I want to start with a number. A small one. 20.

Twenty milligrams per decilitre. That's the difference between what Google's AI Overview told people about post-meal blood sugar thresholds for diabetes — and what the Indian Council of Medical Research (ICMR) actually recommends.

AI said: under 180 mg/dl. ICMR says: under 200. A 20-point difference. Not dramatic. Not obviously wrong to someone reading quickly. Which is, as a McGill University professor put it, precisely the danger.

That sentence should follow everyone building with AI into every meeting they walk into. Because the 20mg/dl error isn't the story. The story is how it got there.

---

The audit that should alarm all of us

In June 2026, a healthcare content consultant audited nearly 500 articles published by five of India's top hospital chains — Apollo, Max, Fortis, Medanta, and Artemis. The findings, shared with The Ken, were startling.

Articles recommended discontinued drugs. Carried outdated emergency protocols. Applied Western clinical benchmarks to Indian patients — a population with meaningfully different metabolic baselines. And many carried unmistakable markers of AI writing.

---

This isn't just a healthcare problem

Healthcare is where the consequences are most visible — a 20mg/dl error has a name, a patient, and a treatment decision attached to it. But the same loop is running in every domain built on content. Here's the mechanic, stripped down.

In AI research, this is called model collapse — when models trained on AI-generated data begin to degrade iteratively. Not dramatically. Not immediately. The training signal gets weaker. The edges of knowledge blur. The model optimises for what sounds right rather than what is right. We're watching the same thing happen to human discourse.

---

Where else is this loop running?

Think about any domain that runs on content — on written knowledge that gets published, indexed, cited, and built upon. The hospital case is just the one we caught, with a number attached to it.

In media and communications — my own domain — the error surfaces more slowly and more quietly. Not as a 20mg/dl difference. As a narrowing of what questions get asked, what narratives feel legitimate, what framing seems "standard." The language gap in Indian regional media makes this especially consequential — over 90% of daily conversation happens in regional languages that most AI systems were never trained to understand deeply.

---

The question we should actually be asking

We spend enormous energy asking whether AI is accurate. That's the wrong first question.

The better question is: whose narrative is AI reinforcing — when the training data is increasingly a mirror of itself?

AI is not just a tool for answering questions. It is becoming a mechanism for defining which questions feel worth asking. It shapes the range of what feels possible, what feels credible, what feels like the right frame for a conversation. And it's doing this silently — not through malice, not through conspiracy, but through iteration.

The cost of not knowing

The 20mg/dl gap in a diabetes guideline is the loop made visible — concrete, measurable, falsifiable. Most of the time, the loop is invisible. And that's the version we should worry about.

For organizations operating in India's media landscape, the stakes are particularly high. AI media intelligence that's built on a corrupted training loop doesn't just miss stories — it actively misrepresents them. The PR monitoring infrastructure that most brands rely on was never designed for this. Neither was the media measurement framework built around it.

The loop nobody designed is running. The question is whether you're inside it — or watching it from a position of genuine intelligence.