What are the main categories of AI image tells?
The guide How to Distinguish AI-Generated Images from Authentic Photographs by Negar Kamali and three co-authors (Northwestern University, June 2024) groups AI image tells into five categories: anatomical implausibilities (hands, eyes, teeth, merged bodies), stylistic artifacts, functional implausibilities (objects that could not work, garbled text and logos), violations of physics (shadows, reflections, depth and perspective) and sociocultural implausibilities (unlikely scenarios). The authors built the guide from 138 images they generated with diffusion models, 9 AI images from social media and 42 real photographs, and they stress that human-visible artifacts are not always present.
| Category | What to look for, with examples from the guide |
|---|---|
| Anatomical implausibilities | Hands, eyes, teeth, merged bodies |
| Stylistic artifacts | Plastic texture, cinematization, hyper-real detail, uneven resolution and colour |
| Functional implausibilities | Objects that could not work, garbled text and logos |
| Violations of physics | Shadows, reflections, depth and perspective |
| Sociocultural implausibilities | Unlikely scenarios |
What is the "AI look"?
The AI look is what Kamali and colleagues call stylistic artifacts: AI images "often look a bit too perfect". The guide lists four:
- Plastic texture. The skin of AI-generated people is often waxy, shiny, cartoonish or glossy.
- Cinematization. AI images often have a cinematic or picturesque style that dramatizes the subject, like a still from a film.
- Hyper-real detail. AI images can render an unnatural level of detail in specific places, most often hair that looks excessively soft, fine-grained and windswept while the rest of the scene is still.
- Inconsistencies in resolution and colour. Parts of the image carry a different level of detail or vibrancy from the rest, like a subject pasted onto a background.
Can one tell prove an image is AI?
No single tell proves anything. The Kamali guide shows real photographs with each stylistic artifact: fashion and cosmetics photography often shows people with perfect skin, and many photographers deliberately shoot in a cinematic style through colour grading, film stock or camera choice. The signature is several tells appearing together, read alongside the image's context and source.
How good are people at spotting AI faces?
People are worse than chance at spotting realistic AI faces. In Nightingale and Farid's study (PNAS 119(8), 2022), 315 participants each classified 128 faces as real or synthesized, from a set of 400 faces made with StyleGAN2 and 400 matched real faces. Average accuracy was 48.2%, against a chance level of 50%. A second group of 219 participants, given training and feedback on every trial, reached 59.0%. A third group of 223 participants rated the faces for trustworthiness on a 1 to 7 scale: synthetic faces averaged 4.82 and real faces 4.48, so the AI faces were rated 7.7% more trustworthy. The authors conclude that synthesis engines have passed through the uncanny valley.
| Experiment | Participants | Task | Result |
|---|---|---|---|
| 1. Baseline | 315 | Classify 128 faces as real or synthesized | 48.2% accuracy (chance is 50%) |
| 2. Training and feedback | 219 | Same task, with training and feedback on every trial | 59.0% accuracy |
| 3. Trustworthiness | 223 | Rate 128 faces from 1 (very untrustworthy) to 7 (very trustworthy) | Synthetic faces 4.82, real faces 4.48: synthetic rated 7.7% more trustworthy |
All three experiments drew faces from the same set of 800: 400 synthesized with StyleGAN2 and 400 real faces matched to them.
Is there a better way to check than looking?
Provenance checks are more reliable than visual inspection. Google's SynthID embeds an invisible watermark in images, video, audio and text made with Google's generative AI products at the moment they are created, and Google says it is designed to survive cropping, filters and lossy compression. The Coalition for Content Provenance and Authenticity (C2PA) publishes an open standard called Content Credentials, which records the origin and edits of a piece of content like a nutrition label. Both only help when the tool that made the image supports them, so a missing watermark is not proof that an image is real.
Sources
Checked 13 November 2026:
- Kamali, N. and three co-authors (2024), How to Distinguish AI-Generated Images from Authentic Photographs, arXiv:2406.08651
- Nightingale, S. J. and Farid, H. (2022), AI-synthesized faces are indistinguishable from real faces and more trustworthy, PNAS 119(8)
- Google DeepMind, SynthID
- C2PA, Content Credentials
Next step
Take one of your own generated images and check it against the four stylistic tells above, then read working with the AI look for how to design with them on purpose. To check where an image came from rather than how it looks, see which platforms label AI images.