{"id":535,"date":"2026-06-23T08:05:00","date_gmt":"2026-06-23T12:05:00","guid":{"rendered":"https:\/\/drugchatter.com\/insights\/?p=535"},"modified":"2026-05-16T22:20:55","modified_gmt":"2026-05-17T02:20:55","slug":"ai-gets-oncology-drugs-wrong-what-pharma-brand-teams-must-track-now","status":"publish","type":"post","link":"https:\/\/drugchatter.com\/insights\/ai-gets-oncology-drugs-wrong-what-pharma-brand-teams-must-track-now\/","title":{"rendered":"AI Gets Oncology Drugs Wrong: What Pharma Brand Teams Must Track Now"},"content":{"rendered":"\n
\"\"<\/figure>\n\n\n\n

Ask ChatGPT whether pembrolizumab is approved for small cell lung cancer. The answer you get depends on which version you query, when you ask, and how the question is phrased. Sometimes the model says yes. Sometimes it cites a clinical trial that supported a different indication. Sometimes it confuses Keytruda with Opdivo. In every case, the model sounds confident.<\/p>\n\n\n\n

That confidence is the problem.<\/p>\n\n\n\n

Oncology is the therapeutic area where AI hallucinations carry the highest consequences. Drug regimens are complex, indication lists shift with every FDA approval cycle, and the line between approved use and off-label use is legally and clinically significant. Patients with cancer are searching AI chatbots for information about their treatment options. Oncologists are using AI-assisted tools for clinical decision support. And pharmaceutical brand teams have almost no systematic visibility into what these systems are saying about their products.<\/p>\n\n\n\n

This article documents the most common AI errors in oncology, explains why they happen, and lays out a monitoring framework that pharma companies can use to protect patients, manage regulatory risk, and stay ahead of AI-driven brand erosion.<\/p>\n\n\n\n

\n\n\n\n

Why Oncology Is the Highest-Risk Therapeutic Area for AI Hallucinations<\/h2>\n\n\n\n

How Rapidly Changing FDA Approvals Create AI Knowledge Gaps<\/h3>\n\n\n\n

The FDA approved 55 novel drugs in 2023, with oncology accounting for the largest single category. Indication expansions, accelerated approvals, and companion diagnostic requirements change the clinical picture for drugs like nivolumab (Opdivo), atezolizumab (Tecentriq), and olaparib (Lynparza) on a rolling basis.<\/p>\n\n\n\n

Large language models are trained on static datasets with cutoff dates. A model trained through early 2023 does not know about approvals from late 2023 or 2024. A model trained through late 2024 may still have sparse or inconsistent coverage of recent label expansions because regulatory documents are underrepresented in training corpora relative to news, social media, and consumer health content.<\/p>\n\n\n\n

The result: a model may confidently state that a drug is approved for an indication that was later withdrawn, or deny an approval that was granted after its knowledge cutoff. Both errors carry pharmacovigilance implications when patients or clinicians act on them.<\/p>\n\n\n\n

Why Combination Regimens Are Especially Vulnerable to AI Errors<\/h3>\n\n\n\n

Single-agent oncology questions are difficult enough. Combination regimens are far worse. Modern oncology routinely involves multi-drug protocols where specific sequences, dosing intervals, and companion biomarker requirements differ by tumor type and line of therapy.<\/p>\n\n\n\n

When a patient asks an AI chatbot whether they can take pembrolizumab with carboplatin for their lung cancer, the model is navigating a matrix of FDA approvals, NCCN guidelines, biomarker requirements (PD-L1 TPS, TMB-H), and line-of-therapy constraints. Getting any one element wrong produces a dangerous answer. AI systems trained on general web content are not built to reason through this matrix reliably.<\/p>\n\n\n\n

The Biomarker Problem: When AI Ignores Companion Diagnostics<\/h3>\n\n\n\n

Several oncology drugs are only approved for patients whose tumors test positive for specific biomarkers. Olaparib requires BRCA1\/2 mutation. Larotrectinib (Vitrakvi) is approved for NTRK gene fusion-positive tumors. Entrectinib covers NTRK and ROS1. Trastuzumab deruxtecan (Enhertu) carries HER2-low and HER2-ultralow designations that require specific IHC scoring.<\/p>\n\n\n\n

