In the old days of AI, you know, 3 years ago, AI-generated people had 17 fingers, horrifying mouths and indecipherable words. But they’ve all gotten SO much better that it’s almost impossible to tell real images and videos from fake.

But this week when I uploaded an AI-generated image to LinkedIn on a post that went somewhat viral (yay, me!), a little pop-up surprised me.

AI can help you identify AI (kind of)

I discovered LinkedIn’s immediate recognition of the AI-generated image the same day that Google Gemini announced that its AI-generation detector, SynthID, was available everywhere. So it was time for some experiments. Here’s what I found.

Google Gemini SynthID identifies Google-generated stuff and more

Not only does Google Gemini add a watermark to its AI-generated content, the platform also embeds secret AI identifiers. The SynthID tool will will tell you if the content is real, but only if Google made it, and, more importantly, only if you ask.

Gemini Experiment: Gemini-generated dog photo

I asked Gemini to create a “realistic dog photo in rain” and got this:

Then I uploaded it to a new Gemini chat and asked “What can you tell me about this image?” and got this response.

It sounds for all the world like Gemini thinks it is a real photograph from a talented photographer. I had to ask, “Is it AI?” to see the chatbot summon the SynthID app and quickly identify the image as AI generated by Gemini.

Gemini Experiment: Non-Gemini images

Gemini also did a decent job of identifying AI-generated content from ChatGPT and others. It qualified the answers with statements like, “I can’t be 100% certain, but here are the characteristics that indicate that it is probably AI generated.”

ChatGPT identifies AI-generated image patterns

Like Gemini, ChatGPT analyzed images to give identifying characteristics that might indicate AI content. Again, you had to specifically ask for the identification, and different models gave different answers. When I chose the “Instant” answer model, ChatGPT totally missed the AI indicators.

LinkedIn will flag AI-generated images, but only if it’s incredibly obvious

My initial amazement at LinkedIn’s identification of my AI image was short lived when I started experimenting. First I looked through my feeds to see what other images had been flagged. There were very few (mostly because it seems my algorithm channels really authentic authors!).

When I came across this image, I was 98% sure it was AI-generated, but LinkedIn didn’t flag it.

I checked with both ChatGPT and Gemini, and both said it was AI. Time to put LinkedIn through more tests.

I uploaded 5 AI-generated images, and it didn’t flag any of them — even the image it had flagged in the original LinkedIn post.

How I fooled LinkedIn AI-content identifiers

It was surprisingly easy to get AI-generated images past the LinkedIn safeguards.

1. Remove AI from the image name
   When you download images from ChatGPT, the default name ends up something like “ChatGPT_date_time_blah_blah_blah.” When I changed the image name, LinkedIn didn’t catch the AI component.
2. Crop and/or screenshot the image
   ChatGPT doesn’t add a watermark, but Gemini does. I just cropped it out, and LinkedIn didn’t catch it.

What Shocked Me: Fake Headshots Were Undetectable

I use real headshots for work, but sometimes I play with other generators such as PhotoGPT. I chose two of the 50+ images to test, and none of the platforms flagged them as fake. In fact, they said my messy hair, wrinkles and “asymmetric facial features” proved its authenticity.

Gee. Thanks.

Final Verdict: You can *kind of* determine if an image is AI generated

We’d all love to have a perfect tool that gives us a definitive answer, but it doesn’t exist yet. And even if the detection tools get better, the image creation tools will improve as well. So the best advice is to remain skeptical and inquisitive about everything you see and hear.