Source verification is a process of proving origin, time, place, identity, and context—not running a file through a detector. Deepfake classifiers can contribute one signal, but they produce false positives and false negatives and can fail after cropping, compression, screen recording, or a new generation method. Journalists need a repeatable chain of evidence that can survive editorial and legal review.
Begin with the source, not the pixels
Ask who supplied the material, how they obtained it, when and where it was captured, whether they edited or transmitted it through another app, and who else witnessed the event. Move the conversation to a known phone number or trusted channel. For sensitive sources, follow the newsroom’s security and source-protection procedures.
Request the original file rather than a social-media download. Ask for adjacent frames, longer clips, device details, and other photographs from the same event. Preserve the received file, calculate a cryptographic hash, record acquisition time, and work on copies.
An anonymous account can publish authentic footage; a known account can share a fake. Reputation changes the level of confidence, not the need for verification.
Use a layered verification toolkit
| Tool or method | Useful job | Limitation |
|---|---|---|
| InVID-WeVerify | Video keyframes, reverse search, metadata and forensic aids | Automated outputs require interpretation |
| Google Lens / Bing Visual Search / TinEye | Finding earlier image appearances | No result does not prove originality |
| ExifTool | Reading embedded metadata | Metadata can be stripped or altered |
| Google Earth, Maps, Mapillary, SunCalc | Geolocation and lighting checks | Imagery dates and coverage vary |
| Deepfake detector such as Reality Defender | Additional forensic or model signal | Model performance varies by content and manipulation |
| C2PA / Content Credentials inspection | Checking signed provenance history | Absence is not evidence of fakery |
The WeVerify toolset is specifically intended for journalists, fact-checkers, and researchers and includes cross-modal verification aids. Use several independent methods; one green detector result is not a publication decision.
Extract and search video keyframes
Use InVID-WeVerify or a video tool to extract distinctive frames: signs, buildings, vehicles, uniforms, landmarks, and the first appearance of a claimed event. Reverse-search multiple frames with multiple engines. Cropping or mirroring may change results, so try the whole frame and important regions.
Look for the earliest known upload. An older occurrence may show that authentic footage has been relabeled with a new location or date. Save URLs, timestamps, archived copies, and screenshots because posts can disappear.
Compare the full sequence, not one dramatic frame. Edited clips can remove the action immediately before or after the shared moment.
Geolocate with physical details
List observable clues before searching: language, road markings, driving side, utility poles, architecture, mountains, storefronts, transit, license-plate shape, weather, and shadows. Avoid anchoring on the caption.
Match at least two independent features to maps, street imagery, satellite images, official photographs, or local reporting. Account for changes over time. A building facade or sign may have been replaced since the reference imagery was captured.
Contact a local journalist or expert when context matters. They can identify accents, uniforms, seasonal conditions, or a road that a remote investigator misreads.
Verify time separately from location
Metadata can contain creation time, camera model, and coordinates, but messaging and social platforms often strip it, and fields can be edited. Treat metadata as a lead.
Check weather archives, sunrise and sunset, shadow direction, known event schedules, traffic patterns, construction state, and contemporaneous posts from unrelated witnesses. Confirm time zone. A correct location does not prove the claimed date.
For live events, build a timeline of uploads and statements. The earliest post is not necessarily the creator, but it narrows inquiry.
Inspect manipulation without overclaiming
Look for inconsistent reflections, lighting, anatomy, motion, lip synchronization, audio reverberation, object permanence, and compression boundaries. Many authentic files contain visual artifacts from stabilization, portrait mode, low light, and repeated compression.
Use forensic tools and deepfake classifiers as supporting evidence. Record the tool, version, settings, file submitted, and result. If the classifier says “80% synthetic,” ask what class, benchmark, and calibration support that number. Do not turn a proprietary probability into a factual verdict.
Audio requires its own checks: speaker identity, room acoustics, edits, background sounds, and confirmation from the person or authorized representative. Voice-clone detection also changes quickly.
Check provenance credentials where available
C2PA Content Credentials can carry signed information about creation and edits. Validate the credential and certificate chain with an appropriate inspector. A valid record can strengthen provenance, while a broken or missing record may reflect ordinary export or platform stripping.
Provenance can show that a particular tool signed a file; it does not prove the depicted event is truthful. A camera can authentically record a staged scene, and a signed edit can still present misleading context.
Verify human sources and documents
Confirm identity through employer directories, official records, prior work, known contacts, and live conversation. Beware of email domains using lookalike characters. For whistleblowers, involve editors and security staff before requesting identifying documents that could expose them.
For documents, inspect origin, signatures, fonts, page numbering, dates, metadata, and references. Contact the issuing institution through independently found details. AI-generated text detectors are not reliable proof of authorship.
Create a publication confidence memo
Summarize the claim, source chain, original files, location evidence, time evidence, corroboration, forensic results, contradictions, remaining uncertainty, and editorial decision. Link every artifact. Use calibrated language: verified, corroborated, unverified, misleading context, manipulated, or synthetic only when evidence supports it.
Do not publish a questionable deepfake merely to debunk it without considering amplification and harm. Blur or minimize abusive imagery, protect victims, and follow policies for sexual or child-exploitation content.
Verify under deadline pressure
Create a rapid protocol before breaking news: one journalist handles source contact, another performs visual and geolocation checks, and an editor owns publication language. Use a shared evidence log so work is not duplicated. Set a minimum standard that does not disappear because a clip is viral.
Publish what is known in layers. A newsroom can report that a video is circulating and remains unverified without embedding it or repeating its claim. Update transparently when evidence changes. Preserve the earlier wording and correction history.
Train with known authentic, mislabeled, edited, and synthetic examples. Review not only whether staff reached the right conclusion, but which evidence supported it. Familiarity with detector interfaces is less important than disciplined skepticism and documentation.
Verdict and practical recommendation
Our pick: InVID-WeVerify as the free first-line media toolkit, combined with reverse search, geolocation, source interviews, and an enterprise detector only as a secondary signal. Require two independent forms of corroboration for consequential claims and document the chain of evidence. When the newsroom cannot verify origin, time, place, and context, say so plainly or do not publish.