{"id":289,"date":"2026-08-06T01:07:00","date_gmt":"2026-08-06T05:07:00","guid":{"rendered":"https:\/\/drugchatter.com\/insights\/?p=289"},"modified":"2026-08-08T13:42:59","modified_gmt":"2026-08-08T17:42:59","slug":"what-chatgpt-gets-wrong-about-drug-safety-in-pregnancy-and-why-pharma-should-care","status":"publish","type":"post","link":"https:\/\/drugchatter.com\/insights\/what-chatgpt-gets-wrong-about-drug-safety-in-pregnancy-and-why-pharma-should-care\/","title":{"rendered":"What ChatGPT Gets Wrong About Drug Safety in Pregnancy \u2014 And Why Pharma Should Care"},"content":{"rendered":"\n
\"\"<\/figure>\n\n\n\n

A pregnant patient searches ‘is it safe to take metformin while pregnant’ in ChatGPT. The answer she receives is plausible, fluent, and wrong in ways that matter. The AI cites no pregnancy category. It doesn’t mention that the FDA retired the A\/B\/C\/D\/X classification system in 2015. It doesn’t reference the Pregnancy and Lactation Labeling Rule (PLLR), which replaced those categories with narrative risk summaries. And it doesn’t say that the American Diabetes Association guidelines for gestational diabetes have shifted considerably since the training data it draws from was assembled. <\/p>\n\n\n\n

This is not a hypothetical. Queries about drug safety during pregnancy are among the most emotionally loaded and high-stakes searches patients make in any medium. When those queries land in AI systems rather than label databases or clinical guidelines, the stakes for pharmaceutical companies go up alongside the risks to patients.<\/p>\n\n\n\n

This article examines how major AI systems handle pregnancy drug safety queries, what that means for pharmacovigilance and regulatory compliance, how brand teams can monitor AI-generated answers about their products, and what pharmaceutical companies should build now before regulators build it for them.<\/p>\n\n\n\n

\n\n\n\n

Why Pregnant Patients Are Turning to ChatGPT for Drug Safety Answers<\/strong><\/h2>\n\n\n\n

How Patients Actually Search for Drug Safety Information During Pregnancy<\/strong><\/h3>\n\n\n\n

Pregnancy changes how patients seek medical information. Anxiety increases. Clinic access narrows. Questions arise at 2 a.m. when no physician is available. The 2023 Kaiser Family Foundation Health Misinformation Tracking Poll found that 25% of U.S. adults reported using AI chatbots for health information, with that figure climbing among adults under 35. Among pregnant women, who often have more frequent health questions and less confidence in self-medication decisions, the adoption of AI health search is likely higher than the general population.<\/p>\n\n\n\n

Search queries in this category cluster around a few patterns. Patients ask whether a specific named drug is ‘safe’ in pregnancy. They ask about trimesters specifically. They ask about alternatives. They ask about what their doctor prescribed but didn’t fully explain. And increasingly, they ask AI systems rather than typing into Google because AI systems produce synthesized, conversational answers rather than a list of links to navigate.<\/p>\n\n\n\n

The queries themselves reveal clinical nuance users don’t realize they’re asking: ‘Can I take Tylenol while pregnant’ is a different clinical question in the first trimester than at 37 weeks, following the 2021 FDA communication flagging potential links between prolonged prenatal acetaminophen exposure and neurodevelopmental disorders. AI systems generally cannot navigate that distinction without explicit prompting. Most don’t flag it unprompted at all.<\/p>\n\n\n\n

What Pregnant Patients Are Actually Asking AI Chatbots About Prescription Drugs<\/strong><\/h3>\n\n\n\n

Query patterns across AI health searches in pregnancy concentrate on a predictable but important drug list: antidepressants (SSRIs especially), antiepileptics, diabetes medications, antihypertensives, thyroid medications, antibiotics, and over-the-counter analgesics. These are drugs where the risk-benefit calculus is genuinely difficult, where clinical guidelines have evolved, and where AI systems are most likely to produce answers that are outdated or oversimplified.<\/p>\n\n\n\n