In testing across public AI chatbots, models frequently describe these drugs without mentioning companion diagnostic requirements. A patient reading that ‘Enhertu is approved for breast cancer’ without the HER2 qualification could draw incorrect conclusions about their eligibility. A physician using AI search to check coverage criteria could receive incomplete information at a critical decision point.<\/p>\n\n\n\n

\n\n\n\n

The Six Most Common AI Mistakes About Oncology Drugs<\/h2>\n\n\n\n

Mistake 1: Hallucinating Indications That Were Never Approved<\/h3>\n\n\n\n

This is the most dangerous class of error. AI models sometimes describe drugs as approved for tumor types where clinical trials failed or were never completed. Atezolizumab (Tecentriq) had its breast cancer accelerated approval withdrawn by Genentech in 2021 after the IMpassion131 trial failed to confirm the IMpassion130 benefit. Models trained on pre-withdrawal data may still associate Tecentriq with triple-negative breast cancer without flagging the withdrawal.<\/p>\n\n\n\n

Pembrolizumab had its accelerated approval for certain gastric and esophageal indications converted or updated multiple times between 2017 and 2023. The layered history of accelerated approval, conversion, and label updates creates a timeline that static-trained models handle poorly. The model may return the correct drug name and the correct cancer type but apply the wrong approval status or the wrong line-of-therapy qualifier.<\/p>\n\n\n\n

Mistake 2: Citing Outdated Dosing or Scheduling Information<\/h3>\n\n\n\n

Dosing for checkpoint inhibitors shifted substantially when flat dosing replaced weight-based dosing. Pembrolizumab moved from 2 mg\/kg every three weeks to a flat 200 mg every three weeks for most adult indications, and later 400 mg every six weeks. Nivolumab added a flat 480 mg every four weeks option alongside the 240 mg every two weeks regimen.<\/p>\n\n\n\n

AI models trained on earlier clinical trial publications, news coverage, or package inserts from previous label versions may return outdated dosing schedules. A patient who cross-checks their oncologist’s recommendation against an AI answer and sees a discrepancy may delay treatment or question their care team. A clinician who uses AI for a quick reference check could receive stale data.<\/p>\n\n\n\n

Mistake 3: Confusing Drugs With Similar Names or Mechanisms<\/h3>\n\n\n\n

The checkpoint inhibitor class now includes pembrolizumab, nivolumab, cemiplimab, dostarlimab, durvalumab, atezolizumab, avelumab, and others. All block PD-1 or PD-L1. All are used in oncology. Their approval profiles, label language, safety warnings, and clinical data differ substantially by drug and by indication.<\/p>\n\n\n\n

AI models trained on web content that discusses these drugs collectively and interchangeably frequently conflate their profiles. Cemiplimab (Libtayo) has different approval status across CSCC, BCC, NSCLC, and cervical cancer compared with pembrolizumab, but a model asked about a PD-1 inhibitor for cervical cancer may blend data from both drugs without distinguishing them.<\/p>\n\n\n\n

Similar confusion occurs within ADC classes, PARP inhibitor classes, and CDK4\/6 inhibitor classes. Ribociclib, palbociclib, and abemaciclib have overlapping breast cancer indications but meaningfully different safety profiles, particularly around QT prolongation (ribociclib), neutropenia timing, and hepatotoxicity monitoring requirements.<\/p>\n\n\n\n

Mistake 4: Getting Black Box Warnings Wrong<\/h3>\n\n\n\n

FDA black box warnings represent the highest tier of safety communication. Getting them wrong in an AI-generated answer is not a minor factual error. It is a potential patient safety event and, under certain interpretations of FDA guidance on promotional communications, could implicate pharmaceutical companies if AI tools associated with them or trained on their materials are involved.<\/p>\n\n\n\n

Immune-mediated adverse reactions for checkpoint inhibitors carry extensive boxed warning language that differs by drug. Axicabtagene ciloleucel (Yescarta) and tisagenlecleucel (Kymriah) carry boxed warnings for cytokine release syndrome and neurological toxicities under REMS programs. AI models queried about these therapies frequently describe the drugs’ mechanism or efficacy without surfacing the REMS requirement or the severity of the warned toxicities.<\/p>\n\n\n\n

