewed sample is small and the confidence intervals are wide. The study’s conclusion was that chatbots gave overall complete and accurate drug information, but that readability was low, many answers were incorrect or potentially harmful, and caution is advised before recommending them to patients. [7]
How often do popular chatbots give problematic health answers?
About half the time on misinformation-prone topics. A 2026 BMJ Open audit led by a UCLA researcher tested five chatbots (Gemini, DeepSeek, Meta AI, ChatGPT, and Grok) with 250 questions in February 2025. The topics were cancer, vaccines, stem cells, nutrition, and athletic performance. [9]
49.6% of responses were problematic: 30% somewhat problematic and 19.6% highly problematic. [9]
Response quality did not differ significantly by chatbot. [9]
Median reference completeness was 40%, and no chatbot provided a fully accurate reference list. [9]
Analysis: on 250 questions, 49.6% is 124 problematic responses, of which about 49 were highly problematic. The topics were chosen because they attract misinformation, so this rate does not describe all health questions. The researchers concluded that “continued deployment without public education and oversight risks amplifying misinformation.” [9]
What happened when Google AI Overviews answered liver test questions?
The Guardian reported on January 2, 2026 that AI Overviews gave misleading answers to some health queries. The example was “what is the normal range for liver blood tests,” where the numbers shown ignored nationality, sex, ethnicity, and age. [10]
On January 11, 2026, TechCrunch reported that AI Overviews no longer appeared for that query or for “what is the normal range for liver function tests.” Variations such as “lft reference range” could still trigger an AI summary. A Google spokesperson said an internal clinical team found that “in many instances, the information was not inaccurate and was also supported by high quality websites.” [10] The two accounts disagree, and we do not resolve the dispute here.
Why these accuracy studies age quickly
Three of the four studies tested models from 2023 to early 2025. [7][8][9] Newer models, search access, and retrieval from labels may change results. The studies show that error rates were material on specific dates. They do not show current rates. Any team that relies on this evidence should test the current model on its own drug.
Comparison Table: Four Evaluations of AI Health and Drug Answers
Study
Date
System tested
Sample
Key result
Source
Long Island University / ASHP
Presented Dec. 2023
ChatGPT (free version)
39 drug questions
10 satisfactory; 8 responses with references all contained non-existent citations
[8]
Andrikyan et al., BMJ Quality & Safety
Queried Apr. 2023
Bing Copilot
500 answers; 20 expert-reviewed
66% of reviewed answers potentially harmful; 22% possibly death or severe harm
[7]
Bean et al., Nature Medicine
Published Feb. 9, 2026
GPT-4o, Llama 3, Command R+
1,298 participants, 10 scenarios
Users identified conditions in fewer than 34.5% of cases; no better than control
[3]
BMJ Open audit
Tested Feb. 2025
Gemini, DeepSeek, Meta AI, ChatGPT, Grok
250 questions
49.6% problematic; median reference completeness 40%
[9]
Where AI and the Evidence Diverge: A Four-Type Taxonomy of Drug Questions
Not every drug question carries the same risk. We propose a four-type classification based on where the correct answer lives. This is an original framework, not an industry standard. It is meant to help teams decide where monitoring effort should go.
Type 1: label-answerable questions
The answer sits in the current prescribing information. Examples are the approved indication, the contraindications, and the dosing range. The test for an AI answer is whether it matches the current label. The failure modes are outdated information and missing warnings.
Type 2: evidence-answerable questions
The answer sits in trials or guidelines, not on the label. Examples are comparative efficacy and place in therapy. The test is whether the claim is supported by a cited publication. The failure modes are invented or misattributed citations, which the ASHP and BMJ Open studies both documented. [8][9]
Type 3: personal-context questions
The answer depends on the person’s medications, conditions, and test results. Interaction questions belong here, as the Paxlovid and verapamil example shows. [8] The failure mode is a confident answer that ignores context the user did not supply. The Oxford trial found that users often struggled to give models relevant information. [3]
Type 4: judgment questions
The answer needs a clinician’s decision, for example whether to stop a medicine or how urgently to seek care. The Oxford trial measured a version of this as disposition, where models alone reached 56.3% and users reached fewer than 44.2%. [3]
How the four types map to monitoring effort
Type
Authoritative source
Can an automated check verify it?