Sertraline (Zoloft) and escitalopram (Lexapro) generate enormous query volumes because SSRI use in pregnancy involves a real tradeoff between untreated maternal depression and a small but documented risk of neonatal adaptation syndrome and, historically, persistent pulmonary hypertension of the newborn. When patients ask ChatGPT whether sertraline is safe during pregnancy, the answer they get depends heavily on what version of the model they’re using, what the model was trained on, and whether they ask the question in a way that retrieves safety information versus general prescribing information. The answers vary enough across sessions to be functionally unreliable.<\/p>\n\n\n\n

Do AI Systems Know the FDA Replaced Pregnancy Drug Categories in 2015?<\/strong><\/h3>\n\n\n\n

This is a concrete, testable failure point. The FDA’s Pregnancy and Lactation Labeling Rule went into effect in June 2015, replacing the letter-category system (A, B, C, D, X) with detailed narrative sections covering human data, animal data, and clinical considerations. Drugs approved before June 2001 have until 2020 to comply. Many older generic drugs still carry legacy labeling.<\/p>\n\n\n\n

When asked about drug safety in pregnancy, major AI systems frequently revert to the old letter categories. This is not simply a stylistic issue. The letter categories collapsed nuance in ways that led to poor prescribing decisions \u2014 a drug could be Category C because animal data showed risk even when substantial human data showed safety. Narrating that a drug is ‘Category C’ to a patient is not just outdated; it’s clinically misleading in ways that could prompt them to discontinue medication that was appropriate to continue.<\/p>\n\n\n\n

Testing across ChatGPT-4o, Gemini 1.5 Pro, and Claude Sonnet in early 2025 showed that all three models sometimes volunteered pregnancy categories when asked about specific drugs, despite those categories being formally retired. Some responses included caveats; many did not. None consistently directed patients to the actual FDA-approved labeling or to resources like LactMed or the OTIS (Organization of Teratology Information Specialists) counseling service.<\/p>\n\n\n\n

\n\n\n\n

How ChatGPT, Gemini, and Claude Handle Pregnancy Drug Queries Differently<\/strong><\/h2>\n\n\n\n

ChatGPT Drug Safety Answers: What the Model Gets Right and Wrong<\/strong><\/h3>\n\n\n\n

ChatGPT’s responses to pregnancy drug safety queries are shaped by its training data mix and by the system prompts OpenAI applies at the API and consumer levels. In general, GPT-4-class models produce responses that are fluent, hedged with disclaimers about consulting a physician, and structured to seem responsible. The problem is not that they are irresponsible in tone. The problem is that specific factual claims are frequently either incomplete or inconsistent across sessions.<\/p>\n\n\n\n

On well-documented drugs with clear black box warnings \u2014 thalidomide, isotretinoin (Accutane), valproic acid \u2014 ChatGPT performs reasonably. These are drugs with extreme, well-publicized teratogenicity, and their contraindication in pregnancy is embedded deeply enough in training data that the model retrieves it reliably. The risk failures appear on the edges: drugs where the evidence is genuinely evolving, where FDA communications are recent, or where clinical practice has diverged from older labeling.<\/p>\n\n\n\n

Acetaminophen is the clearest example. The FDA’s 2023 updated communication on prenatal acetaminophen exposure \u2014 specifically around prolonged use \u2014 does not appear consistently in ChatGPT responses. Newer FDA Drug Safety Communications and MedWatch alerts from 2022 onward fall outside or near the edge of training cutoffs for most model versions, which means patients asking about acetaminophen in pregnancy may receive advice that precedes the current cautionary language from the FDA.<\/p>\n\n\n\n

How Claude Handles Pregnancy Drug Safety Compared to ChatGPT and Gemini<\/strong><\/h3>\n\n\n\n

Anthropic’s Claude models show a different pattern. Claude tends to be more explicit about uncertainty and more willing to decline confident recommendations, defaulting to language that frames answers as general information rather than clinical advice. This is appropriate epistemic caution, but it creates a separate problem for pharmaceutical monitoring: Claude’s responses are often less specific about named drugs, which makes AI monitoring for brand mention share-of-voice harder to track in this model.<\/p>\n\n\n\n

When Claude is asked about specific branded drugs in pregnancy \u2014 Humira in rheumatoid arthritis patients considering pregnancy, for example, or Dupixent in pregnant patients with atopic dermatitis \u2014 it often provides accurate high-level information but may not distinguish between the biosimilar landscape and the originator product in ways that matter for brand monitoring. A brand team tracking how often ‘adalimumab’ versus ‘Humira’ appears in AI responses would see Claude frequently substituting INN (international nonproprietary name) for brand name, which shifts share-of-voice metrics systematically.<\/p>\n\n\n\n

