Verification order is the whole skill: provenance first, pixels last. Finding the earliest copy and reading its caption settles more cases than every forensic filter combined, and it produces evidence you can show someone. Pixel forensics produces a colourful heatmap and an argument.

The beginner mistake is running error level analysis on a downloaded JPEG and announcing the image is fake. The second is looking for deepfakes: the overwhelming majority of deceptive media is real footage with a false caption — right pixels, wrong war, wrong year, wrong country.

find-the-original-image

geolocate-from-pixels

Then always ask what the image would look like if the claim were true, write it down, and check for those things specifically. Verification tests a hypothesis; hunting for anomalies fails, because anomalies are everywhere.

Tools and their failure modes: reference/tool-catalogue.md. Ordered by cost: reference/verification-checklist.md.

No tooling, reasoning you can explain to an editor or a court, immune to recompression. This is where to spend your time. Shadow convergence is the strongest. Sunlight is parallel, so in a perspective image, lines drawn from each shadow's tip through the base of the object that cast it must all meet at one point — the projection of the light source. Draw three or four. An object whose line refuses to meet the others was probably not in the original scene; sloped ground is the confound, so use one plane only. Then: shadow direction and penumbra hardness should be consistent across an outdoor scene; specular highlights in eyes, glass and polished metal should agree on where the lights are; a reflection must show what is in front of it, correctly placed and reversed; parallel lines should converge on a common vanishing point with eye level consistent for people on one ground plane; and real lenses leave an optical signature — consistent depth of field, chromatic aberration at high-contrast edges, vignetting, a noise floor that varies with brightness. An element carrying none of that, in an image that has it elsewhere, was added.

ELA re-saves the image at a known JPEG quality and displays the difference, on the theory that a region compressed a different number of times responds differently. Four reasons it produces confident nonsense: it responds to content, so edges and texture light up while flat sky and skin go dark, meaning every image has "suspicious bright regions"; one re-save destroys it, so ELA on a social-media download describes the platform's encoder and nothing earlier; it cannot localise a modern edit, because content-aware fill, generative editing and a full re-save leave no differential history to find; and it fails in both directions, with bright regions on untouched images and clean output on manipulated ones both routine.

Where it earns its place: on a single-generation JPEG straight from a camera, a pasted region from a differently-compressed source can genuinely show up. Narrow case. Use it as one weak input, only on least-processed files, never as the basis of a published claim. More defensible relatives — quantisation-table comparison against camera signatures, and double-compression detection — are also defeated by platform processing.

C2PA binds a cryptographically signed manifest to a file recording capture and edit history. Where it exists it is the strongest provenance evidence available, because it is verifiable rather than inferential. A valid manifest means the signer asserts this history, the file is unchanged since signing, and you know who to hold responsible — not that the content is true. A signed photograph of a staged scene is a signed photograph.

Absence means almost nothing: most cameras do not sign, most editing pipelines do not preserve manifests, and platforms strip them during re-encoding. Missing credentials are the default state, not a red flag. Same for the IPTC digital-source-type field used to label synthetic media, and for model-specific invisible watermarks — a positive is strong where you can check it, a negative only says one vendor's mark was not found.

Anything resting on a model's current weaknesses will be fixed. Prefer tells grounded in physics and structure.

Durable, because they need a world model the generator does not have: impossible lighting (inconsistent shadow directions, missing shadows under objects, a subject lit from a direction with no source); structural incoherence in background objects (a bicycle frame that does not connect, a railing whose baluster spacing changes, stair treads that do not line up, a strap that vanishes and resumes, patterned fabric whose pattern ignores the folds); text degradation, especially small, repeated or peripheral text; contact and occlusion errors, such as a hand around a cup that does not enclose it, or feet not meeting the ground; absent camera physics — no sensor noise, no chromatic aberration, uniform focus, and too little high-frequency detail, which is what "over-smooth skin" actually is; and no plausible provenance at all.

Ages badly — check, but do not rest on: finger and tooth counts, ear asymmetry, garbled foreground text, mangled jewellery, suspiciously symmetrical faces.

Detector tools. A confident score with no auditable reasoning. They false-positive on compressed, resized, upscaled, heavily edited and low-light real photographs, false-negative against generators newer than their training data, and are adversarially fragile — mild recompression moves scores. Run more than one, treat them as a weak signal, never publish a conclusion resting on one. If your evidence is a percentage from a website, you have no evidence.

Grade each claim separately — one image can have a confirmed origin, a contradicted caption and unconfirmed authenticity at once.

A video circulates showing an explosion, captioned as a strike on a named city that morning.

pulls the best rendition rather than a screen recording. shows a frame rate and encoder string typical of a platform re-encode, so signal-level tests are off the table — noted, not lamented.

Keyframes extracted at scene changes. Frame four hits nothing. Frame nine, the only one with a skyline, hits a news site from eighteen months earlier in a different country. Provisional answer in ten minutes. But the earlier page's video is shorter, so which is the parent? The suspect version has a hard cut and a noise-floor jump in the audio four seconds in, and the ambience after the cut does not match the visible space. Added audio, not added video.

Dead end worth recording: an AI detector scored the frames as likely synthetic. They are a re-encoded crop of real broadcast footage — exactly the input that fools these tools. Discarded.

Findings: confirmed recontextualised, with corroborating the earlier page's date; audio probably manipulated; authenticity of the underlying video cannot be assessed at signal level given the copy available.

Reporting standard. State what you verified, what you could not, and what each conclusion rests on — never a bare "fake" or "real". A defensible line reads: the image was published at least three years before the claimed event, on a news site, credited to a named photographer; the location matches that credit and not the claim; no manipulation was detected, but the only available copy was a platform re-encode, so signal-level tests were not meaningful. "Our analysis shows this is fake" is not. Include the tests you ran and their negative results, the file's processing history, and your assumptions, then hand the package to .

Publishing an accusation of fabrication against a named person carries defamation risk in most jurisdictions, and "our tool said so" is not a defence. Uploading material to online forensic services discloses it to those services and, for some, to public galleries — check whether a tool publishes submissions before submitting anything sensitive. If the media depicts a crime, a victim, or intimate content: minimise copies, do not redistribute, and in the case of child sexual abuse material stop immediately and report to the appropriate authority rather than analysing it. See ../../ETHICS.md.