Five Visual Tells That Give Away AI-Generated Images — and Why They're Disappearing

📚 This is Part 11 of the AIGC in Plain Language series Series contents: 01 Content Understanding | 02 RAG | 03 Three Tips for Choosing | 04 Tool List | 05 Document Q&A | 06 Pitfalls | 07 Industry Observations | 08 Running Models Locally | 09 Token Billing | 10 Inference Acceleration | 11 AI Image Generation (this article)

The first ten parts were all about text models. This one switches modalities: images.

You can spot an AI image at a glance, but you can't say why

Scroll past an image, and you'll likely judge "this was made by AI" in about three seconds.

But if someone asks, "What gives it away?", most people freeze and can only mutter, "It's just a feeling."

That "feeling" is actually very concrete. It's not mysticism; it's five categories of visual evidence you can break apart. This article takes them apart one by one, and also explains why AI images have become increasingly hard to recognize over the past two years.

First, the root cause: AI draws "averages"

What diffusion models do during training is essentially learn what a million images have in common, then sample a new image from that "common appearance."

Analogy: overlay ten thousand highly-rated photos and take the average.

What kind of face do you get? Skin without pores (pores are in different places, so they average away), no acne scars, no asymmetry, perfect even lighting, and perfectly standard facial proportions.

This average face is technically "most correct," but it has no human touch. It has sanded off everyone's unique features.

The most direct source of the "plastic look" is right here: it's not that the image is poorly drawn; it's that it's too average.

The "imperfections" on a real face—pores, fine lines, slight left-right asymmetry, oil sheen, small moles—are precisely the things that are hardest to average away, and they are the most authentic parts of real photos.

Five tells, broken down one by one

Tell 1: Skin looks airbrushed, and over-airbrushed.

Real skin has texture: pores, fuzz, blemishes, uneven oil. Under most default settings, AI skin is uniform, translucent, like plastic or wax.

The reason is above—texture is high-frequency detail; its position is random, so averaging destroys it. And the model's training objective is precisely "look like most samples."

Tell 2: Lighting doesn't obey physics.

This is the hardest piece of evidence. In the real world, the direction, intensity, and color temperature of light in a scene are usually consistent (even multi-light setups follow one set of logic): if the light source is above-left, all shadows fall down-right, and all highlights are on the left side.

Models don't understand physics; they only understand "does this resemble a photo?" So they draw:

• Window light on the face, but catchlights in the eyes point the wrong way
• Backlight on the hair, but no light source in the background
• Left side bright, right side dark, but the nose highlight is dead center
• Shadows from multiple objects that contradict each other

Checking shadow direction against the light source is the fastest way I identify AI images.

Tell 3: Local details collapse, and they collapse suddenly.

The whole image looks perfect, but zoom in and it falls apart. The areas most prone to collapse share a common trait: high structural complexity + few training samples + thin, crossing shapes.

High-risk area	Why
Fingers	Too many poses, complex occlusion
Text	Needs precise character structure; models only "draw something that looks like letters"
Glasses temples / earrings	Thin and long, must correctly intersect the face shape
Shoelaces / necklaces	Crossing structures, easy to fuse into one blob
Background crowd	Faces in the distance, too few pixels, smeared into mush

This kind of error has a characteristic: distant, small, secondary areas tend to break first. Because in latent space, small regions are compressed more aggressively, and they get fewer pixels and less information to begin with.

Tell 4: Composition is too tidy, like a textbook.

The training set was itself filtered to keep images with "clear subjects and standard composition"; later preference scoring reinforced that. What models learn is this "safe composition."

So AI images often have: subject dead center, horizon perfectly level, excessive symmetry, and backgrounds so clean they're unnatural.

Real photos tend to be imperfectly composed — subject slightly off-center, horizon slightly tilted, background a bit messy. This "imperfection" is what makes a photo feel human.

Tell 5: Colors are too bright, too saturated, too sharp.

This ties to "human scoring." Models are tuned with human preference data later on, and when people look at an image at first glance, they tend to give high scores to "bright, vivid, sharp, high-contrast" images.

