How to verify images and videos for journalists and fact-checkers

When do newsrooms need deepfake detection?

Newsrooms need deepfake detection when content arrives faster than it can be traced, which is most of breaking news. Verification desks, fact-checkers, social media editors and investigative reporters all face the same situations.

• A dramatic photo from a disaster, protest or conflict spreads before any reporter is on the ground.
• A video of a public figure saying something explosive appears on a little-known account.
• An audio “leak” of a politician or executive circulates in messaging apps.
• A news-looking website publishes stories with images and quotes nobody can source.
• Recycled media: real footage from an old event presented as new.

The last case is a reminder that many misleading visuals are not AI-generated at all. A real photo with a false caption needs source tracing, not an AI image detector.

What can go wrong when you fact check AI images?

You can publish a fake, or you can dismiss something real. Publishing a synthetic image as authentic damages credibility fast. Calling an authentic image “AI” can be just as harmful: it discredits witnesses and feeds what legal scholars Bobby Chesney and Danielle Citron called the “liar’s dividend”, where the existence of deepfakes lets people deny real evidence.

Detectors struggle most with exactly what newsrooms receive: screenshots, re-uploads, cropped and compressed files. Every pass through a social platform removes metadata and degrades the pixels detectors rely on.

How to verify images and videos for journalists: a step-by-step workflow

To verify a viral video or image, work from the source outward and keep notes as you go. This sequence follows standard open-source verification practice.

1. 01Find the earliest version. Run a reverse image search (Google Lens, TinEye, Bing) on the image or on key frames of the video, and sort results by date.
2. 02Contact the uploader. Ask how, when and where they captured it, and request the original file, which keeps metadata and full quality.
3. 03Check provenance. A content credentials checker reads C2PA and EXIF/XMP data that may name the camera or the AI tool used.
4. 04Geolocate and chronolocate. Match landmarks, signs, terrain and weather; check shadows against the claimed time of day.
5. 05Inspect visually. The guides on how to tell if an image is AI-generated and how to spot a deepfake video list what to look for.
6. 06Run AI detection on the best copy you have, and read the listed signals, not just the verdict.
7. 07Weigh everything, decide what you can say with confidence, and tell readers what you could and could not confirm.

How to report on a suspected AI image or deepfake

Report on synthetic media in a way that informs readers without spreading the fake further. A debunk that reposts the original at full size can do more harm than good.

• Show the image or video only when needed, smaller, clearly labeled as false or unverified on the image itself.
• Explain how you checked it: the earliest source, the provenance data, the visual clues and the detection result.
• Describe confidence honestly. “We could not confirm where or when this was taken” is a legitimate finding.
• Separate the questions: whether the content is generated, and whether the claim attached to it is true.
• Update the story if new evidence appears, and say what changed.

Which AnyFact detectors fit journalists, and what do they return?

Four detectors cover newsroom work. Each starts with provenance, then an open-source detection model, with Claude as a second opinion only in the grey zone. You get a verdict (probably human, uncertain or probably AI), the signals, an explanation in English, French or Arabic, the method’s limits and a PDF report for your files.

The AI website checker is useful for suspected content farms: a domain registered recently, with pages of generated text and images, is a strong lead for a story.

What AI detection cannot settle for a newsroom

AI detection cannot tell you whether an image shows what its caption claims. It cannot date footage, identify a location or confirm who is speaking. It answers one narrow question, whether content shows signs of generation, and it can be wrong in both directions.

Use it the way you use any single source: as a lead that needs corroboration. When verification stays inconclusive, saying so plainly is good journalism.

Frequently asked questions