Oct. 15, 2026
There’s a debate as to whether the average person can recognize AI writing on sight…
I recently heard Nick Richtsmeier (interviewed by Jonathan Stark) claim that research evidence shows we cannot reliably distinguish AI writing from human writing:
“Increasing evidence shows that you can’t tell…the average reader cannot tell the difference [between AI and human writing].”
I’m not sure what evidence he’s referring to, but a quick search shows that some researchers certainly do make this claim…
For example, Dongwon Lee claims that humans can identify AI-generated text with only ~53% accuracy (i.e., similar to guessing randomly).
A quick Google Scholar search for the author and topic did NOT reveal the source material. The closest thing I could find was Uchendu, Lee, Shen, Le, Huang, & Lee (2023), where they cite evidence that text messages can be hard to distinguish as human- vs. AI-generated. That sounds far more plausible to me than the claim that humans can’t spot AI text in general.
Similarly, Cheng et al. (2025) show evidence that humans detect AI-generated text with accuracy rates comparable to random guessing.
However, with a sample size of five participants, this is not exactly the world’s most persuasive finding.
(Hint: You cannot base conclusions about people in general on findings from five individuals alone. Maybe these five individuals just happen to suck at spotting at AI text?)
Now, why did I choose to pick on these two sources in particular when writing this post?
Because these are the two sources that Google Gemini cited when confidently informing me that humans CANNOT identify AI-generated text!
One more among a million reasons to be skeptical of AI output, even when it backs up its claims with research citations.
Granted, there might be better evidence out there that does show humans can’t spot AI-generated text.
I wouldn’t know, because I haven’t looked into it more deeply yet…
On the other hand, Ethan Mollick points to research reporting that people who use AI frequently CAN recognize its writing with extremely high accuracy.
For example, Russell et al. (2025) claim that individuals who use LLMs regularly were able to correctly classify text as human- or AI-generated with 99.67% accuracy.
Now, that result is so shockingly high that I’m skeptical of it as well! (Or maybe I’m just skeptical in general?)
Regardless, although my intuitions align with Mollick on this one, it ultimately doesn’t matter.
Because we can test this for ourselves…
Click here to take the 30-second test and prove if you can distinguish human writing from AI writing. =)
