I conducted an experiment yesterday to see if people could differentiate between a real cat and a Midjourney AI-generated cat. I posted the image comparison on LinkedIn and Facebook, asking for responses over 12 hours. The results from 49 respondents were fascinating.
55% of people correctly identified the AI cat (“A”), saying:
– There were too many whiskers
– The whiskers were too perfect
– the image was too dreamy
– “I’ve got a cat and it doesn’t look like that.”
37% of people thought the real cat (“B”) was fake, with some saying that something looked “off” about it. You’re right, I chose a Pinterest photo of a weird-looking cat to throw you off.
8% thought I was calling their bluff and said both cats had to be AI-generated. Nah, I’m not planning to lose my credibility here.
I devised this experiment (and a smaller one the previous day) after watching an online video that wanted to teach people how to identify AI images. The intention of the video was good, but the AI-generated photo it used was so obviously fake with poorly rendered body parts and blurry details.
I wondered, “What if I could render something with minimal visual mistakes? Could people detect it?” So, I set to work with Midjourney.
With the two cats, I managed to confound nearly half the respondents, but I thought I could have done better. And now I will be able to, thanks to the user feedback on Cat “A” that taught me which nuances I need to be aware of!
The user feedback also affirms my own teaching from my Gen AI workshops: We need to be subject matter experts to use Gen AI well, and to defend ourselves against malicious threats (eg. deepfakes, not cats). In other words, if you are an expert on cats, it’s not that easy to trick you…for now.
Thanks for all the responses, folks, I am now designing a more challenging quiz and making it more difficult for you to detect the next AI image.
You can read the votes and comments from the previous post here: https://lnkd.in/g-Sued69