How to Tell if an Image Is AI-Generated: A Practical Checklist

Why Bother Checking Manually

AI detectors are one click away, so why learn to inspect images yourself? Because detectors are probabilistic and context-blind — they will confidently flag a studio product photo and confidently pass a well-made deepfake. Human inspection reads meaning: whether the physics, the context, and the file itself tell a consistent story. The strongest verdicts combine both.

This guide is the checklist we use ourselves, ordered from fastest to most involved. Each step takes seconds; together they catch the vast majority of synthetic images in the wild today.

Step 1 — Read the File, Not the Picture

Before looking at pixels, look at the container. Download the image (don't judge from a screenshot — platforms strip metadata on upload, so judge the original file where possible) and inspect it:

• EXIF block: absent entirely, or Software naming a generator (Midjourney, Stable Diffusion, DALL·E) — near-conclusive. A real camera profile (make, model, lens, plausible exposure) — supportive but forgeable, as we explain in the EXIF guide.
• C2PA manifest: visible to tools like ExifTool as JUMBF/C2PA structures. If present and valid, it states the file's true history cryptographically.
• Dimensions and footprint: perfectly square 1024×1024 or 2048×2048 outputs from a "camera," or PNG files named like diffusion outputs, raise the first eyebrow.

On social platforms you often can't get the original file — that's expected. Move to step 2 with the pixels you have.

Step 2 — Hands, Text, Small Objects

Diffusion models render plausible rather than correct, and they still fail hardest on things with strict structures that humans parse unconsciously:

• Hands and fingers: count them. Extra, missing, or fused fingers; impossible joint directions; fingernails that melt into skin. Modern models are much better — so also check the anatomy logic: does the grip actually hold the object?
• Text: signs, menus, price tags, book spines. Letterforms that morph mid-word, inconsistent baselines, or characters that are decorative gibberish. Text remains the single most reliable pixel-level tell.
• Small mechanical objects: eyeglass arms, earrings, chair legs, instrument keys, bicycle spokes — things with precise topology. Look for parts that merge, pass through each other, or attach to nothing.
• Teeth and ears: classic fusion zones — blended molars, ears that flow into hair.

Step 3 — Physics: Reflections, Shadows, Light

Generators model appearance, not optics. Interrogate the scene's physical consistency:

• Reflections: in mirrors, windows, water, eyes. A reflection must show the scene from the reflected surface's point of view. Synthetic images frequently reflect a different scene or blur what should be legible. Eye catchlights that disagree with the room's light sources are a strong tell in portraits.
• Shadows: each object needs a shadow consistent with the light source in direction, softness, and color. Look for people whose shadows are missing, doubled, or point a different way than everyone else's.
• Specular highlights: shiny surfaces — car paint, glass, wet streets — should share highlight direction across the whole frame.
• Perspective: parallel lines (building edges, tiled floors, table sides) must converge consistently. Watch for backgrounds that "fit" without obeying the foreground's vanishing points.

Step 4 — Textures at 100% Zoom

Open the image at full resolution and study the micro-textures — this is where the statistical signature of generation lives, and it's exactly what detectors measure (see how detection works for the theory):

• Skin: real skin has pores, fine hair, uneven texture across zones. Synthetic skin trends toward porcelain smoothness or painterly uniformity.
• Hair: individual strands should separate and tangle. Watch for hair that becomes painted mass, or strands that merge with clothing at the boundary.
• Fabric: weave direction, thread irregularity, seam behavior. AI fabric often has directionless texture that drapes wrong.
• Foliage and crowds: distant leaves or background faces that resolve into fractal mush — individually plausible, collectively patterned.
• Edges: object boundaries that glow, halo, or smear — a compression-like artifact that doesn't match the file's actual compression level.

Step 5 — Context and Source

The strongest evidence is often not in the image at all:

• Reverse image search: does an earlier, different, or clearly labeled version exist? An "authentic" photo that traces back to a prompt gallery is settled.
• Account history: a "photojournalist" account three weeks old posting exclusively viral content is its own signal.
• Narrative physics: claims that are oddly perfect — impossibly convenient framing of a news event, weather that contradicts the stated date, famous people doing unlikely things. Staged perfection is a synthetic signature at the story level.
• Cross-platform triangulation: legitimate events generate many independent photographs. A world-event with exactly one circulating image deserves suspicion on arithmetic alone.

Putting It Together: A Scoring Approach

In practice, weigh the signals:

• Conclusive (any one): valid C2PA manifest declaring generation; generator string in original-file EXIF; reverse-search origin in a prompt gallery.
• Strong (two or more): fused anatomy, unreadable text in-situ, reflections showing the wrong scene, and account-context anomalies together.
• Weak (never alone): smooth skin, square aspect ratio, missing EXIF, "a feeling." Each has innocent explanations — heavy retouching, cropping, platform stripping, good lighting.

Knowing the Limits

Honesty about the ceiling: well-executed generation plus deliberate humanization — of exactly the kind this site documents — defeats every step above some of the time. Textures can be re-noised, EXIF rewritten, physics inpainted by hand. Detection and evasion are an arms race with no permanent winner; the practical defense of truth is provenance at the source and multiple independent witnesses, not any single test.

For the machine side of this coin, read our comparison of AI image detectors — where they succeed, and exactly where they disagree with each other.