Five Visual Tells That Give Away AI-Generated Images — and Why They're Disappearing
As AI images get harder to spot, the practical skill shifts from casual spotting to careful scrutiny of shadows, micro-details, and metadata. For Western developers building AI image tools or content pipelines, this breakdown offers a concrete mental model for diagnosing output quality and for steering prompts away from the default "plastic" look.
AI image generators are easy to recognize in about three seconds, but most people can't explain why. The reasons break down into five concrete visual tells, all rooted in how diffusion models train: they learn the average appearance of a million photos, so they smooth away real texture, produce physically inconsistent lighting, collapse on small complex details like hands and text, default to textbook-safe composition, and push colors too bright and saturated because human preference scoring rewards that.
The article ranks the tells by reliability: checking shadow direction is the fastest hard evidence, followed by zooming in on hands and text, then scanning backgrounds for small objects. Skin texture and composition are softer signals and only count as suspicion, not proof. Newer models are fixing these flaws through architectural shifts (from U-Net to DiT), stronger text encoders that understand photographic language, and heavy human-preference alignment, so on small social-media images the tells are nearly invisible now.
For creators who want outputs that look less AI, the most effective moves are writing "imperfect" prompts (candid, natural light, grain, skin texture), lowering CFG guidance from 7 to 3–5, deliberately breaking symmetrical composition, and adding subtle noise or JPEG recompression. The article also warns that this tell list is temporary, that "looks like AI" is not reliable evidence, and that the bigger risk is aesthetics flattening as everyone uses the same models.
The tells are not bugs in drawing skill so much as artifacts of averaging many samples; higher-fidelity generation requires models that deliberately inject imperfection, not just more training data.
The shift from U-Net to DiT marks a real architectural turning point: global attention at every step enforces scene consistency in a way that layer-stacked convolutions cannot match.
Human preference alignment is a double-edged sword: it makes outputs more realistic but also reinforces a narrow, saturated aesthetic that risks flattening visual culture.
The article's own caveat undercuts its usefulness as a detection manual: every tell is being patched, so the list is more a snapshot of current model weaknesses than a permanent forensic tool.
Prompt engineering advice reveals a feedback loop: users routinely sabotage themselves by requesting "perfect" qualities that directly trigger the plastic look, showing that understanding the model's inductive biases matters more than adding negative prompts.
The tells get more granular in practice: headline text now renders correctly, but small in-scene text still falls apart, and mismatched or smeared paired accessories offer a quick check. A prioritized 10-second inspection order puts shadow direction first, then hands and text, then background faces and accessories, while skin texture and over-perfect composition are only secondary cues. Prompt guidance identifies the same contradiction the article describes: adding smoothness, HDR and 8K actively injects AI flavor, while suppressing noise and grain strips out the texture that makes images believable.
Text has fallen off the fastest in the last two years; large poster titles can basically be written correctly now. But small text in the background is still a disaster area: street signs, book spines, ingredient lists on the back of packaging — zoom in and it's all 'things that look like characters.' There's also a simple trick: check whether the left and right earrings, or the two glasses temples, are the same style; often one side is one way and the other side is another.
I've seen that left-right asymmetry too: with a pair of earrings or two glasses legs, the model often only draws one side carefully, and the other side gets smeared over.
A 10-second judging order, sorted from what gives it away first: 1. Look at shadow direction first - all shadows must point the same way and opposite the light; this is the fastest. 2. Zoom in on hands, and on any text in the image. 3. Check faces, earrings, and glasses temples in the distant background. 4. Only last, check whether the skin has pores and fine lines. 5. Composition that's too perfect or colors too saturated only counts as a bonus, not decisive evidence. Conversely, to make your own images less fake: Add 'candid, natural light, slight noise, film grain, skin texture' to the positive prompt; remove 'smooth, perfect, ultra-clear, HDR, 8K' from the positive prompt, and remove 'noise, grain, blur, low quality' from the negative prompt. 'Smooth, perfect, ultra-clear' is actively adding AI flavor; 'noise, grain' suppressed as flaws is exactly real texture. It's pushing in both directions; many people write it one way while complaining their images look fake.