Main owner
1. Label-answerable
Current prescribing information
Largely yes
Regulatory Affairs, Medical Information
2. Evidence-answerable
Trials, guidelines, PubMed
Partly; citations can be checked
Medical Affairs
3. Personal-context
Label plus patient facts
Only with a defined test case
Medical Information, Pharmacovigilance
4. Judgment
Clinician
No; review needed
Medical Affairs, Legal
What AI Companies Are Doing: ChatGPT Health and Claude for Healthcare
What is ChatGPT Health?
ChatGPT Health is a health-focused space inside ChatGPT that OpenAI announced in January 2026. Reporting on the launch states that more than 260 physicians across 60 countries gave feedback over two years, providing input on model outputs more than 600,000 times. In the US, users can import medical records from providers through a partnership with b.well. [5]
OpenAI states that the product is “explicitly not intended for diagnosis or treatment,” and the product is not FDA approved. [5] Rollout was staged and excluded the European Economic Area, Switzerland, and the UK at launch. [5]
What is Claude for Healthcare?
Anthropic announced Claude for Healthcare on January 11, 2026. US Claude Pro and Max subscribers can connect electronic medical records through HealthEx, which consolidates records from more than 50,000 health systems, and lab results through Function Health. [6] Anthropic states that users choose what to share, can revoke access at any time, and that health data is not used for model training. [6]
Anthropic also announced connectors for providers, insurers, and pharma, including ClinicalTrials.gov, PubMed, and Medidata. [6]
What changes when a chatbot can read the user’s own medical records?
It moves some questions from Type 2 and Type 3 toward grounded answers about the person’s own medications and results. That may improve accuracy for those questions. It also raises privacy stakes. KFF found that 41% of AI health users (13% of all adults) had uploaded personal medical information such as test results or doctors’ notes, and 77% of adults were concerned about the privacy of medical information shared with AI tools. [1]
Analysis: 13% of all adults divided by 32% of adults who are users gives 40.6%, which matches KFF’s reported 41%.
Who Regulates a Consumer Health Chatbot? What FDA and EMA Said in January 2026
What did FDA’s January 2026 guidances change?
On January 6, 2026, FDA updated its General Wellness and Clinical Decision Support Software guidances. They supersede versions from September 2019 and September 2022. The guidances are non-binding and took effect on publication. [12]
The clinical decision support changes concern tools for health care professionals. A consumer product falls under the general wellness framework only if it avoids medical or clinical claims such as screening, diagnosing, or monitoring a disease. [12] The Foley summary does not analyze generative AI chatbots directly. [12]
Do the FDA and EMA AI principles cover consumer chatbots?
On January 14, 2026, EMA and FDA jointly published ten principles for good AI practice in the medicines lifecycle. They address AI used in evidence generation and monitoring from early research through manufacturing and safety monitoring, and they are aimed at medicine developers and marketing authorization applicants and holders. [13] The EMA page does not address consumer chatbots. [13]
The result is a gap. Regulators have set expectations for companies that use AI in development. No document we reviewed sets accuracy expectations for a general chatbot that answers questions about a company’s marketed drug.
Scenario Analysis: What If AI Becomes the Default First Stop?
The scenarios below are interpretation. We built them from the documented facts above and state the assumption behind each. They are not forecasts.
Scenario A: AI stays a supplement
Assumption: use stays near today’s level and people still verify with a clinician. Under this scenario, AI answers shape the questions people bring to appointments. The documented evidence that 41% of users look up information before deciding whether to see a provider is consistent with it. [1]
Scenario B: AI becomes the default for routine drug questions
Assumption: users stop checking labels and pharmacists for routine questions. Under this scenario, Type 1 errors matter most, because they are the easiest to verify and the most costly to ignore. The ASHP and BMJ studies show that Type 1 and Type 3 errors occurred at material rates in 2023. [7][8]
Signal to watch
Watch whether surveys begin to ask about drug-specific questions. Current surveys do not isolate them. [1][2]
Scenario C: AI answers with the patient’s records attached
Assumption: record connections spread beyond early adopters. [5][6] Under this scenario, personal-context accuracy improves for users who connect records, and privacy and consent become the main issues. KFF found that 77% of adults are concerned about privacy. [1]
Signal to watch
Watch for peer-reviewed tests of record-grounded chatbots against pharmacist review. We did not find one in this review.
What This Means for Medical Affairs and Medical Information
How should Medical Information teams treat AI-generated questions?
As a new upstream source of inbound questions. People who ask an AI first will arrive at a medical information line with the AI’s claims, correct or not. Teams need to know what those claims are. A medical information team that tracks only its own call logs sees the question after the AI has framed it.
