How to Tell If an Image Is AI Generated

Check three things, in this order: hands and any text at 100%, whether the light falls the same way on every object, and the file — its metadata and whether it matches the pixels. Then count. As a working rule, two independent anomalies in different categories is strong evidence of generation; one soft anomaly is inconclusive; none means unverified, not human. Everything here you can see yourself, with no detector in the loop.

Why the "count the fingers" checklist stopped working

The 2022 tells were anatomy failures — six fingers, three arms, an ear melting into a jaw. That era's models were bad at hands because hands are hard: many parts, self-occlusion, odd poses. The advice spread from it was a list of anatomy errors, and plenty of pages still run it.

Anatomy got fixed first, because it was the most visible failure and the easiest to train against. What survives is inconsistency — texture repeating at the wrong scale, light disagreeing with itself, edges resolving too cleanly. QuillBot's detector page names exactly that class of artifact: "texture repetition, lighting inconsistencies, and the structural slips generators leave behind" [2].

Provenance changed the arithmetic too. Google says SynthID has watermarked more than 180 billion images and videos and 240,000 years of audio [8], and OpenAI's tool reads SynthID watermarks and C2PA Content Credentials in images and audio from ChatGPT, its API and Codex [6]. The file can now be more decisive than the picture.

One flag: "anatomy is mostly fixed" is inference, not a published finding — guides still describe distortions generally [1] while vendors describe consistency artifacts specifically [2].

The seven things you can actually check

Each item answers three questions: where to look, what counts as a hit, and when the hit is a lie.

1. Hands and limbs

Look at: every visible hand at 100%, counting digits and knuckles; then the wrist angle and which side each thumb sits on.

Counts as: an impossible digit count, digits fused with no separation line, an extra joint, a thumb on the wrong side, a wrist bent past what a forearm allows.

Lies when: the hand is cropped, gripping something, moving or small in frame. Real people also have four fingers and fused joints.

2. Text and symbols

Look at: signs, labels, clothing prints, book spines, screens, plates, logos.

Counts as: letters touching into one shape, strokes starting or stopping mid-letter, characters outside any alphabet, a long string that reads as nothing in any language.

Lies when: text is small, angled, out of focus or printed on fabric. Stylised fonts break strokes on purpose, and you cannot call a script gibberish if you cannot read it.

3. Material and texture

Look at: skin at cheek and forehead, fabric weave, knitted or tiled patterns, hair at strand level.

Counts as: a pattern repeating at a scale the material cannot support, skin smooth in one patch and detailed beside it, hair merging into a solid mass.

Lies when: noise reduction erased the pores, the lens was wide open, or the subject really is plastic or airbrushed.

4. Light and shadow

Look at: the brightest highlight on the face, the shadow under the chin, the shadow on the ground, then reflections in eyes, glasses, metal, wet floor.

Counts as: highlights on both sides of a face with no light that could make them, a shadow pointing away from every visible source, softness contradicting the light's apparent size.

Lies when: two real sources, bounce off a bright wall, or on-camera flash. Relighting moves highlights legitimately.

5. Depth and perspective

Look at: the transition into background blur, relative size of objects at different distances, long straight lines — kerbs, ceilings, railings.

Counts as: blur starting and stopping unevenly around a subject, background objects scaled as if at the wrong distance, two vanishing points for lines that should be parallel.

Lies when: portrait-mode bokeh, telephoto compression, tilt-shift, panoramas.

6. Noise and compression

Look at: a flat area — sky, wall, shadow — at 100%, then a busy area. Compare their noise, then find patches suspiciously clean inside a noisy frame.

Counts as: noise stopping at a boundary no lens made, one region cleaner than the rest, blockiness changing alignment across the frame, repeated block-scale patterning over a repainted area.

Lies when: the file was re-saved, resized, denoised or upscaled — each leaves its own texture.

7. The file

Look at: the metadata block, the pixel dimensions, and whether the two agree.

Counts as: EXIF naming a camera that cannot produce those dimensions, no camera data from a source that should keep it, a Software field naming a generator.

Lies when: most platforms strip metadata on upload, which makes missing EXIF the weakest signal here.

When the file is the evidence

Open properties, not the preview: right-click, Properties, or drop the file into an EXIF viewer. You want camera make and model, lens, exposure, ISO, focal length, date. Be suspicious of a camera model beside dimensions it cannot shoot, a modified timestamp older than the created one, a Software field naming a generator, or no camera data on a file said to come from a camera.

Then check pixels against the claim. A file claiming a phone camera but measuring exactly 1024 by 1024 is not a phone photo: phone sensors produce their own aspect ratios, and perfect squares are generation defaults. That is a pattern inference, not a documented rule — some apps do export squares.

Compression is the third layer. Re-saving leaves a block pattern; a screenshot adds a second grid on top, usually misaligned. If block edges do not line up across the frame, the file has been through two encoders.

Three false positives you will hit

Heavy re-compression. Platforms re-encode everything. Blockiness, softened edges and crushed detail are normal after one upload, and they look exactly like generation artifacts.

