Explorer

Anyone Remember Watson?

Jan 25, 2026

OpenAI and Anthropic both made healthcare announcements this month, which means it is time for everyone to have a very normal and measured conversation about AI and medicine.

Just kidding. Obviously.

OpenAI announced ChatGPT Health, a dedicated health experience inside ChatGPT. According to OpenAI, more than 230 million people ask ChatGPT health and wellness questions each week, which is a wild number even by current AI-number standards. TechCrunch covered the announcement and noted that OpenAI says Health conversations will be separated from normal chats and not used to train its models.

Then Anthropic announced Claude for Healthcare, aimed at providers, payers, and patients. That announcement leaned more into workflows like prior authorization, research, report generation, and connecting Claude to healthcare datasets like PubMed, ICD-10, CMS coverage data, and NPI data.

And then TechCrunch followed up with a piece asking the question I think is probably the right one: doctors think AI has a place in healthcare, but maybe not as a chatbot.

I have… thoughts.

TL;DR Takeaways

People are already asking chatbots health questions, so pretending this is not happening is not a serious answer.
A more private, health-specific experience is probably better than people dumping symptoms and lab results into a generic chat window.
The highest-value use cases may not be “AI doctor,” but boring workflow stuff: summarizing records, preparing appointment questions, reducing paperwork, prior authorization, and helping clinicians find the right information faster.
This still needs guardrails. Healthcare is not the place for “move fast and hallucinate things.”
We should remember IBM Watson before we declare that the robot doctor has arrived.

The Chatbot Problem

I am pretty sympathetic to the argument that these products are responding to existing behavior, not creating it from scratch.

People already Google symptoms. People already ask Reddit. People already send screenshots of lab results to group chats and ask their friend who took two biology classes in college what it means. Now they are asking ChatGPT and Claude (and probably Gemini?).

That does not mean it is safe. It just means the consumer behavior is real.

TechCrunch quoted Dr. Sina Bari describing a patient who came in with a ChatGPT answer claiming a medication had a 45% chance of pulmonary embolism. The statistic apparently came from a niche tuberculosis-related paper that did not apply to that patient. That is the nightmare version of AI health advice: technically sourced from somewhere, confidently presented, and totally wrong for the person sitting in the exam room.

At the same time, Dr. Bari still told TechCrunch he was excited about ChatGPT Health because people are already doing this, and formalizing it could add privacy protections and safeguards. That is basically where I land too.

If 230 million people are asking health questions every week, then the answer cannot just be “don’t.” It has to be “how do we make this less dangerous?”

The Boring Use Cases Are Probably the Good Ones

The part of this that seems most promising to me is not replacing doctors. It is removing some of the nonsense that keeps doctors from doctoring.

TechCrunch points out that administrative work can take up a huge chunk of a primary care physician’s time. Anyone who has dealt with healthcare bureaucracy for more than eight minutes can believe that. Forms, summaries, prior authorization, insurance rules, referrals, chart review, more forms, slightly different forms, and then, as a treat, a portal message that says “please allow three business days for a response.”

This is where Anthropic’s healthcare announcement is more interesting to me than the “ask a chatbot about your symptoms” version. If Claude can help with prior authorization packets, summarize a chart, pull relevant guidelines, or draft documentation that a clinician reviews, that seems much more grounded.

It is still risky. You still need audit trails, human review, clear responsibility, privacy controls, and a workflow that does not quietly turn “AI suggestion” into “automatic decision.” But at least the job is narrower.

We Have Been Here Before

The thing I cannot stop thinking about is IBM Watson.

I remember Watson beating people at Jeopardy back in 2011. It was a huge cultural moment (for us nerds at least). This computer could understand natural language clues, buzz in, and beat Ken Jennings and Brad Rutter. For a minute there it felt like we were watching the computer from Star Trek beat people at a game show.

But Watson was not supposed to stop at Jeopardy. The game show was the demo. The pitch afterward was that this kind of system could move into real-world decision support, including cancer care. It was going to digest medical literature, understand patient records, recommend treatments, and generally become the next great doctor-adjacent superbrain. Which, in the abstract, sounds like one of the best support use cases for AI to me. AI can retrieve and parse massive amounts of information that no human brain ever could. It could, theoretically, be aware of every possible drug interaction. It could, theoretically, see trends and patterns across multiple disparate studies. No human doctor can memorize every drug interaction. No human doctor can read every single study and medical trial and then connect the dots to some other random study they read 8 months ago.

STAT published a long investigation in 2017 arguing that IBM had pitched Watson as a revolution in cancer care, but it was nowhere close. The problems were not small. Medical records were messy. Local standards of care varied. The system depended heavily on human training. Its recommendations reflected the institutions and assumptions baked into that training. It was hard to update. It was hard to integrate. It was hard to prove that it improved outcomes.

In other words: medicine turned out to be harder than Jeopardy.

What Is Different This Time?

To be fair, today’s AI systems are not Watson. LLMs are much better at language, summarization, document handling, and conversational interaction than the older expert-system-ish Watson era. The infrastructure is better. The user behavior is already there. The tooling around retrieval, citations, connectors, and privacy controls is more mature.

But the hard parts of healthcare did not disappear.

Medical records are still messy. Clinical context still matters. Patients still have weird edge cases because bodies are inconsiderate like that. Insurance is still a Kafka novel with billing codes. Regulations still matter. Privacy still matters. And “the model sounded confident” is still not the same thing as evidence.

So my cautious optimism is conditional.

I can see AI being genuinely useful if it is used to help patients prepare for appointments, understand plain-language explanations, organize their own records, and ask better questions. I can see it helping clinicians move faster through documentation, chart review, coding, prior authorization, and research. I can see it making the healthcare system slightly less hostile to normal people who do not speak fluent Portal Message.

But I am much less excited about anything that looks like a general-purpose chatbot playing doctor.

The difference matters. “Help me understand this lab result so I can ask my doctor better questions” is very different from “tell me what treatment I should choose.” “Draft this prior authorization for clinician review” is very different from “deny this claim automatically.” “Summarize this chart” is very different from “diagnose this patient.”

The closer AI gets to diagnosis, treatment, coverage decisions, or triage, the higher the bar should be. Not vibes. Not demos. Evidence.

My Current Read

I think AI probably does have a real place in healthcare. Maybe a big one.

But the lesson from Watson is that healthcare does not care how impressive your demo was. A system can win Jeopardy, look magical on stage, and still fall apart when it meets real clinical workflows, incomplete records, institutional variation, safety requirements, and humans who are sick, scared, busy, or all three.

The useful future here is probably less “robot doctor” and more “boring assistant that reduces friction while keeping humans responsible.”

That is not as flashy. Nobody is putting “prior authorization copilot” on a Super Bowl ad unless things get very weird.

But if it means doctors spend less time fighting paperwork, patients get better explanations, and clinicians can find the right information faster without handing medical judgment over to a black box, that is a future worth being cautiously optimistic about.

Just maybe keep Watson’s ghost in the room while we build it.