According to this semi-viral post, ChatGPT can analyze four photos of someone and estimate their height within an inch, based on the background. So, are we at a point where AI can reliably perform basic photogrammetry tasks?

To find out, I fed ten photos of myself with varying context into ChatGPT and Claude. I also tried Gemini but was told “I can't help with images of people yet.” Several of the images provided straightforward context, like a door, pickup truck, or even a scale bar. See images below.

Despite the informative context, both tools consistently provided ridiculous results and occasionally arrived at something logical. For instance, ChatGPT was provided photos 4 (without the scale bar), 5, 6, and 8, ultimately estimating that I'm 7’4” based on 4. When told that didn’t make sense, it switched to photo 5 and estimated my height as 5’1”. For the record, I’m a little over 6’1” in those cushy vans.

In another round, ChatGPT was provided three photos (including 4 with the scale bar) and quickly determined I was 6’1” after automatically detecting the scale bar. See the entire response here. However, during other sessions ChatGPT didn't detect the scale bar and when specifically asked about the scale, indicated it couldn't read the related markings.

In another round, ChatGPT used photos 1,5,7, and 8 to estimate my height as 5’10” to 6’2”. No scale, just basic context. OK, nice work robot, nice work.

Claude was provided four photos (including 4 with the scale bar) and estimated my height as 6’6” when informed the double door in photo 2 is 94” tall. Claude was then provided with photo 4 and asked, “Based on the scale bar next to this person, how tall are they?” It determined I was 5’10” to 5’11” tall. Even a nap-deprived kindergartener could glance at the image and tell that’s wrong.

Overall, I’m not impressed and find it hard to believe the results of the original poster. I think, “Even a blind squirrel finds a nut now and again,” is a proper summation. Having said that, these tools were not designed for such purposes and I don’t think it’ll be long before AI plays a part in forensic photogrammetry and video analyses… we’re just not there yet.

Keep exploring,

Lou Peck
​Lightpoint | Axiom

P.S. If you want to learn how to accurately determine speed from video or model crush and scene evidence using only photos, an online version of my photogrammetry course (with daily office hours) was just launched. Five seats are still available.