I used to struggle to find suitable images for this blog. Then late last year, I had the opportunity to study the foundations of Artificial Intelligence (AI) thanks to Oracle. Equipped with that knowledge, I began to create AI generated images and things have gone uphill since then.
Recent technological advancements, especially in AI, has made it very easy to generate images. With the right text prompts, you can generate images that are a good reflection of your thoughts and intents. More importantly, these advancements, have made it difficult to tell between real and artificial images. Bad actors know this and have taken full advantage to commit cybercrime and spread misinformation.
What do real images look like?
Before we look at a few tell-tale signs of AI generated images, lets remind ourselves of what real images look like. Think of the last photo you took. What prominent features did it have? It most likely had consistent lightning. This means it had natural lightening, shadows and reflections that align with the light source.
Next it most likely had natural texture which created non uniform patterns or grains. Additionally, it most likely had correct geometry, that is, consistent straight lines especially in architectural scenes. With this in mind, let’s turn our attention to some dead giveaways of AI generated images.
Is it “too Perfect”
AI generated images are unable to reproduce many features of real images. There is often a mismatch between lightening on the face and the background or the use of fantasy or cinematic looking backgrounds. If there are human figures in the image, they will lack texture in the face and have oversaturated skin color. Additionally, there will be inconsistent alignment between these figures and their shadows.
Check the Human Anatomy
Diffusion models, which are designed to create AI generated images are excellent at creating big picture details of the human anatomy such as the body shape and size. Unfortunately, they are notoriously bad at small details such as the fingers, teeth and eyes.
The next time you wonder whether a human image is real or artificial, check if it has overlapping teeth, lifeless eyes or unnaturally shaped pupils. In addition, check if it has unnatural hand positions, fingers blending into each other, disproportional or irregular number of fingers.
Unreadable text in images
Another limitation of diffusion models is its inability to generate coherent text. This is because, unlike Large Language Models such as Chat GPT which are designed to understand and create human like text, diffusion models are designed to create high quality photos, videos and audios. When you zoom in on the text in the image below, you realize it is complete gibberish.
Verify the image source
Finally, a good way to check the validity of an image is to check its source. This becomes especially important when none of the tips above have proved useful. One way to do this is to do a quick check on the internet to identify other places the image appears on. In general, the same AI generated image will appear less frequently on many web sites. Lastly, verify the original source of the image. Confirm its coming from a verified social media account or credible website.
Take the Test
You are now equipped to identify AI generated images and less likely to fall for the tricks of bad actors. Always bear in mind the basic features of real images such as consistent lighting and texture. This will serve as a pointer to correctly spotting AI generated images. Also, don’t forget the limitations of diffusion models. They lack the ability to correctly illustrate the finer details in images such as buttons and buckles or generate text. Finally, if it’s too difficult to tell verify the source of image. This is a sure way to tell if it can be trusted.
Let’s put this knowledge to the test. Challenge yourself. Can you tell what is real and what is AI in this quiz? Leave your comments in the chat below.