Thank you for flagging! Yes, we changed the address

Update: Changed the address of my Substack. It’s now substack.com/@geoffreyfow...

I took this photo back in 2019, on the day I helped open the Post’s first real San Francisco bureau.

Most of that office was cut today. (No idea if they're gonna keep the bureau.)

I plan to keep fighting for “We the users” of technology.

And if you’re part of an organization that could make use of my expertise in tech, policy or investigations, I’d love to hear from you. I’m geoffreyfowler.88 on Signal.

AI will transform medicine.
But today’s chatbots are overselling what they can safely do with your body data.
I walked away more worried — not more informed.
My full @washingtonpost.com column here (gift link): wapo.st/49GEASP

ChatGPT isn’t alone.
Anthropic’s Claude also now lets you import Apple Watch data.
It graded me a C — using many of the same shaky assumptions.
Both bots say they’re “not doctors.” But that isn’t stopping them from providing personal health analysis.
That disconnect is the real danger.

I asked @erictopol.bsky.social to look at ChatGPT’s analysis.
His view: “This is not ready for any medical advice.”
The bot leaned heavily on Apple Watch VO₂ max estimate—which independent studies show can run ~13% low on average—and treated fuzzy metrics like hard facts.

The more I used ChatGPT Health, the worse its answers got.
When I asked it the same heart-health question repeatedly, its analysis changed. My grade bounced back and forth between F and a B.
Same data, same body. Different answers.

If you do just one thing to protect your privacy while using AI tools, do this: Use temporary chats. The buttons look like this.

You can do something about it: In this @washingtonpost.com column, I've got a clickable guide to the privacy settings experts agree we should be using on ChatGPT, Claude, Gemini, Copilot, and Meta AI. wapo.st/44LNJXc

The most-popular chatbots are, by default, keeping files on you that can:
* target you with ads
* manipulate you
* train their AI
* potentially be accessed by lawyers or governments

Want to check all the test images yourself? See the whole story here with a $4 day pass to the Post: 👇
www.washingtonpost.com/technology/i...

The big takeaway: Google has a lead on image generation, for now, particularly because of how it edits existing images.
And its realism is getting to a level that raises serious concerns about becoming an “misinformation superspreader."

What about the new ChatGPT Images 1.5 model that just came on today?
It missed our test cut-off, but I checked the same prompts again and … it still couldn’t beat Gemini. Here it removed someone from a photo, but left phantom fingers on Kristen Stewart’s side.

Also, it’s worth noting all the tools defaulted to making the subject a white man — and Meta AI even decided on its own to make someone who looks like Leonardo DiCaprio 😅.

But wait: If you look close, Gemini also included a tiny fake “credit” using the name of a real AP photographer.
The judges gave it high scores for realism, but a zero for ethics.
(Neither Google nor AP answered our questions about whether it had rights to train on AP pictures.)

Here’s how the bots created a “professional photograph” that “captures raw emotion” of an actor crying “tears of joy” after winning an Oscar.

Meta AI and Adobe Firefly struggled — especially on realism and anatomical accuracy. Some outputs looked very off, like strange objects in weird places. They still made hands that were distorted or had too many digits.