Mistake 5: Recommending Generic Substitution Where None Exists<\/h3>\n\n\n\n

Biologics are not generically substitutable in the same way as small molecules. Biosimilar interchangeability designations require specific FDA determinations that not all biosimilars receive. Yet AI models, when asked about cost-saving alternatives for drugs like trastuzumab (Herceptin) or bevacizumab (Avastin), sometimes recommend biosimilars as if interchangeable substitution were straightforward, without explaining pharmacy-level substitution rules, state law variation, or the distinction between interchangeable and non-interchangeable biosimilars.<\/p>\n\n\n\n

For branded small-molecule oncology drugs with generic competition, models sometimes recommend switching to a generic version when the brand and generic have different formulations, different bioavailability profiles, or where the branded product has a patient assistance program that makes cost comparison more complex than the model represents.<\/p>\n\n\n\n

Mistake 6: Providing Off-Label Information as If It Were Approved Use<\/h3>\n\n\n\n

Off-label use in oncology is common and clinically legitimate. NCCN guidelines routinely include off-label regimens with strong evidence backing. The problem is not off-label use itself. The problem is when AI systems describe off-label use without flagging it as such, making it appear that an unapproved use is FDA-sanctioned.<\/p>\n\n\n\n

Everolimus (Afinitor) is approved for specific subtypes of breast cancer, renal cell carcinoma, pancreatic neuroendocrine tumors, and TSC-related conditions. It has been studied off-label in a range of other tumor types. A model describing everolimus use in an off-label context without flagging its regulatory status conflates clinical investigation with regulatory approval in a way that could mislead patients and potentially attract FDA scrutiny if linked to branded content.<\/p>\n\n\n\n

\n\n\n\n

Which Oncology Drugs Get Mentioned Most in AI Search and Why It Matters<\/h2>\n\n\n\n

Checkpoint Inhibitors Dominate AI Share of Voice in Oncology<\/h3>\n\n\n\n

Keytruda (pembrolizumab) is the most broadly approved oncology drug in history by indication count. As of 2024, Merck’s flagship had over 40 FDA-approved indications. That breadth means it appears in AI responses across an enormous range of cancer-type queries. When someone searches ‘best immunotherapy for lung cancer,’ Keytruda appears. When someone searches ‘immunotherapy for melanoma,’ Keytruda appears. When someone searches ‘cancer treatment without chemo,’ Keytruda appears.<\/p>\n\n\n\n

That share of voice has brand value, but it also creates disproportionate exposure to hallucination risk. A model that gets a Keytruda detail wrong reaches more patients than one that gets a rare tumor drug wrong. Merck’s medical affairs team monitors these outputs, though the degree to which any pharma company has a systematic, real-time AI monitoring program varies considerably.<\/p>\n\n\n\n

How Opdivo Competes With Keytruda in AI Responses<\/h3>\n\n\n\n

Bristol Myers Squibb’s nivolumab (Opdivo) covers many of the same tumor types as pembrolizumab and was first to market in several indications. AI models trained on oncology literature often treat the two drugs as near-equivalent, which is partly clinically accurate but ignores meaningful differences in companion diagnostic requirements, combination approvals, and label language.<\/p>\n\n\n\n

When a clinician or patient queries AI for a PD-1 recommendation in first-line NSCLC, the model’s answer should distinguish pembrolizumab’s PD-L1 TPS threshold requirement from nivolumab plus ipilimumab’s TMB-H alternative pathway. Models regularly conflate these, producing answers that blend the two drugs’ data without preserving the regulatory and clinical distinctions.<\/p>\n\n\n\n

Tracking AI Mentions of Enhertu, Trodelvy, and the ADC Wave<\/h3>\n\n\n\n

Antibody-drug conjugates are the most rapidly evolving class in oncology right now. Trastuzumab deruxtecan (Enhertu, developed by Daiichi Sankyo and AstraZeneca), sacituzumab govitecan (Trodelvy, Gilead), and fam-trastuzumab deruxtecan-nxki are generating enormous clinical attention and patient interest.<\/p>\n\n\n\n

AI coverage of ADCs is inconsistent. Training data for these drugs is thinner than for checkpoint inhibitors because they are newer. The HER2-low and HER2-ultralow distinctions for Enhertu are recent and represent a paradigm shift that many general-purpose AI models have not fully incorporated. A patient with HER2-low breast cancer asking an AI about their options may not receive accurate information about Enhertu’s eligibility for their specific HER2 score.<\/p>\n\n\n\n

