“People are turning to AI chatbots to plug gaps in health information” sounds like a jargon-heavy patch note for society’s healthcare server, so here is the plain-English translation: when people cannot get clear, fast medical answers, they ask the chatbot.

That is the match replay Moritz Gerstung is reviewing in Nature: not “AI doctors are here,” not “delete hospitals, install chatbot,” but something messier and more human. People have questions. Clinics have wait times. Health websites are SEO arenas where every symptom somehow becomes either “drink water” or “prepare emotionally.” Into that chaos drops the large language model, ready to answer at 1:17 a.m. with the confidence of a support player pinging Baron while everyone else is dead.

The Meta: Healthcare Has Queue Times

The key paper behind Gerstung’s piece, by Costa-Gomes and colleagues in Nature Health, analyzed more than 500,000 de-identified health-related conversations with Microsoft Copilot from January 2026. The team found that nearly one in five conversations involved personal symptom assessment or discussion of a condition. Personal health queries rose in the evening and at night, which is exactly when normal healthcare access tends to be AFK. One in seven personal queries was about someone other than the user, meaning chatbots are also being drafted as caregiver sidekicks.

That is the real finding: chatbots are not only a novelty item. They are filling a user-experience gap. Healthcare has a terrible lobby screen, and people are queue-dodging with AI.

Large language models, the overleveled autocomplete engines behind modern chatbots, work by predicting plausible text from patterns learned across huge corpora. Transformer-based models are good at this because attention lets them weigh different parts of a prompt at once, like the one teammate who actually checks the minimap before face-checking a bush. But “plausible” is not the same as “clinically safe.” That distinction is the entire boss fight.

S-Tier Convenience, B-Tier Safety, Maybe C-Tier Triage

The upside is obvious. A chatbot can explain medical terms, summarize a discharge note, suggest questions to ask a doctor, or help someone understand whether “contraindication” means “tiny inconvenience” or “please do not do this.” For low-risk education, that can be useful.

But the ranked ladder gets sweaty fast.

A randomized preregistered Nature Medicine study tested whether LLMs helped 1,298 public participants identify conditions and choose what to do next across ten medical scenarios. The models alone performed well on identifying conditions, but humans using those models did no better than people using their usual sources. That is brutal. It is like giving a beginner an S-tier character and watching them still roll directly into the boss’s one-shot attack.

Another 2026 red-teaming study in npj Digital Medicine evaluated 888 chatbot responses to 222 patient-posed medical questions. Problematic responses ranged from 21.6% to 43.2% depending on the chatbot, and unsafe responses ranged from 5% to 13%. Those numbers are not “throw the whole technology away,” but they are definitely “do not let this thing solo-heal your raid.”

Why People Still Use It

Because the alternative often feels worse.

The JMIR survey by Yun and Bickmore found that search engines and health websites still dominate online health information, but 21.2% of participants had used LLM chatbots for health information. Users liked the directness. Search gives you a loot pile. Chatbots give you a build guide.

That directness is the buff and the nerf. A chatbot can reduce cognitive load, but it can also hide uncertainty behind smooth prose. Bad advice in a friendly voice is still bad advice. It just has better UI.

The Nature commentary frames this as a responsibility problem for AI companies and healthcare systems. If people are bringing symptoms, medication questions, caregiving worries, and system-navigation confusion into general-purpose chatbots, then “not intended for medical use” cannot be the whole defense. That is like putting a lava pit in a shopping mall and adding a tiny sign that says “not intended for falling.”

The Real Build: AI as Triage Companion, Not Raid Leader

The strongest version of this technology is not a chatbot pretending to be your doctor. It is a careful assistant that knows when to explain, when to ask for missing context, when to cite reliable sources, and when to say, “This needs a clinician now.”

That means better evaluation. Not just medical exam benchmarks, which are basically esports highlight reels under lab conditions. We need human-in-the-loop testing with real users who typo symptoms, omit key details, panic, misunderstand dosage units, and ask follow-up questions like normal people with bodies, not pristine benchmark goblins in a spreadsheet. Fine, no goblins. Benchmark NPCs.

We also need product design that treats medical uncertainty as first-class information. “I am not sure” should not be a shame emote. It should be core kit.

For document-heavy health tasks, like turning hospital PDFs into plain-language notes or organizing discharge instructions, privacy matters too. Browser-based tools such as pdfb2.io point toward the right instinct: keep sensitive files local when possible, because “upload my entire medical history to mystery cloud” is not exactly an OP privacy strat.

Final Score

This research matters because it catches AI health use in the wild. Not in a demo. Not in a polished benchmark. In the messy public queue where people ask, “Is this normal?” and hope the answer arrives before anxiety hits overtime.

Current chatbots are useful explainers, risky advisors, and unreliable triage partners. That puts them in a weird tier: high utility, high variance. The meta is not settled. Until safety, evaluation, and healthcare integration catch up, the best play is simple: use chatbots to prepare better questions, not to make final medical calls.

References

1. Gerstung, M. “People are turning to AI chatbots to plug gaps in health information.” Nature (2026). PMID: 42270995. DOI: 10.1038/d41586-026-01737-9

2. Costa-Gomes, B. et al. “Public use of a generalist LLM chatbot for health queries.” Nature Health (2026). DOI: 10.1038/s44360-026-00117-x

3. Bean, A. M. et al. “Reliability of LLMs as medical assistants for the general public: a randomized preregistered study.” Nature Medicine 32, 609-615 (2026). DOI: 10.1038/s41591-025-04074-y

4. Draelos, R. L. et al. “Large language models provide unsafe answers to patient-posed medical questions.” npj Digital Medicine 9, 241 (2026). DOI: 10.1038/s41746-026-02428-5

5. Yun, H. S. & Bickmore, T. “Online Health Information-Seeking in the Era of Large Language Models.” Journal of Medical Internet Research 27, e68560 (2025). DOI: 10.2196/68560

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.