Differentiating AI-Gen vs Photo

Sharing some tips on how to differentiate

Recently I came across this article from Bloomberg. In the article it commented that Google search engine used Generated AI’s images as part of its search results. This is not ideal as it may lead to disinformation and distort reality when images that are generated and not reflecting accurately of reality.

However there is a fun twist to the article, it actually includes a quiz of TEN pair of photos. Readers get to try their hands on differentiating which photo in each pair is generated by AI and which is the real photo. I am happy and lucky to say I scored a full score! But I have to say, there were a few pairs that was very tough to differentiate between. (If you will like to do the quiz first before getting to the rest of the issue, here it is.)

In this issue, I thought of sharing a few tips with you on how I (luckily) manage to differentiate the actual images vs Generated AI images.

a) Grainy vs High Definition

Generated Images are very high definition! You can that it is very smooth and clear as compared to an actual image. That was the deciding factor for me for the “Joan of Arc” pair in the quiz. If the complexion of human faces are too smooth and clear, there is a good chance it is AI-generated.

b) Symmetry

Generated images tend to have very good symmetry, both bilateral and radial. Human body , or should I say naturally created items are never symmetrical bilaterally, there will be some differences. Usually the symmetrically closer it is, as in the left and right side are very similar, it is a sign of good aestethic. So if you look at “Napoleon Bonaparte” pair, you will notice that the eyes are different (very slightly) for the real image/painting vs AI-generated. From what I understand, most people’s eyes are not really the same.

b) Repeat Items

Look at the background, especially if there are repeated items. In this quiz, “Winston Churchill” pair will be the best to talk about. <Spoiler> If you notice the background, books spines gets repeated across the bookshelves behind. Its pretty “lazy” of the AI to do that, but it also makes me wonder why that is the case. If any of my subscribers have any speculations can let me know. :)

Not in this quiz, but recently a friend of mine shared an image of phad thai and mango sticky rice.

The dead giveaway was that there were pieces of mango in the phad thai, however I did notice that the mango pieces are squarish, smooth round corners, or similar cut and just varies in sizes. The prawns which were supposedly heavier than the noodles and the rest can sit very nicely on top of it, which seems to defy the laws of physics, imo. All these are, to a certain extent, giveaways.

c) Numerous Varying Features

“Maya Angelou” pair come to my mind to explain this. Generated images try to get away with minimum features, meanwhile reality is cruel as they tend to add more features to the face as we age. So a real image actually has a lot more features rather than the bare minimum. But this factor has to be combined with the “symmetry” factor then you can make a better judgement imo.

Conclusion

These are some tips I can think of that has helped me to differentiate AI generated vs actual images through painting and photography. Having some domain knowledge on how computers behave vs humans/nature does help in a bit. I constantly ask myself what kind of training data is widely available, and how AI are trained to read them and the likely image characteristics it will generate based on the prompts.

We as humans, wants to avoid falling prey to misinformation or disinformation, so in an age of AI generated content in different medium, it is time we start innoculating ourselves by constantly learning and practice on how to question and differentiate AI generated images from reality.

I will love to hear some tips how you differentiate generated images vs actual images. Let me know your thoughts too! Please share them in the comments below. :)

Do consider supporting my work! You can make a “book” donation and drop me some wisdom! :)

Recommended Content

Bloomberg Article on Generated Images getting into Google search results. <Article>