Can you distinguish traditional photos from generative AI?

Have you tried an AI image generator to make creatives?

You’re probably wondering: will it be realistic? Will it give me an impactful ad?

Sure, you can get that spatial opera theater or Darth Vader browsing a cereal aisle in a matter of minutes. But the truth is, making ads still requires more realistic images and photos taken in professional settings. An image of a dog with sunglasses will rarely do the job. 

Let’s begin with ad impact: the Memorable Ad Maker creates ads with AI using only a text prompt, and instantly tells users which ads will perform better. 

That means you can make sure that you have the right creative before you spend your media dollars. Quite important in a year when every dollar spent on ads has to bring additional revenue.

OK, so Memorable ensures creative effectiveness. What about photorealism? Here’s our hypothesis:

Many of the images generated by the Memorable Ad Maker will be impossible to differentiate from images coming from a traditional photo shoot.

To test it, we went on and made a little challenge: we showed people 4 images, where one was traditional photography and 3 were made by Memorable AI Ad Maker.

Before you look at the results, click here to play yourself!

Turns out, most people confused the images made by Memorable AI with a traditional photo. More than 70% of the votes went to AI-generated images. 

Did you guess correctly? The traditional photography was image #1 📷

Some of our generations are so realistic that most humans couldn’t tell the difference. Now it’s up to marketers to simply choose one with strong predictive KPIs and of their liking.

Since most people have trouble identifying AI content from traditional ones, it's only natural that developers started generating AIs to actually detect AI-created content. Even OpenAI has just released a classifier trained to distinguish between AI-written and human-written text.

This is not new for Memorable: our founder has published research with models that help detect deep fakes.

PS for the stats nerds out there

Statistics lovers might still be curious about how likely this outcome is in a context of equal images—i.e. where the AI is indistinguishable from the traditional photo.

We were too. So we did a fun additional exercise: we ran 1000 simulations of what people would choose if all images had equal chances of being considered real. That means generating 1000 universes in which the AI can't be detected by humans.

Then we looked at the probability of getting the results we got in the LinkedIn survey if we were under that scenario. So we estimated the likelihood of finding a sample where the traditional photo gets more than 13 votes.

As it turns out, that probability is 33%—the purple bars in the graph below, representing 330 of the 1000 simulations. That means that our survey results are not super likely in a context where the 4 pictures have equal chances of being considered real. 

Come try the Memorable Ad Maker.

Got questions? Want to get some advice or share your creations? Join our Memorable Innovators Community. You’ll get insights on AI in ad creative, as well as early access to our features.