Gemini Drug Safety Responses: Google’s AI and Pregnancy Information<\/strong><\/h3>\n\n\n\n

Gemini, Google’s AI system, has the structural advantage of integration with Google Search’s real-time index, at least in some response modes. This means Gemini can, in principle, retrieve more current FDA communications than a model relying solely on pre-training data. In practice, Gemini’s pregnancy drug safety responses reflect this inconsistently. For FDA Drug Safety Communications published within the last year, Gemini sometimes retrieves current information and sometimes produces responses consistent with pre-update guidance.<\/p>\n\n\n\n

Google’s health information policies and its medical information panels have historically emphasized sourcing from established medical authorities \u2014 NIH, CDC, Mayo Clinic. Gemini tends to reflect this in its response architecture: answers about drug safety in pregnancy often surface information from these institutional sources. The risk is that those sources may not reflect the most current FDA labeling or the most recent clinical guideline updates from ACOG (American College of Obstetricians and Gynecologists) or SMFM (Society for Maternal-Fetal Medicine).<\/p>\n\n\n\n

Perplexity AI and Drug Safety: The Citation Problem<\/strong><\/h3>\n\n\n\n

Perplexity AI’s model generates cited responses, which creates a different monitoring dynamic. For pharmaceutical brand teams, Perplexity is both more transparent and more revealing: you can see exactly which sources the AI is synthesizing, which means you can audit whether your drug’s labeling, your company’s clinical publications, or your patient support resources appear in the citation set. You can also see which competitor sources are being cited and whether generic drug resources or price comparison sites are being surfaced over branded prescribing information.<\/p>\n\n\n\n

The citation model also exposes a reliability gap: Perplexity sometimes cites drug information databases, patient forums, or third-party pharmacy sites that may carry outdated drug safety information. A cited answer isn’t necessarily a correct one. For pregnancy drug safety specifically, forum content from BabyCenter, What to Expect, or Reddit’s r\/BabyBumps sometimes appears in cited Perplexity responses \u2014 sources with no regulatory standing and variable accuracy.<\/p>\n\n\n\n

\n\n\n\n

Can AI Hallucinations About Pregnancy Drug Safety Trigger FDA Risk?<\/strong><\/h2>\n\n\n\n

What Counts as a Reportable Adverse Event When AI Is the Information Source?<\/strong><\/h3>\n\n\n\n

This is where pharmaceutical regulatory affairs teams need to pay close attention. The FDA’s postmarketing adverse event reporting requirements, under 21 CFR Part 314.81 and Part 600.80, require manufacturers to report adverse events they become aware of. The question of whether an AI-generated misinformation-driven medication decision could constitute a reportable event is not yet resolved in regulatory guidance, but it is being discussed.<\/p>\n\n\n\n

The more immediate risk is indirect. If a pregnant patient discontinues a prescribed medication because ChatGPT told her it was unsafe \u2014 and she or her provider then reports an adverse event related to the untreated condition \u2014 the chain of causality runs through AI-generated misinformation. If a drug company’s pharmacovigilance team isn’t monitoring AI outputs for their products, they are missing a signal channel that is already influencing patient behavior.<\/p>\n\n\n\n

FDA’s 2023 discussion paper on AI and machine learning in drug development and postmarketing surveillance flagged AI-generated content as an emerging pharmacovigilance challenge without providing specific guidance. The European Medicines Agency has moved slightly faster, with its 2023 AI workplan including language about AI-generated information in patient communication monitoring. Neither agency has issued binding guidance on how AI chatbot outputs should be handled by pharmacovigilance teams, but the regulatory trajectory is clear: companies that build this capability now will be ahead of requirements, not behind them.<\/p>\n\n\n\n

Real FDA Warning Letters Related to Drug Misinformation Online<\/strong><\/h3>\n\n\n\n

The FDA has issued warning letters related to drug misinformation online, providing a proxy for where AI-generated misinformation risk ultimately lands. In 2022, the FDA issued a warning letter to a company marketing a product with unsubstantiated pregnancy safety claims on its website. The agency’s Office of Prescription Drug Promotion (OPDP) monitors promotional claims across digital platforms. Social media monitoring is explicitly part of OPDP’s published workplan.<\/p>\n\n\n\n

