is-this-photo-real

Is this photo real

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.

Triage: what question are you actually answering

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.

Method

1. Get the best copy. Every re-encode, resize and screenshot destroys forensic signal. Chase the original upload or the agency version, not the platform rendition. Hash it, work on copies. If all you have is a screenshot, say so and lower every downstream conclusion.
2. Provenance. Run find-the-original-image; for video, extract and search keyframes. You want an earlier appearance, a different caption, a photographer credit, and an on-page date corroborated through read-deleted-pages. An earlier copy with a different caption ends the case.
3. Metadata. Run secrets-in-file-metadata: editing chain, thumbnail-versus-image comparison, timestamp inconsistencies, whether MakerNotes fit the claimed device.
4. Provenance signing. Check for C2PA Content Credentials.
5. Internal consistency. Signage language, plates, currency, uniforms, vehicle models, season, weather and shadows against the claimed date and place, via geolocate-from-pixels. Ordinary detective work, more productive than forensics.
6. Physical consistency. Lighting, shadows, reflections, perspective.
7. Signal-level forensics. Noise residuals, JPEG quantisation and double compression, clone detection, colour-filter-array traces. Easy to over-read, worthless on a platform-processed file.
8. Write what you verified, not a verdict.

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

Lighting, shadow and geometry — the checks that hold up

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.

Error level analysis, and why it is mostly used wrongly

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 and Content Credentials

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.

AI generation — durable tells and tells that rot

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.

Video

• Container and encoder. ffprobe -show_format -show_streams and MediaInfo give the encoder string, frame rate, rotation matrix and track structure; values typical of a platform re-encode mean you do not have an original. ffprobe -show_frames exposes frame types — duplicated frames mean frame-rate conversion or inserted slow motion, an unexplained keyframe mid-way through a static shot can mark a splice, and interlacing or telecine artifacts reveal a pipeline nobody mentioned.
• Generation loss. Blockiness, banding and mosquito noise stack with each re-encode. Heavily degraded footage is old, widely copied, or both — and every signal-level test on it is void.
• Audio. Lip-sync drift, room acoustics that do not match the visible space, ambience that does not change when the camera goes indoors, noise-floor jumps at edit points. The weakest link in most fabricated video and the least examined.
• Face-swap behaviour. Flicker or blur at hairline and jaw, face lighting not tracking head movement, teeth and tongue degrading during speech, the face at a different resolution from the frame, breakdown on profile turns and hand occlusion.
• Keyframes to reverse search. The most common outcome of a video verification is finding the video, older, elsewhere. Extraction commands live in find-the-original-image.

Where this goes wrong

• Every filter has a base-rate problem. Run six forensic tools on an authentic photograph and something will look anomalous. Anomaly is the normal condition of real images.
• The platform did it. Resizing, re-encoding, chroma subsampling and metadata stripping produce artifacts people attribute to manipulation. Establish processing history before interpreting any artifact.
• You will be handed the worst copy — a screenshot of a repost of a crop. Most signal-level analysis is invalid on it, and the honest report says so.
• "Not manipulated" is not a finding. Absence of detected manipulation is a statement about your tests, not about the image.
• A real photo can be entirely misleading. Selective framing, staged scenes and a true image with a false caption all pass every forensic test.
• Debunking amplifies. A detailed refutation spreads the original claim — an editorial judgement worth making deliberately. Material is also sometimes seeded to be discovered and debunked, or to see who investigates; see investigate-without-getting-made.
• Identification from resemblance is the highest-consequence error here — "this is person Z because they look alike" is not a finding.

Confidence grading

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

• Confirmed original and correctly described — earliest copy located from a plausible originator, date independently corroborated, location visually verified, internal details consistent, nothing contradicted in metadata or geometry.
• Confirmed recontextualised — the same image demonstrably published earlier with a different, better-sourced caption. The most common positive finding here.
• Probably manipulated — a specific, describable physical or geometric inconsistency you can point at, ideally with a source for the inserted element. Not a heatmap.
• Probably synthetic — multiple durable generative tells, no provenance history, no camera-consistent compression or metadata, no earlier copies. Name the tells.
• Unconfirmed — no earlier copy found, nothing wrong found. Where most cases end.
• Cannot be assessed — the available copy is too processed for the tests the question needs. Say which tests were invalid and why.

Worked example

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

yt-dlp pulls the best rendition rather than a screen recording. ffprobe 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 read-deleted-pages 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 write-the-intel-brief.

Pivots

Legal and ethical notes

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.