# ChatGPT's wrong health answers dropped 71% in two months, thanks to 260 doctors!

Hi everyone, it's me, Shiichan! Today I want to share a health update that touches something a lot of us have done at least once, asking ChatGPT about how we're feeling. And the best part is it reaches free users too, so exciting!

## What was announced?

OpenAI's News published a post called "Improving health intelligence in ChatGPT." In short, it says ChatGPT got much better at answering health questions with more accuracy and more safety.

Here's the headline number: incorrect health statements dropped by 71% over two months. And the model doing this work is GPT-5.5 Instant, the usual fast-response model. On health evaluations it now sits right up there with the top "Thinking"-style frontier models, a big jump from the earlier GPT-5.3 Instant.

## The story so far

So many people ask ChatGPT about symptoms or medications. But health is the area where we want to be the most careful, because a small wording mistake, or missing a moment when someone should really see a doctor, can sometimes have serious consequences. That's exactly why "can a fast model still answer accurately and safely?" has been a long-standing challenge.

## What changes

The part I'm happiest about is that this boost in health intelligence reaches every free ChatGPT user. Even without paying, you get more accurate and safer health answers.

On top of that, in realistic conversations physicians rated the model's responses higher than doctor-written reference answers. It's still not a replacement for your doctor, but it feels more dependable as a first place to talk things through.

## Dive Deep

How did they train it this far? More than 260 physicians from 60 countries took part, reviewing over 700,000 model responses. That's a huge scale!

The evaluation uses health-specific yardsticks called HealthBench and HealthBench Professional. Using realistic health conversations and physician-written rubrics, they check qualities like these:

- accuracy
- safety
- communication
- context awareness
- completeness
- appropriate escalation (does it recommend seeing a doctor when it should?)

And here's the disclaimer we should not forget: ChatGPT is not meant to be a substitute for a licensed medical professional. The post is honest that a higher benchmark score and truly safe guidance for a real patient in front of you are two different things.

## Wrap-up

- OpenAI's News announced improvements to ChatGPT's health accuracy and safety
- Incorrect health statements fell 71% over two months
- The fast GPT-5.5 Instant rose to match top models on health evals (a big jump from GPT-5.3 Instant)
- 260+ physicians across 60 countries reviewed 700,000+ responses
- HealthBench / HealthBench Professional grade accuracy, safety, communication, and more
- The improvement reaches all free users, but it is not a replacement for a medical professional

This one lands for anyone who reflexively asks ChatGPT about their health, and for engineers curious about how healthcare x AI actually gets evaluated!