The extension of OPDP scrutiny to AI-generated outputs is a logical next step, particularly if an AI system trained on or citing a pharmaceutical company’s own content produces a claim that diverges from approved labeling. If an AI system cites a drug company’s website as the source for a pregnancy safety claim that contradicts the approved label, the regulatory exposure for that company is real, even if the AI generated the synthesis.<\/p>\n\n\n\n

Off-Label Pregnancy Use: How AI Systems Discuss Unapproved Indications<\/strong><\/h3>\n\n\n\n

Off-label drug use during pregnancy is common and often clinically appropriate \u2014 many drugs used to treat pregnancy complications, including preterm labor, have complex regulatory histories. Progesterone supplementation, hydroxyprogesterone caproate (Makena), and terbutaline are examples where off-label use, approved labeling, and withdrawal of approvals intersect in ways that AI systems handle poorly.<\/p>\n\n\n\n

Makena’s story is instructive. AMAG Pharmaceuticals (later Covis Pharma) received accelerated approval for hydroxyprogesterone caproate in 2011 to reduce preterm birth risk. The confirmatory trial, PROLONG, failed to show efficacy. The FDA’s advisory committee voted in 2023 to withdraw approval, and the drug was ultimately withdrawn from the market. Patients asking AI systems about progesterone injections for preterm birth prevention may still receive references to 17-OHPC or Makena as an option, depending on when the model was trained. This is a live AI hallucination risk with direct clinical implications.<\/p>\n\n\n\n

\n\n\n\n

Tracking AI Brand Share-of-Voice for Pregnancy Drug Mentions<\/strong><\/h2>\n\n\n\n

How Often Does an AI Mention Your Drug vs a Competitor’s in Pregnancy Queries?<\/strong><\/h3>\n\n\n\n

Share-of-voice monitoring in AI systems is fundamentally different from traditional digital share-of-voice measurement. In search, you count impressions, click-through rates, and ranking position. In AI, you are measuring mention frequency, context of mention, sentiment around the mention, and whether the AI is recommending, warning against, or substituting your product in a given query context.<\/p>\n\n\n\n

For pregnancy drug safety specifically, the relevant competitive set depends on the therapeutic area. In the antidepressant market, the question is whether AI systems preferentially mention sertraline or escitalopram when asked about ‘safe antidepressants during pregnancy’ \u2014 and whether those mentions are positively or negatively framed. In the antihypertensive market, the question is which agents \u2014 labetalol, nifedipine, methyldopa \u2014 AI systems recommend for gestational hypertension, and whether branded or generic framing dominates.<\/p>\n\n\n\n

Pharmaceutical brand teams that run systematic query monitoring across AI platforms can measure this. The methodology requires running a defined query set against each major AI system (ChatGPT, Gemini, Claude, Perplexity, Copilot), logging responses, extracting drug mentions, and coding the context of each mention across a consistent taxonomy: recommended, contraindicated, mentioned as alternative, mentioned with safety warning, not mentioned. Running this quarterly provides a trend line that search monitoring cannot.<\/p>\n\n\n\n

Tools like DrugChatter’s AI monitoring platform<\/a> are built for exactly this use case \u2014 enabling pharmaceutical teams to track how AI systems discuss their drugs across platforms, detect sentiment shifts, and compare brand versus generic mention rates without building a custom monitoring stack.<\/p>\n\n\n\n

Do LLMs Recommend Generic Drugs More Often Than Branded Products in Pregnancy Contexts?<\/strong><\/h3>\n\n\n\n

The evidence suggests yes, with important caveats. AI systems are trained on large web corpora where generic drug names dominate in medical literature, clinical guidelines, and patient education materials. Clinical guidelines almost universally use INN names. FDA labeling uses both. Drug databases like Drugs.com and Medscape use both but often lead with generic names. The net effect is that AI models have more training signal around generic names than brand names for most drug classes.<\/p>\n\n\n\n