Screenshots. A screenshot re-encodes pixels and discards metadata: you lose section 7 entirely and gain section 6 artifacts that were never in the original. What is left is your eye, and the eye is a weaker instrument than it feels: across roughly 287,000 image evaluations by more than 12,500 participants, people correctly identified AI-generated images 63% of the time, against 62% overall [11].

Restored or retouched old photos. Scanner noise, dust removal, AI upscaling and face restoration leave one signature: smooth patches, invented pore texture, edges sharper than the surrounding grain. An AI-restored 1960s print fails several checks here while being a photograph.

Images that are only partly AI

The hardest case: a real photo with a regenerated face, a swapped background, a removed object.

The seam is where you look. QuillBot describes inpainting as leaving "seams the AI photo detector can read: texture that repeats where it shouldn't, lighting that doesn't agree with the rest of the frame, and edges that resolve too cleanly" [2]. That is also the manual checklist — trace the boundary and compare texture scale, light direction and edge sharpness on each side.

Second signal: noise discontinuity. The replaced area was synthesised, so its grain is generated rather than captured and will not match the grain around it.

Detectors are weakest here and admit it. QuillBot notes that manipulated images "score lower than fully generated images" [2]; DeepAI warns that "heavily rendered digital art, airbrush work, and AI-based editing tools can influence the result" [3].

AI-generated vs Photoshop vs a real photo

Photoshop edits what is there. Removed objects leave real pixels from elsewhere, so grain continues across the seam, perspective is inherited, and the light is the original light. Cloning repeats a photographed pattern at the photographed scale.

AI generation invents the frame. Nothing is inherited, so the failure mode is disagreement between parts: texture at the wrong scale, shadows pointing two ways, detail dissolving on inspection.

A real photo is internally consistent by construction: grain matches optics, blur matches aperture, shadows match one light, and the file carries a camera fingerprint — if it survived upload.

The rule: Photoshop errors are local and physical; AI errors are global and statistical.

When these checks don't work

They fail on small crops, screenshots, stylised art where realism was never the goal, and anything through three rounds of compression. They fail against laundering — a generated image printed, photographed and re-uploaded loses most of what this page looks for.

Then switch to provenance. SynthID is Google's watermarking system, embedded at generation time and built to survive cropping, filters, colour changes and lossy compression [9]. Google's portal checks images, video and audio for it and states outright that it "is not a general AI detector" — it only speaks for content from companies that adopted SynthID [10]. A positive is proof of AI involvement; a negative proves nothing. How to use it, what it accepts and where it stops is covered in SynthID Detector: how to use Google's AI watermark check.

FAQ

How do I identify an AI-generated image? Three passes: anatomy and text at 100%, then light and depth consistency, then metadata. Count independent anomalies rather than weighing one — a single odd detail is usually compression or a real-world exception.

What does an AI photo look like? Less "fake" than "too resolved": skin without pores, hair without strands, text that reads as shapes rather than letters. The anomaly is often an absence, not an error.

Can ChatGPT identify an image? It can describe one; describing is not verifying. OpenAI's tool checks SynthID watermarks and C2PA Content Credentials in images and audio from ChatGPT, its API and Codex [6], and does not currently support text verification [6].

Can Google detect AI images? Through provenance, not pattern-matching: SynthID checks run in Search, Gemini and Chrome, plus a portal at synthid.com [9][10]. It covers Google and partner models including OpenAI, NVIDIA and Kakao [7], and will not flag generators that do not participate.

Can you check if an image is AI-generated online for free? Yes, as a second opinion. Detection is probabilistic: DeepAI notes that digital art and AI editing can sway results [3], Hive's demo reports a confidence figure plus a breakdown by likely generator [4], and vendor pages compete on claims like "super-human accuracy" [5]. A score is evidence, not a verdict.

Sources

1. Microsoft — Is it real or AI? How to spot an AI-generated image: microsoft.com/en-us/microsoft-365/microsoft-designer/how-to-tell-if-images-are-ai-generated
2. QuillBot — AI Image Detector: quillbot.com/ai-image-detector
3. DeepAI — AI Image Detector: deepai.org/ai-image-detector
4. Hive — AI-Generated Content Detection demo: thehive.ai/demos/ai-generated-content-detection
5. Sightengine — Detect AI-generated images: sightengine.com/detect-ai-generated-images
6. OpenAI — Verify OpenAI-generated content: openai.com/research/verify/
7. Google blog, October 7, 2026 — Google expands SynthID Detector for AI content: blog.google/innovation-and-ai/models-and-research/google-deepmind/synth-id-ai-content/
8. Google blog, October 15, 2026 — TIME has named Google SynthID one of the Best Inventions of 2026: blog.google/innovation-and-ai/models-and-research/google-deepmind/synthid-time-best-inventions/
9. Google DeepMind — Google SynthID overview: deepmind.google/models/synthid/
10. SynthID Detector portal, supported formats and additional terms: synthid.com
11. Roca et al., "How good are humans at detecting AI-generated images? Learnings from an experiment", arXiv:2507.18640: arxiv.org/abs/2507.18640