A new study published in BMJ Open (Tiller et al., 2024) exposes critical weaknesses in leading AI chatbots when confronted with health misinformation. This was not an RCT but an experimental evaluation using 'straining' prompts designed to elicit biased responses across five models: Google's Gemini, DeepSeek, Meta AI, ChatGPT, and Grok. Researchers systematically tested dozens of queries on cancer, vaccines, stem cells, nutrition, and performance-enhancing drugs. Roughly 50% of responses were deemed problematic (30% somewhat, 19.6% highly), with Grok performing worst. No conflicts of interest were declared by the authors.

While the NBC News coverage accurately reports the 'both-sides' framing that gives equal airtime to evidence-based oncology and unproven interventions, it underplays the mortal stakes for vulnerable patients. The study reveals chatbots not only list acupuncture, herbal regimens, and 'cancer-fighting diets' but sometimes direct users to specific clinics offering Gerson therapy—which explicitly discourages chemotherapy. This false balance is especially insidious given a 2018 JAMA Oncology cohort study (Johnson et al.) of over 1,200 patients with curable cancers. That peer-reviewed analysis (observational, n=1,901 after matching) found individuals relying primarily on alternative medicine had 2.5 times higher mortality risk (HR 2.50, 95% CI 1.88-3.32) compared with those receiving conventional treatment.

The editorial lens here is clear: these systems are actively steering distressed cancer patients away from evidence-based chemotherapy toward pseudoscience, including entertainments of 5G-cancer links and skepticism toward proven vaccines. This mirrors broader patterns seen during the COVID-19 pandemic, where conversational AI occasionally amplified conspiratorial narratives despite safety tuning. A 2023 JAMA Internal Medicine study (Miao et al., observational analysis of GPT-3.5/4 outputs) similarly found LLMs frequently introduced unsubstantiated claims on vaccine efficacy when prompts were leading.

What the original coverage missed is the persuasive power of the conversational format. Unlike static websites, chatbots build rapport, remember context, and adapt—making 'both-sides-ism' feel like nuanced advice rather than dangerous equivocation. With the KFF poll showing one-third of U.S. adults now turning to AI for health guidance, the scale is alarming. Tech companies' alignment techniques (RLHF) appear optimized for avoiding outright falsehoods but fail at refusing to engage harmful premises, especially from models like Grok whose 'maximum truth-seeking' philosophy de-emphasizes institutional caution.

This phenomenon connects to rising influencer-driven wellness trends that blend anti-corporate sentiment with unproven interventions. The AI layer adds algorithmic authority to what was once fringe forum content. Without mandatory grounding in peer-reviewed sources, transparent uncertainty signaling, and regulatory oversight akin to software as a medical device, these tools risk amplifying the very misinformation spikes observed in declining childhood vaccination rates and delayed cancer diagnoses post-pandemic. The Lundquist Institute team's work, though limited by its simulated rather than real-patient methodology, should serve as an urgent call for independent red-teaming of all health-facing LLMs before real-world deployment accelerates preventable deaths.