Methodology doesn't support the conclusion.

Try using images that it wasn't trained on.

I would have to agree, this feels like fake science. They don't even seem to really present a reasonable initial problem.

This seems like a man / women / skinny / fat detector.

If you just classified people into those categories I wonder if you would achieve a similar level of accuracy.

It would be great if people stopped submitting misleading ads for the AI industry.

I just fed 4o and o4 mini a photo of my torso and gave it my height and asked it to estimate my body weight and body fat %. It was 20-40 lbs too low, so way off.

So that is obviously a significant relationship between ChatGPT estimate and the Henselman Guide BFP numbers. But just being a significant relationship doesn't necessarily mean it's useful. I'd bet that BMI (which the author repeatedly mentions is outdated and not very useful), is almost certainly also significantly correlated with BFP values.

Eyeballing that plot, it looks like ChatGPT estimates are +/- 10%. I'm wondering what the cutoff for usefulness is. I wouldn't' be surprised if BMI is that level of correlated or better.

The error was analyzed in the text.

Men: median signed error: +0.8%, median absolute error: 2.4%

Women: median signed error: +3.5%, median absolute error: 5.7%

That doesn't do anything to address my concerns. I was talking about the total spread, not the mean error, but in either case, a comparison to how well BMI correlated would be needed to address the issues I raised.

Dexa is expensive, you can get a modern smart scale for 50€, the cheapest accurate method would be the Caliper tongs.

What can it be good for? Video surveillance with metadata enrichment? Profiling? Losing body fat takes time. A smart mirror that displays body fat?

Just by looking at the graph I know that this will be rubbish - because of the line

When you draw a line this means that your model expects a linear relationship between the variables. Otherwise this is either a feel good line, or incompetence.

Given the images and text are definitely in the model training set, I'm surprised the results are this inaccurate.

Can AI estimate psychological profile based on a photo of the face (face structure)?