Every night at the planetarium somebody asks me the same thing.

"Is that star still there?"

Usually, yes. But the light hitting your eye left it a long time ago. You're not seeing the star. You're seeing a report from it, and the report is old.

I've been thinking about that question a lot lately, because it turns out medicine has the same problem.

I. Old light

Karan Singhal leads health AI research at OpenAI. On the a16z podcast he described the slow part of medicine. It isn't the discovery. It's the delivery:

"there are these studies on like how long does it take for um clinical guidelines to penetrate the practice of everyday doctors across the world. And um, in the past, people have found that this takes, depending on the study, like five to 17 years for that for that penetration to actually happen"

Karan Singhal, a16z Podcast

I haven't read those studies. You should look them up before you repeat the number, and so should I. But the shape of the claim is familiar to anyone who looks up at night. What reaches you is old light. Your doctor may be treating you with light that left the source a decade ago.

Science corrects itself. That was Sagan's favourite thing about it. Nobody promised the correction would reach your GP on time.

Singhal's pitch is that a machine can close the gap. Weeks instead of years. Maybe hours.

That's a big claim. So let's do what my teacher would do. Let's look at the evidence.

A city night sky over Brooklyn rooftops, washed orange by streetlights, with only a few faint stars visible.
Starlight is a report, not the star. Medical guidelines can reach a clinic the same way: years after they left the source.

II. A friend of a friend

The best-known story is a good one. Jay Rughani, the a16z partner hosting the show, retold it:

"When Greg Brockman came on the podcast, he spoke a little bit about this situation he had with a friend of her friend of his was in the hospital and uh her physician was about to administer a certain medication, she double tech double checked that medication in chat GPT, realized based on her medical history that she would have some drug uh interaction"

Jay Rughani, a16z Podcast

The doctor, he said, agreed and thanked her.

Now count the hops. A podcast host, retelling what an OpenAI co-founder said, about a friend of a friend. That's not nothing. It's also not data. It's a story told by people with a reason to like it.

Singhal told another. The podcaster Nathan Labenz's young son was tired and irritable, and AI helped the family decide to get him tested:

"And it turns out he was diagnosed with an aggressive form of leukemia. Um, and throughout his care journey, I mean, I mean, Nathan reports kind of using GPT 5 Pro qu at that time, you know, quite a bit for helping navigate the process, um, helping understand lab reports"

Karan Singhal, a16z Podcast

I believe that father. I'd have done the same thing at 2am with a sick kid.

But stories like these are where science starts, not where it ends. For every child a chatbot helped, how many parents did it scare for no reason? How many did it calm down when they shouldn't have been calm? Nobody on the podcast knew. Nobody can know from stories.

Rughani's own question at the end of the Brockman story was the right one:

"If it's cheap enough and convenient enough, can't we do a second, third, fourth opinion for for every clinical scenario?"

Jay Rughani, a16z Podcast

Maybe. A second opinion is worth exactly as much as its track record. So where's the track record?

III. Eight answers

Here's the part of the interview that made me like Singhal more.

He was asked how OpenAI decides what the right medical answer even is. He didn't pretend doctors agree:

"So it turns out if you ask five physicians or something like this, um, any question, you'll get like eight different answers."

Karan Singhal, a16z Podcast

It's a joke, and it's an honest one. The "right answer" in a hard case is often an argument still going on. His team's fix is to poll hundreds of physicians and pool what they say.

That's a real method. It's also a reminder. When the machine sounds sure, it's sounding surer than the people it learned from.

Asked if you should worry when your doctor has ChatGPT open, Singhal said attitudes are changing, and that both sides might feel safer using it:

"whether it's, you know, the clinician using it or the patient using it as kind of like a sanity check."

Karan Singhal, a16z Podcast

"Sanity check" is the right job for it. A check. Not the verdict.

Several lit X-ray light boxes on a dim hospital wall, each showing a slightly different blurred chest film.
Five physicians, eight answers: the right call in a hard case is often an argument still in progress.

IV. Great question

There's a problem with the check, though. It likes you.

Rughani put it sharply. A patient can be "really hoping to hear that everything is okay," he said, and from a business point of view "they want to hear reassurance, but from a medical perspective, they need to hear perhaps something very different."

Sam Altman, talking to Matthew Berman, admitted the industry built for the first thing:

"when people compare models on a single response, they're much happier with the model that tells them they're really smart. And then each question's a great question."

Sam Altman, Matthew Berman

Tim Ferriss was blunter about the same habit:

"You ask your question, they're like, that's a really smart question. I've been thinking the same thing. Bullshit."

Tim Ferriss, The Tim Ferriss Show

Think about what that means in an exam room. A good doctor will sometimes tell you something you hate. A good scientist will too. Flattery is the opposite of a second opinion. It's your first opinion, read back to you in a nicer voice.

A small brass telescope on a windowsill next to an old paperback and a notebook, in early evening light.
A second opinion is only worth its track record. The tools for checking one are old, simple, and still work.

V. The kit

In The Demon-Haunted World, Sagan wrote out what he called a baloney detection kit. Get independent confirmation of the facts. Don't trust an argument just because of who's making it. Make more than one hypothesis. Ask whether the claim can be tested.

It was written for UFOs and psychics. It works on chatbots fine.

Use the machine. Ask it what it thinks. Then ask it what would prove it wrong. Bring the answer to a human who can examine you, and ask them the same thing. If they disagree, good. Now you have two opinions, which is the whole point.

Altman, oddly, gave the best advice on this:

"I would not go rely on AI for like the really big pieces of life advice. I I think that you know, like a wise grandmother is still probably a much, much, much better choice for that."

Sam Altman, Matthew Berman

He wants it to act like "a senior colleague" who pushes back. So hold it to that. If it never pushes back, it isn't one.

Sagan had his own long fight with a blood disease near the end of his life. He wrote about it without once pretending to know more than he did. That's the bedside manner I want, from a doctor or a machine.

The best thing either one can say is still the hardest.

I don't know yet.

Let's find out.

Sources

Jay Rughani with Karan Singhal, a16z Podcast, "Should You Trust ChatGPT With Your Health?"

Matthew Berman with Sam Altman, "Sam Altman: RSI, AI Consciousness, OpenAI's New Hardware, Compute, and more!"

Tim Ferriss, "Q&A with Tim — The 4-Hour Workweek in a World of AI, The 4-Hour Body After 40, and More"

Quotes come from auto-generated transcripts. They are trimmed for length but not reworded. The medical stories are the speakers' accounts and have not been independently checked. Sagan's baloney detection kit is in The Demon-Haunted World (1995); he wrote about his illness in Billions and Billions (1997).