Unsolicited requests
FDA has had a draft guidance on responding to unsolicited requests for off-label information since December 2011. [14] A question typed into a public chatbot is not a request to the company, so the guidance does not apply to it directly. The practical issue is what the company says when the same question reaches its own medical information service.
What should Medical Affairs measure?
Medical Affairs should measure whether AI answers about its drugs match the current label and cite real publications. The documented failure modes are non-existent references, missing interaction warnings, and answers that omit context. [7][8][9]
What This Means for Pharmacovigilance
Does AI-mediated advice create a pharmacovigilance blind spot?
It may. KFF found that among AI users, 42% who asked about physical health and 58% who asked about mental health did not follow up with a provider. [1] Analysis: 29% of adults used AI for physical health and 42% of them did not follow up, which is about 12.2% of all adults. For mental health, 16% times 58% is about 9.3%. These are derived estimates. They assume the follow-up rates apply to all users of each topic.
The surveys do not say whether these users later experienced adverse events. We have no evidence on that point, and we do not claim one. The narrower point is that a growing share of health decisions happen before any clinician contact, so a report that would once have started in a clinic may start in a chat window.
What PV teams can do now
Track whether AI answers about a drug state its boxed warning, major interactions, and contraindications as the label does. Treat a missing warning as a finding to route to Regulatory and Medical Information.
What This Means for Brand, Regulatory, and Legal Teams
What should brand teams know about AI answers?
AI systems describe brands using whatever sources they retrieve or learned. The Oxford, ASHP, and BMJ results show that the output can differ from the label. [3][7][8] Brand teams cannot edit the answer. They can make sure the authoritative sources an AI would retrieve are current and easy to find.
What should Regulatory Affairs and Legal track?
Regulatory should track label changes against what AI systems say, because label changes can leave outdated statements in circulation. Legal should note that the regulatory framework for consumer chatbots is not settled. The January 2026 FDA guidance does not address generative AI directly, and the EMA and FDA principles do not cover consumer chatbots. [12][13]
How do physicians fit in?
Physicians use AI too. The AMA’s 2026 survey of nearly 1,700 physicians found that 81% use health AI, more than double the 2023 rate. The most common use was summarizing medical research and standards of care, at 39%. [11] Analysis: more than double implies the 2023 rate was below 41%. Patients and physicians may now both start from AI answers.
How to Monitor What AI Says About a Drug: Method and Limits
What a monitoring study needs to record
Another researcher should be able to repeat the test. A sound design records these items.
The AI platform, model, and version, with the date and time of each query.
The exact prompt, whether web search was on, and whether citations were requested.
The label version and other sources used to verify the answer.
How duplicate responses were handled and how discrepancies were classified and graded by severity.
Where DrugChatter fits
DrugChatter’s chatmon monitoring runs a library of drug questions across several AI models on a schedule and compares the answers with authoritative sources such as the label. That is one example of the method described above. It is not a substitute for primary evidence. This article does not report DrugChatter test data. Every figure comes from the cited studies and surveys.
Limits of this analysis
Usage figures from OpenAI are company-reported. [4][5]
The accuracy studies tested older models and small samples. [7][8][9]
No source we reviewed measured drug-specific questions at population scale.
Our calculations use the reported percentages and inherit their uncertainty.
Key Takeaways
AI is a first stop for a large minority of US adults: 32% used a chatbot for health information in the past year. [1]
OpenAI reports 40 million daily and over 230 million weekly health users, and seven in ten conversations occur outside clinic hours. [4][5]
In a randomized trial, users of language models identified conditions in fewer than 34.5% of cases, no better than control, while the models alone scored 94.9%. [3]
In a pharmacist-led test, 10 of 39 ChatGPT drug answers were satisfactory, and every reference ChatGPT supplied included a non-existent citation. [8]
In a 2026 audit, 49.6% of answers from five chatbots on misinformation-prone topics were problematic. [9]
Regulators have not set accuracy expectations for consumer chatbots answering drug questions. [12][13]
Pharma teams can monitor Type 1 label-answerable questions most reliably, and should escalate Type 3 and Type 4 findings for human review.
FAQ
Is ChatGPT FDA approved for medical advice?
No. Reporting on the ChatGPT Health launch states that it is not FDA approved and is explicitly not intended for diagnosis or treatment. [5]
Can AI chatbots give accurate drug information?
Sometimes. Bing Copilot answers averaged 88.7% accuracy in one study, yet experts judged 66% of the 20 answers they reviewed as potentially harmful. [7] ChatGPT gave satisfactory answers to 10 of 39 questions in another. [8] Accuracy depends on the model, the date, and the question.