For pregnancy drug safety, this has practical implications. A query about ‘safe blood pressure medication during pregnancy’ is more likely to produce a response naming ‘labetalol’ or ‘nifedipine’ than ‘Trandate’ or ‘Procardia.’ A query about ‘antidepressants safe in pregnancy’ will typically surface ‘sertraline’ before ‘Zoloft.’ Brand monitoring teams need to account for this systematic shift and measure it explicitly \u2014 tracking whether their brand name appears at all in AI responses, not just whether the INN appears.<\/p>\n\n\n\n

The exception to this pattern is drugs with strong brand-name cultural saturation \u2014 Ozempic is a good example, though not a pregnancy drug. In therapeutic categories where the brand name has become more common than the INN in patient-facing discourse (because of advertising, media coverage, or cultural prominence), AI systems show greater brand name retrieval. Pharmaceutical companies with older generic-dominated drugs have less leverage here than those with newer, heavily marketed brands.<\/p>\n\n\n\n

How to Track AI Mention Frequency for Your Drug Across ChatGPT, Gemini, and Claude<\/strong><\/h3>\n\n\n\n

A workable AI share-of-voice monitoring program for pharmaceutical companies involves four layers. First, define the query set: map the questions patients and physicians actually ask about your drug in pregnancy contexts, drawing from search analytics, call center data, and medical information request logs. Second, run those queries systematically across AI platforms at regular intervals, capturing full response text. Third, extract drug mentions and code context using a consistent taxonomy. Fourth, trend the data over time and flag anomalies \u2014 sudden drops in mention frequency, emergence of safety language that wasn’t present before, or shifts in which competitor drugs are being mentioned alongside yours.<\/p>\n\n\n\n

The technical infrastructure for this is not trivial but is well within the capability of a pharmaceutical digital intelligence team or an external vendor. The API access required to run queries programmatically is available for ChatGPT (OpenAI API), Claude (Anthropic API), and Gemini (Google API). Perplexity offers API access. The challenge is consistency \u2014 AI responses vary across sessions even with identical prompts, which means adequate sample sizes are required to measure share-of-voice reliably.<\/p>\n\n\n\n

DrugChatter’s monitoring infrastructure<\/a> handles this multi-platform query architecture natively, providing pharmaceutical companies with aggregated share-of-voice data across AI systems without requiring internal infrastructure build.<\/p>\n\n\n\n

\n\n\n\n

Pharmacovigilance in the Age of AI Search: What Drug Companies Must Build Now<\/strong><\/h2>\n\n\n\n

Can AI Outputs Be Used as Pharmacovigilance Signal Sources?<\/strong><\/h3>\n\n\n\n

The answer is yes \u2014 with significant methodological qualifications. Traditional pharmacovigilance draws on spontaneous adverse event reports, clinical trial data, literature surveillance, and structured patient registries. Social media listening has been an emerging signal source for over a decade, with the FDA’s 2011 draft guidance on social media pharmacovigilance marking the formal acknowledgment that patient-generated digital content carries signal value.<\/p>\n\n\n\n

AI-generated content occupies a different position. AI outputs are not patient reports \u2014 they are synthetic text that may or may not reflect actual patient experience. But two indirect pharmacovigilance use cases are credible. First, AI outputs can reveal what patient questions and concerns are driving queries \u2014 questions about a specific side effect or drug interaction that appears with high frequency in AI query logs suggests that patients are seeking that information, which is itself a signal worth tracking. Second, AI systems trained on patient forum data (Reddit, PatientsLikeMe, condition-specific communities) may synthesize patient safety signals embedded in that training data in their responses to clinical questions.<\/p>\n\n\n\n

The FDA’s Patient-Focused Drug Development (PFDD) guidance series has increasingly emphasized patient-reported outcomes and real-world evidence. AI query pattern analysis is a natural extension of that methodology \u2014 with appropriate controls for the fact that AI query volume reflects information-seeking behavior, not clinical outcomes.<\/p>\n\n\n\n

What Reddit and Patient Forums Feed Into AI Drug Safety Responses About Pregnancy<\/strong><\/h3>\n\n\n\n

Reddit is a meaningful component of LLM training data. OpenAI’s data licensing agreement with Reddit, announced in May 2024, confirmed what AI researchers had long assumed: Reddit content is a significant source of training data for large language models. Subreddits like r\/BabyBumps, r\/pregnant, r\/gestationaldiabetes, r\/multipleSclerosis, and condition-specific communities where pregnant patients discuss medication decisions have generated millions of posts and comments that now potentially shape AI drug safety responses.<\/p>\n\n\n\n

