Common visual verification problems and how to fix them

What to take away

A genuine photograph can carry a false claim. Move its date, its location or its crop and the viewer receives misinformation with no fabricated pixels. Start with provenance and context. Manipulation detection comes later, and often not at all.

Problem 1: asking only whether the file is fake

"Real or fake" collapses six separate questions. A camera-recorded scene can be miscaptioned. A composite can be honest illustration.

Fix: test source, creation method, place, time, depicted event and edits as separate findings. Publish the parts you confirmed instead of one vague label.

Problem 2: analyzing a screenshot as the original

A screenshot drops the URL, upload date, account, caption and resolution. It can also place a real post inside a fabricated interface.

Fix: search the text, username and visible identifiers for the live or archived page. Ask the sender for the page URL and the original file. Keep the screenshot as evidence of what circulated, not of what existed.

Problem 3: trusting metadata without a chain

Device clocks drift. Software rewrites files. Platforms strip fields. A GPS value may record where a file was processed rather than what it shows.

Fix: compare metadata against source history, visual features, daylight and weather. State which fields were present in which file version.

Problem 4: accepting one landmark match

Buildings, ridgelines and road furniture repeat. Search tools push the reviewer toward a famous place, and then every clue seems to fit.

Fix: match a bundle of fixed features and the spatial relationships between them. Compare road curvature, building order, slope, poles and view direction. Then hunt for geometry that contradicts your candidate, and write down the alternatives you rejected.

Problem 5: treating reverse-search absence as proof of novelty

Indexes miss private posts, closed platforms, deleted pages, crops and low-resolution files. No result does not mean the image was made this week.

Fix: run several engines on full frames, crops, keyframes and mirrored versions. Search exact captions and visible text. Check the uploader's history and the event terms.

Problem 6: turning a detector score into a verdict

The NIST Open Media Forensics Challenge briefing describes an evaluation platform for research on detecting inauthentic images and tracing digital origins across generative, deepfake, CGI and anti-forensic techniques. Continuing evaluations reflect a moving technical problem.

They do not hand newsrooms a detector that settles a file. For the full workflow behind these judgments, see the image and video verification guide.

Fix: preserve the input, record the model and version, and confirm the file fits the tool's intended use. Test innocent explanations such as compression or resizing. For a consequential finding, bring in a qualified analyst and weigh the score against source and context.

Problem 7: ignoring sequence and audio

A short clip can omit the act that prompted a response. A loop makes one event look repeated. Swapped audio changes the perceived place or motive.

Fix: find the longest version and step frame by frame around every cut. Transcribe the audio and check it against visible events. Seek other angles. Describe what the recording shows, not what happened outside it.

Problem 8: verifying publicly before protecting people

Posting a face, landmark or coordinate to crowdsource help can expose a witness, a survivor, a home, a shelter or a medical site. Uploading sensitive media to third-party tools creates new copies you no longer control.

The Red Cross handbook on data protection in humanitarian action sets the principle a newsroom borrows: protecting personal data is part of protecting a person's life, integrity and dignity. It treats social media as a key risk for groups in volatile settings.

Verification methods belong inside that safety-aware process, especially with footage of abuse or conflict.

Fix: assess harm before you assess the file. Restrict access, strip unnecessary identifiers from working copies, use secure channels, and ask for location help privately. Publish only the detail a reader needs.

Problem 9: hiding uncertainty in the caption

An article can explain that the date is unresolved while the caption reads "Video shows Tuesday's attack." Many readers meet the thumbnail or social card and nothing else.

Fix: carry one evidence status through every surface where the claim appears.

Common questions

Are visual artifacts proof of AI generation?

No. Compression, motion, stitching, sharpening, low light and ordinary editing all produce odd features. Artifacts tell you to look closer, not what you found.

Is a watermark proof of origin?

No. It may identify a republisher, be copied, or be added after the fact. Trace the file back and confirm the credited creator before you treat it as a source.

Can weather confirm an exact minute?

Usually not. Nearby observations can support or contradict a time range when conditions are distinctive. Treat it as one input among several, never as a timestamp.

When should publication wait?

Wait when the unresolved element is central to the story, when being wrong causes real harm, or when release could endanger a person or an operation. If the answer is unclear, ask an editor rather than the crowd.