\n

‘AI-generated drug information is now a front-line patient touchpoint. By 2025, an estimated 40% of patients diagnosed with a serious illness will consult a generative AI tool before their first specialist appointment.’ \u2014 Accenture Life Sciences, 2024 Digital Health Patient Survey<\/p>\n<\/blockquote>\n\n\n\n

\n\n\n\n

Can AI Hallucinations Trigger FDA Regulatory Action Against Pharma Companies?<\/h2>\n\n\n\n

How the FDA Views AI-Generated Drug Information<\/h3>\n\n\n\n

The FDA has not issued a final guidance document specifically addressing pharmaceutical company liability for AI-generated content as of mid-2025. It has issued draft guidance on prescription drug promotion on social media, on the use of AI in drug development, and on generative AI more broadly within FDA’s own operations. The intersection of third-party AI systems, manufacturer promotional content, and patient-facing AI tools remains in regulatory gray territory.<\/p>\n\n\n\n

What is clear is that FDA’s OPDP (Office of Prescription Drug Promotion) has historically taken a broad view of manufacturer responsibility for promotional materials. If a pharmaceutical company trains an AI tool on branded content, deploys an AI chatbot on a drug’s website, or sponsors an AI-assisted patient support program, existing promotional guidance applies. Hallucinated safety omissions in that context would likely constitute a violation of 21 CFR Part 202.<\/p>\n\n\n\n

Real FDA Warning Letters Involving Digital Drug Information<\/h3>\n\n\n\n

FDA OPDP has issued warning letters addressing digital and social media channels, though not yet specifically targeting AI-generated content. Warning letters to Duchesnay (Diclegis), Horizon Pharma (Ravicti), and multiple other manufacturers over the past decade addressed inadequate risk disclosure in digital formats, misleading banner ads, and insufficient safety information in character-limited formats like Twitter.<\/p>\n\n\n\n

The regulatory logic in those letters applies directly to AI outputs. An AI chatbot that mentions a drug’s efficacy without surfacing the required safety information fails the same fair balance test that got previous companies into trouble. Brand teams that do not monitor what AI systems say about their drugs cannot demonstrate awareness of potential violations, let alone remediate them.<\/p>\n\n\n\n

Pharmacovigilance and AI: What the EMA Has Said<\/h3>\n\n\n\n

The European Medicines Agency published a reflection paper on the use of artificial intelligence in pharmacovigilance in 2023. It noted that AI systems that analyze patient-reported information, including social media and digital health data, could support signal detection. It did not address the inverse problem: AI systems generating inaccurate drug information that could suppress or confuse adverse event reporting.<\/p>\n\n\n\n

If a patient consults an AI and receives incorrect information about a drug’s side effect profile, they may not recognize or report an adverse event. If the AI minimizes a known toxicity, the patient may attribute a serious symptom to something else. That breakdown in adverse event recognition represents a pharmacovigilance gap that regulators in both the U.S. and Europe have not yet quantified.<\/p>\n\n\n\n

Off-Label AI Promotion: The Emerging Compliance Risk<\/h3>\n\n\n\n

FDA prohibits pharmaceutical manufacturers from promoting approved drugs for unapproved uses. AI systems that associate a manufacturer’s drug with off-label indications, or that describe off-label uses as if approved, could constitute off-label promotion if the AI tool is in any way connected to the manufacturer.<\/p>\n\n\n\n

The risk is not limited to manufacturer-owned AI tools. If a pharma company’s medical affairs team uses an AI-assisted content generation tool to draft website copy, and that tool inserts an unapproved indication based on training data, the resulting content is still the manufacturer’s promotional material. The AI’s role in generating it does not transfer liability away from the company.<\/p>\n\n\n\n

\n\n\n\n

How Patients Actually Ask AI About Their Oncology Drugs<\/h2>\n\n\n\n

The Real Queries Cancer Patients Type Into ChatGPT and Perplexity<\/h3>\n\n\n\n

Patient query patterns in oncology AI searches differ substantially from the clinical information needs that most pharmaceutical companies optimize their content for. Patients do not search ‘pembrolizumab prescribing information.’ They search:<\/p>\n\n\n\n