So the model learns: lean toward vivid, and the score goes up.

The result is that familiar "AI look": over-saturated skies, glowing edges, HDR feel, and clarity like the lens was just wiped.

So why is it getting harder to tell now?

If you compare a 2023 model with a current one, you'll clearly see the improvement. Three reasons, none mystical:

First, the architecture changed. From U-Net to DiT (Diffusion Transformer): U-Net is convolution-heavy, and its receptive field grows only by stacking layers; DiT can see the whole image at every step, so global consistency is much steadier. This is an architectural change, not a tuning tweak.

Second, text understanding got stronger. With better text encoders, models now understand photographic language like "backlight," "candid," and "shot on phone" more accurately—ask for imperfection, and it will actually give you imperfection.

Third, and most critical: alignment with human preference. Massive amounts of "which one looks more real" human labeling directly suppress "AI-ness" as a negative sample.

So right now: on small images, social media sizes, and fast scrolling, it's already hard to tell. Only when you zoom in on fingers, text, or shadow logic do the flaws show.

A 10-second judging sequence: start where it's easiest to trip up

All that said, when you actually need to judge, you just need an order. My own habit is to go from the "hardest evidence" to the "softest impression":

1. First, check shadow direction. A scene has only one dominant light direction; all shadows must point the same opposite way, and highlights must line up with the light source. This is the fastest and least error-prone check.
2. Zoom in on hands and any text in the frame. Count fingers, check joint orientation; if text appears, zoom in to see whether strokes merge or are "letters that aren't letters."
3. Scan the background. Faces of people in the distance, earrings, glasses temples, shoelaces—places where the model invests the least often break the most visibly.
4. Only then look at skin. Whether there are pores, fine lines, uneven oil. This step is already close to "feeling," so don't use it first.
5. Composition and color last. Too centered, too symmetric, too vivid—those only add weight, they are not decisive evidence.

The wrong order leads to misjudgment: look at skin first, think "pretty real," then look at composition, think "a bit fake," and end up with a conclusion you're not sure about.

Remember: only the first three steps can find a "specific error" that counts as evidence; the last two only count as suspicion.

If you want your own images to "look less AI"

Four operations that actually work, sorted by effect:

Method	What to do	Effect
Write "imperfect" into the prompt	Add: candid shot, natural light, slight noise, film grain, skin texture, phone snapshot	Most effective
Lower guidance strength (CFG)	Drop from 7 to 3–5; the model is more "relaxed," less rigid	Obvious
Break symmetric composition	Explicitly write off-center, top-down, handheld angle, subject to the side	Medium
Add noise / recompress in post	Add 1%–2% noise, or recompress as JPEG, to break the "over-sharp" look	Medium

The numbers in the table are empirical ranges, not official recommendations—different models have different sweet spots, so tune to your own model.

A more common pitfall is in the prompt itself, and it pushes in both directions:

• Piling "smooth, perfect, ultra-HD, HDR, 8K" into the positive prompt—that's basically asking for plastic
• Putting "noise, grain, blur, low quality" in the negative prompt—that's suppressing real texture as a defect

Many people write prompts that way and then complain their images look fake.

A dose of cold water

First, this list of tells will expire. Every single item above is being specifically patched by the next model revision. It's fine as a popular science guide, but not as a forensics standard.

Second, "looks like AI" is not evidence. There have already been cases where real photos or hand-drawn illustrations were mistaken for AI. True verification looks at generation traces, metadata, and consistency across multiple images—not "feels like it."

Third, the thing to watch most isn't fake images; it's "average aesthetics." When more and more people generate images with the same model, what we see becomes more and more alike. Technical problems will be fixed; this one won't.

---

👋 This series is called AIGC in Plain Language, with the goal of explaining AIGC in plain language: no jargon memorization, no ranking lists, each piece explains one thing clearly, and includes a useful judgment method or template you can copy directly. If you find it useful, follow along.

When did you last recognize an AI image? Was it fingers, text, or shadow direction? Tell us in the comments, and I'll put together a "spot it in one glance" comparison table.