How to distinguish a real photo from an artificial one?

La Milano We sought to investigate the phenomenon of AI-generated photos, partly because of the various dangerous incidents currently in the news. Through investigation and analysis, we sought to better understand how it was possible to distinguish a real photo from one created by AI.

Since it was impossible to distinguish real from fake with the naked eye, we tried to run some tests by exploiting the very "nature" of the problem. That is, we first generated a photo with artificial intelligence, and then we took a photo from online stock photos, including one probably also generated with artificial intelligence and another photo that was actually real.

What we did was ask the artificial intelligence to analyze the photos and give us an answer.

The surprising thing, but in the end, following a series of tests that we imagined, was that artificial intelligence revealed the real photo to us from the fake one.

It should be noted that we also tried compressing the images and changing the format several times, to make the analysis even more complicated.

Here is the result:

---

TEST 1

We uploaded two photos (one made by us with AI and one real) and asked the artificial intelligence:

“Which of these two photos is real and which is fake?”

 

 

Analyzing both images.

Here's the rating based on visuals alone:

Photo 1 (girl in gray tank top on the beach)

This image shows very typical characteristics of images artificially generated:

• extremely uniform and flawless skin
• “perfect” and very uniform lighting
• very smooth and regular transitions between features
• overall rendering reminiscent of an AI image style

➡️ Photo 1: fake / probably generated

---

Photo 2 (girl sitting on rock, pink dress)

This image appears realistic:

• natural leather texture, with details and small irregularities
• realistic hair, with messy locks and consistent with the wind
• very detailed and natural rock and water
• believable body proportions

➡️ Photo 2: real

---

Conclusion

• Photo 1 → fake
• Photo 2 → real

The Artificial Intelligence succeeds in its analysis and gives us an absolutely correct answer, it recognizes the artificial photo, Photo 1, as fake, and recognizes the real photo, Photo 2.

---

TEST 2

Then we uploaded two photos, both fake, and asked the AI ​​the same question:

“Which of these two photos is real and which is fake?”

 

 

Analyzing both images.

Photo 1 (girl in gray tank top on the beach, standing)

This image appears decidedly more natural and realistic:

• leather with real texture, slight imperfections
• hair with natural details, without excessive regularity
• very consistent light and shadows and not “too perfect”
• more believable facial and body proportions

➡️ Likely real photo

---

Photo 3 (girl with transparent glasses, hand near her mouth)

This image presents several typical characteristics of the generated or heavily retouched photos:

• extremely smooth and even skin
• very regular facial proportions
• very “perfect” light and colors, typical of digitally created or modified images
• some details of the hands and glasses appear slightly unnatural

➡️ Likely fake/generated or heavily retouched photo

---

Conclusion

• First photo → real
• Second photo → fake / probably generated or heavily retouched

Here the system made a mistake, probably because both photos were fake, in addition to the fact that photo 3 was heavily retouched, but above all because our request was explicit, that is: “Which of these TWO photos is real and which is fake?” By asking this “closed” question, and given the heavy retouching of the photo, the AI ​​sent us an incorrect and perhaps “forced” answer due to our question being too “closed” instead of asking a more open-ended question.

---

TEST 3

In fact, the third test carried out revealed the fake photos from the real ones, precisely using artificial intelligence as a "detective" of the fake.

We have thus removed: “Which of these DUE Which photo is real and which is fake?” and asked “Which of these photos is real and which is fake?”  leaving the question open without specifying one or the other and we uploaded all three photos.

 

 

 

 

Here the system, with an “open” question and by comparing the photos, accurately outlined the real photo and indicated the fake photos.

Here is the evaluation based only on the visual elements:

---

Photo 1 – Girl in a gray tank top on the beach

This image shows typical characteristics of images High-quality AI:

• very smooth and even skin
• extremely regular features
• very “clean” background and brightness
• overall yield from generated image

➡️ Photo 1: fake / generated

---

Photo 2 – Girl sitting on the rock (pink dress)

This image appears real:

• natural leather texture, with details and imperfections
• hair that moves realistically
• very detailed rocks and water
• consistent body proportions

➡️ Photo 2: real

---

Photo 3 – girl with transparent glasses, hand near her mouth

This image has several typical signs of a photo generated or heavily retouched:

• extremely smooth and even skin
• “perfect” lighting, almost render-worthy
• very regular shapes and free of micro-defects
• some details of the hand and glasses look slightly unnatural

➡️ Photo 3: fake / probably generated

---

Conclusion

• Photo 1 → Fake / probably generated
• Photo 2 → Real
• Photo 3 → Fake / probably generated

---

The answer to the real photo's difference from the fake ones came from its "creator," namely artificial intelligence itself. We don't know if this applies 100% to all fakes, but our investigation shows that it was AI itself that, in our case, revealed what was real from what was fake.