AI Tells — What to Look For

When you're reviewing generated images and videos, you need to spot the things that scream "AI made this." Some are obvious; some are subtle but break the illusion. This page is the reference.

The Generation Runner already auto-checks for the worst AI tells and re-runs flagged outputs. But not every tell gets caught automatically, and you'll need to flag the rest manually.

The four categories

flowchart TB
    AT[AI Tells]
    AT --> A[Anatomy issues]
    AT --> B[Composition issues]
    AT --> C[Texture / surface issues]
    AT --> D[Logic / context issues]
    A --> A1[Extra fingers]
    A --> A2[Melted faces]
    A --> A3[Garbled limbs]
    A --> A4[Eye asymmetry]
    B --> B1[Stock-footage feel]
    B --> B2[Symmetrical composition]
    B --> B3[Floating subjects]
    C --> C1[Plastic skin]
    C --> C2[Painted look]
    C --> C3[Over-saturated colors]
    D --> D1[Wrong number of subjects]
    D --> D2[Impossible props]
    D --> D3[Floating objects]

Category 1: Anatomy issues

The biggest red flags. These get caught by the Generation Runner's auto-QA pass — but spot-check them anyway.

Extra fingers

The classic. A hand with 6 fingers, 7, sometimes 4 in obvious places. Most common when the prompt has multiple hands visible.

Two left hands (or two rights)

Both hands rendering with the same chirality. Often paired with the extra-finger problem.

Melted face

Features sliding off-center, asymmetric in a "not human" way, eyes at different heights, mouth and chin merging.

Garbled limbs

Arms with extra joints, legs bending the wrong way, hands attached at impossible angles.

Eye asymmetry

Slight differences are fine (real people are slightly asymmetric). Drastic differences — eyes at clearly different heights, one half-closed while the other is open in a frozen frame — are tells.

Teeth disasters

Too many teeth, teeth visible through closed lips, the upper and lower teeth not matching. Usually a problem in smiling scenes.

These are the ones the Generation Runner watches for. If a flagged image makes it past 3 rerun attempts, the runner gives up — you may need to manually rerun with a higher attempt cap or edit the prompt.

Category 2: Composition issues

These are subtler but break the "real candid phone footage" feel.

Stock-footage feel

Too clean. Too composed. The lighting is perfect, the subject is centered, the background is artfully blurred. Real phone videos have noise, asymmetry, slightly off framing.

Symmetrical composition where there shouldn't be

Both eyes equidistant from frame edges, the head perfectly centered. Looks like a passport photo, not a casual moment.

Floating subjects

The subject doesn't quite connect to the ground / surface. There's an invisible gap between feet and floor, or hands and a held object don't quite meet.

Backgrounds that pull focus

A background element that's sharper / more interesting than the foreground subject. Usually a stylized environment that the model decided to make beautiful.

Category 3: Texture / surface issues

These read as "polished but wrong."

Plastic skin

Too smooth. No pores, no fine texture, no hair. Looks like a 3D render rather than a photograph.

Painted look

The whole image has a slight brushstroke or oil-painting quality. Light hits surfaces in a too-uniform way.

Over-saturated colors

Reds that are TOO red. Skin tones that have unnatural warm or cool tints. Backgrounds that are too vivid.

Hair that's a single texture

Real hair has variation — different strands catch light differently, some flyaway strands, texture changing across the head. AI-generated hair often looks like one uniform helmet.

Category 4: Logic / context issues

These need a beat of looking before you spot them.

Wrong number of subjects

The prompt asked for one person; the output has two. Or there's a second person in the background that wasn't called for.

Impossible props

A coffee mug with two handles, a chair with three legs, a phone with no buttons in a context that needs them.

Floating objects

The product is suspended just off the surface, no visible contact. Hands holding things at awkward distances.

Hallucinated text

Background signs with garbled letters, price tags with nonsense numbers, product labels with random characters. This is low-priority — don't bother fixing unless the text is the focal point. Just accept it.

Wrong era / season / setting

Christmas decorations in a "summer kitchen" shot. Vintage props in a "modern bathroom." Usually the model got over-creative with the setting.

What you can ignore

These are usually fine even if they look slightly AI:

• Slight color shifts between candidates of the same scene (Veo's interpolation produces minor variation)
• Minor background imperfections that don't draw the eye
• Slight differences in skin tone between scenes (post-production color correction handles this)
• Hallucinated text in backgrounds (see above — low priority)
• Tiny background details that wouldn't matter at scroll speed

What's NOT an AI tell

Sometimes things look weird but they're actually fine:

• Realistic blemishes — moles, pimples, tiny scars. These read as "real," not "AI."
• Slight messiness — hair out of place, slightly wrinkled clothing, casual posture. Realer than perfection.
• Asymmetry — real faces are asymmetric. A slightly higher brow or off-center mouth is human.

Don't flag a candidate as "AI tell" just because something is imperfect. Imperfection is what makes phone videos look real.

Priority order when judging

When you're reviewing 4 candidates, ask in this order:

1. Are there anatomy issues (extra fingers, melted face)? → rerun if all 4 have them
2. Is the composition what the storyboard called for? → if no, rerun with prompt fix
3. Does the face match the reference? → if no, check ref image, rerun
4. Is the wardrobe correct? → if no, prompt fix needed
5. Does the lighting / environment match the workflow? → minor mismatches OK, major needs rerun
6. Skip everything else — texture / hallucinated text / minor color shifts → accept

Setting attempts higher when stubborn

The Generation Runner defaults to 3 regen attempts per node. If you have a scene with consistently bad hands across all 3 attempts, you can ask for more:

You: rerun Scene 05's image gen with up to 10 attempts — the hands keep
     coming out wrong and I want to give it more tries before pivoting
     to a prompt edit.

Claude: [reruns Scene 05 with attempts=10]

If after 10 it still fails: the prompt is wrong, not the model. Edit the prompt to reduce hand visibility (frame the scene with hands out of view) or specify hand position more explicitly.

When you're ready

→ Next: Image Prompt Rules — the rules every image prompt must follow to avoid AI tells in the first place.