This has direct implications for pregnancy drug safety. Patient forum discussions about medication during pregnancy are emotionally driven, often anecdotal, and sometimes flat-out incorrect. They also capture real patient experience that clinical literature misses. The synthesis of these sources in LLM training means that AI systems may reflect both the anxieties and the genuine insights present in those communities \u2014 without a physician’s ability to contextualize either.<\/p>\n\n\n\n

For pharmaceutical companies, this creates a monitoring imperative: understand what patients are saying on forums about your drug in pregnancy, because that content may be shaping what AI systems say about your drug in pregnancy. Forum sentiment analysis and AI output monitoring are two sides of the same intelligence problem.<\/p>\n\n\n\n

How Patient Forum Sentiment About Pregnancy Drug Safety Ends Up in AI Answers<\/strong><\/h3>\n\n\n\n

The mechanism is imprecise but real. LLMs are trained on text corpora where information about a topic is represented proportionally to how much text covers that topic. If a drug has extensive negative patient forum coverage \u2014 discussions of birth defects, miscarriages attributed to the drug, fear-driven discontinuation \u2014 that text increases the probability that the model will generate cautionary language about the drug when queried, independent of the clinical evidence base.<\/p>\n\n\n\n

Paroxetine (Paxil) is an example worth studying. Paroxetine received an FDA pregnancy category D designation in 2005 following an analysis suggesting a small increase in cardiac malformation risk in newborns. The label change generated substantial media coverage, patient forum activity, and clinical discussion. Subsequent research has produced mixed findings on the magnitude of that risk. But the volume of negative coverage in the training data for any LLM trained on pre-2020 web content is very high, which means AI systems may produce paroxetine-in-pregnancy responses that reflect the 2005-era alarm rather than the more nuanced 2024 clinical picture.<\/p>\n\n\n\n

Pharmaceutical companies whose drugs have undergone this kind of safety signal amplification in patient discourse should be particularly active in monitoring how AI systems discuss those drugs \u2014 and in ensuring that current, authoritative information is well-represented in accessible digital sources that AI systems can potentially retrieve.<\/p>\n\n\n\n

\n\n\n\n

What Eli Lilly, Novo Nordisk, and AstraZeneca Should Be Doing About AI Drug Monitoring<\/strong><\/h2>\n\n\n\n

How Big Pharma Is Responding to AI Search and Brand Risk in 2025<\/strong><\/h3>\n\n\n\n

Publicly disclosed AI monitoring programs among major pharmaceutical companies remain limited. A handful of companies have acknowledged building AI monitoring capabilities within their medical information or digital intelligence teams, but formal program disclosures are rare. What is visible from job postings, conference presentations, and vendor announcements suggests that the most sophisticated programs are in early-stage commercial deployment at a small number of large pharmaceutical companies.<\/p>\n\n\n\n

Eli Lilly, given the enormous AI visibility of Zepbound and Mounjaro (tirzepatide), has obvious incentive to monitor AI responses \u2014 both for safety-adjacent misinformation about off-label use in non-diabetic patients and for competitive intelligence against Novo Nordisk’s semaglutide products. The pregnancy angle for GLP-1s is particularly active: patients with obesity or type 2 diabetes who become pregnant, or who are considering pregnancy while on GLP-1 agonists, generate significant query volume about whether to continue medication during pregnancy. Current FDA guidance recommends discontinuing GLP-1 agonists at least two months before planned pregnancy, but AI systems handle this recommendation inconsistently.<\/p>\n\n\n\n

Novo Nordisk faces the same challenge for Ozempic and Wegovy. The FDA label for semaglutide includes explicit contraindication in pregnancy, but the volume of patient queries about semaglutide and pregnancy is high enough \u2014 given the drug’s cultural penetration \u2014 that AI responses on this topic are likely being seen by a meaningful number of pregnant patients or patients planning pregnancy.<\/p>\n\n\n\n

Which Drugs Are Most Frequently Mentioned by AI in Pregnancy Safety Contexts?<\/strong><\/h3>\n\n\n\n

Based on the clinical prevalence of conditions requiring management during pregnancy and the query volume those conditions generate, the drugs most likely to appear in AI responses about pregnancy drug safety fall into predictable therapeutic clusters:<\/p>\n\n\n\n