Do AI chatbots make up medical references?
They can. In the ASHP study, all 8 responses that included references contained at least one non-existent reference. [8] In the BMJ Open audit, no chatbot gave a fully accurate reference list. [9]
Do people follow up with a doctor after asking AI?
Often not. Among users who asked about physical health, 42% did not follow up with a provider. Among users who asked about mental health, 58% did not. [1]
Are AI health answers better than a web search?
In the Oxford trial, participants using language models did no better than a control group that used sources of their choice. [3]
Why do people use AI for health questions instead of a clinic?
Among users, 65% wanted quick information, 36% preferred private research, 19% cited cost, and 18% cited access barriers. [1]
Does Google AI Overviews give reliable health information?
The evidence is disputed. The Guardian found misleading liver test ranges, and Google removed AI Overviews for some of those queries. Google said its clinicians found the information was often not inaccurate. [10]
What do the FDA and EMA AI principles cover?
They cover AI used by medicine developers and marketing authorization holders for evidence generation and monitoring across the medicines lifecycle. The EMA page does not address consumer chatbots. [13]
Do doctors use AI?
Yes. In the AMA’s 2026 survey, 81% of physicians reported using health AI. [11]
How can a drug manufacturer find out what AI says about its product?
Run a fixed set of real patient and clinician questions across the major AI platforms on a schedule, record the platform, model, date, and prompt, and compare each answer with the current label and cited publications. The method is described above. Escalate discrepancies by severity.
References
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/
KFF. (2024). KFF Health Misinformation Tracking Poll: Artificial intelligence and health information (fielded June 3-24, 2024). https://www.kff.org/public-opinion/kff-health-misinformation-tracking-poll-artificial-intelligence-and-health-information/
Bean, A. M., Payne, R. E., Parsons, G., Kirk, H. R., Ciro, J., Mosquera-Gomez, R., Hincapie M, S., Ekanayaka, A. S., Tarassenko, L., Rocher, L., & Mahdi, A. (2026, February 9). 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 (abstract via Oxford University Research Archive: https://ora.ox.ac.uk/objects/uuid:ccb8b891-c738-4890-a4d5-45959d0e04d9; secondary report: The Register, February 9, 2026, https://www.theregister.com/2026/02/09/ai_chatbots_medical_advice_sucks/)
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
Unite.AI. (2026, January 8). OpenAI launches ChatGPT Health for 230 million weekly users. https://www.unite.ai/openai-launches-chatgpt-health-for-230-million-weekly-users
Fortune. (2026, January 11). Anthropic unveils Claude for Healthcare and expands life science features, partners with HealthEx to let users connect medical records. https://fortune.com/2026/01/11/anthropic-unveils-claude-for-healthcare-and-expands-life-science-features-partners-with-healthex-to-let-users-connect-medical-records/
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
Long Island University. (2023, December 5). Study finds ChatGPT provides inaccurate responses to drug questions [Press release on research presented at the ASHP Midyear Clinical Meeting, December 3-7, 2023]. PR Newswire. https://prnewswire.com/news-releases/study-finds-chatgpt-provides-inaccurate-responses-to-drug-questions-302005250.html
Van Beusekom, M. (n.d.). AI chatbots provide poor answers to medical questions half the time, study finds. CIDRAP, reporting Tiller, N. B., et al., BMJ Open, 16, e112695, https://doi.org/10.1136/bmjopen-2025-112695. https://www.cidrap.umn.edu/misc-emerging-topics/ai-chatbots-provide-poor-answers-medical-questions-half-time-study-finds
Ha, A. (2026, January 11). Google removes AI Overviews for certain medical queries. TechCrunch. https://techcrunch.com/2026/01/11/google-removes-ai-overviews-for-certain-medical-queries
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
Foley & Lardner LLP. (2026, January). Digital health policy: FDA relaxes restrictions over wearables and AI decision making tools in two new 2026 guidances. https://www.foley.com/insights/publications/2026/01/digital-health-policy-fda-relaxes-restrictions-over-wearables-and-ai-decision-ma/
European Medicines Agency. (2026, January 14). EMA and FDA set common principles for AI in medicine development. https://www.ema.europa.eu/en/news/ema-fda-set-common-principles-ai-medicine-development-0
U.S. Food and Drug Administration. (2011, December). Responding to unsolicited requests for off-label information about prescription drugs and medical devices [Draft guidance]. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/responding-unsolicited-requests-label-information-about-prescription-drugs-and-medical-devices
