How to spot AI

Shadows fall in directions that disagree with each other, or a reflection is missing the thing that should be in it. Diffusion models compose plausible regions, not a single physical light setup.

The subject is flawless and the background quietly falls apart: a doorframe that changes width, a crowd member with a fused arm, a railing that stops existing behind a shoulder.

Skin pores, fabric weave, gravel or foliage tile or smear at 100% zoom. Detail is generated, not captured, so it lacks the irregularity of a sensor.

Perfect symmetry, centred subject, no lens dirt, no motion blur, no awkward crop. Real photographs carry evidence of a person holding a camera.

Not hands any more — teeth counts, earring pairs, watch faces, glasses arms, jewellery that merges into skin.

Generators leave statistical traces in the frequency spectrum that classifiers can learn. This is real and strong, but it needs tooling — it is invisible to the naked eye.

Frequency-based detection of generated images (arXiv)

Signage, labels and book spines used to dissolve into pseudo-letters. Current models advertise accurate text rendering, so this now only catches older or open-weight checkpoints.

OpenAI on 4o image generation and text renderingBlack Forest Labs on FLUX.2

Textures, freckles, logos and hairlines shimmer or re-draw between frames. Whole detection families are built on frame-to-frame inconsistency.

Temporal artifacts in generated video detection (arXiv)

A cup, a hand, a passer-by leaves frame and comes back changed — or never comes back. Objects occluded mid-shot are the most reliable place to look.

Cloth, hair, liquid and crowds move with the right vibe and the wrong mass. Collisions resolve too softly; things pass through each other slightly.

Mouth shapes lag or over-articulate, and consonants that need lip closure (b, p, m) are the usual giveaway. Commonly cited, less formally quantified.

Perfectly smooth motion with no operator error, or a pan whose parallax does not match the depth of the scene.

Measured overuse of words like “delve”, “intricate”, “underscore” and “boasts” in text written since 2023. A real, quantified shift in published English.

Science Advances: measurable LLM word-frequency shiftCOLING 2025 study

Tidy tricolons, “it's not just X, it's Y”, a summary paragraph nobody asked for, and balanced hedging that refuses to land on a position.

Sentence lengths cluster; paragraphs are the same size. This is what perplexity-based detectors measure, and light paraphrasing defeats it.

Citations, quotes, statistics and URLs that look right and do not resolve. A strong signal of an unchecked model — but humans invent sources too.

The internet's favourite tell and one of its weakest: the human base rate was always high, and vendors have tuned the behaviour down.

Illustrative, high-saturation, slightly airbrushed. Prompt-faithful but literal: everything asked for is present, centred and lit like a stock photo. Embedded text collapses.

Caveat — Superseded by native GPT-image generation in 2025, which fixed most of the text and anatomy artifacts.

Photographically convincing, legible in-image text, coherent multi-object scenes. Residual habit: a faintly warm, evenly exposed studio quality and unnaturally tidy scene logic.

Caveat — Vendor-claimed improvements; the remaining 'tidy' impression is a heuristic, not evidence.

Sora 1 clips drift: morphing limbs, background continuity loss after a few seconds. Sora 2 adds audio and much better physical plausibility, pushing failures into long-shot continuity and crowd behaviour.

Caveat — Clip length and prompt complexity change the failure rate more than the model version does.

Neutral, documentary-leaning colour and very clean edges. Fewer stylistic fingerprints than Midjourney, which makes it harder, not easier.

Caveat — Outputs may carry SynthID watermarking, which tooling can detect even when the eye cannot.

Built for editing, so the tell is usually the edit: a subject preserved perfectly while the surrounding lighting, grain or perspective does not quite match the plate.

Caveat — Edited real photos are a genuinely mixed case — part real capture, part generation.

Strong camera language and native audio. Watch for physics that resolves too gently and for ambience that is generically 'correct' rather than specific to the place.

Caveat — Independent frame-level comparisons across video models are still thin.

Text only. Careful, structured, heavily hedged prose; explicit caveats and balanced 'on the one hand' framing. Rarely commits to a bold claim without qualification.

Caveat — Style is prompt-steerable, so this describes defaults only. Anthropic ships no public image or video generator.

Long, exhaustively structured answers with heavy enumeration. Early reasoning models sometimes leaked deliberation phrasing (“wait, let me reconsider”) into final output.

Caveat — Open weights mean output style depends heavily on who deployed it and how.

Fluent, long-context, list-friendly English. Occasional register mismatches and calques from Chinese-language training data in longer pieces.

Caveat — The clearest folklore entry here — widely observed by practitioners, not formally studied.

Excellent prompt adherence and skin rendering. Practitioners report a recognisable skin and micro-contrast signature and slightly plasticky highlights, especially in the fast [schnell] variant.

Caveat — The 'FLUX look' is community consensus rather than published research.

The strongest house style of any generator: cinematic rim light, shallow depth of field, dramatic haze, and a narrow beauty standard for faces. Aesthetic homogeneity is a documented property.

Caveat — v7 explicitly improved hand and body coherence, so anatomy-first checking is no longer effective on it.

The classic artifacts survive here: fused fingers, warped ears and teeth, duplicated background limbs, and mushy detail away from the subject. Old checkpoints are still in daily use.

Caveat — Specific anatomy quirks per checkpoint are practitioner lore, not measured findings.

C2PA Content Credentials

An open, cryptographically signed manifest travelling with a file, recording which tool made or edited it. Adopted by OpenAI, Google, Adobe and camera makers.

Limit — Metadata is stripped by screenshots, re-encoding and most social platforms, and absence of credentials proves nothing.

C2PA specification

SynthID

Google's imperceptible watermark applied across Gemini, Imagen and Veo output in text, image, audio and video, detectable by its own verification tooling.

Limit — Only covers Google-generated media, and research has demonstrated removal and evasion attacks.

Google DeepMind: SynthID

OpenAI provenance metadata

OpenAI attaches C2PA Content Credentials to images produced in ChatGPT and the API.

Limit — OpenAI itself describes these as helpful indicators rather than a guarantee, and they survive only while the metadata is intact.

OpenAI: C2PA in images

Reverse image and context search

Still the highest-yield check available to a person with no tooling: find the earliest appearance, the original crop, and whether the event happened at all.

Limit — Fails on genuinely novel generations that were never published anywhere else.

Google Images