How to Verify Images and Video in the AI Era
Save the original context
Record the post URL, account, caption, upload time and the exact claim attached to the visual. An authentic image can still be used with a false date or location.
Search for earlier appearances
Use reverse image search or visual search to find older copies, alternative crops and pages that may identify the original event.
Compare landmarks and environment
For day-to-day verification, the practical consequence is straightforward: Road signs, weather, shadows, terrain, storefronts, transit maps and architecture can test a location claim more reliably than a dramatic caption.
Treat metadata as supporting evidence
For day-to-day verification, the practical consequence is straightforward: EXIF and provenance metadata can help when present, but metadata may be stripped or altered. Do not let missing metadata become proof of manipulation.
Check motion and audio separately
A real video can be paired with unrelated audio, and synthetic audio can be placed over authentic footage. Verify the visual sequence and the voice claim as separate evidence streams.
Use AI artifacts cautiously
Distorted hands, lip-sync errors and odd reflections can be clues, but generative media is improving and compression can create similar defects. Visual intuition should trigger further checking, not a final verdict.
Find authoritative copies
For public events, search official livestreams, full speeches, pool video, local broadcast archives or original photographer uploads to restore missing context.
Check edits and sequence
Short clips can reverse meaning by removing what happened immediately before or after. Find a longer version and note cuts, speed changes, overlays and subtitles.
Use provenance features where available
Tools such as Google’s “About this image” can surface earlier uses and contextual information. Treat provenance indicators as one input among several.
Escalate high-stakes visuals
When a visual could move markets, trigger violence, damage a person’s reputation or cause a financial transfer, require independent confirmation through known channels before acting.
Keep a short evidence log with the frame searched, earliest matching URL, location clues and unresolved contradictions. That record is more useful than a screenshot of a search result because another person can repeat the same steps.
Compression artifacts, unusual edges and metadata changes can suggest further investigation, but social platforms routinely recompress files and strip metadata. A detector score should not outweigh a traceable original upload, corroborating footage from another angle or a verified location. For synthetic-media concerns, prioritise provenance and independent observation before visual intuition.
Inspect edits without overclaiming
Reverse image search is strongest for finding earlier uses, not for declaring an image true or false. Extract several video frames because different moments may produce different matches. Search distinctive background objects as well as the main subject; reposts frequently crop or overlay the most recognisable part of an image.
Visual verification is easier when you treat three questions independently. Provenance asks where the file came from and whether an earlier version exists. Geolocation asks whether visible landmarks, road layouts, signs or terrain match the claimed place. Chronolocation asks whether weather, shadows, construction, vegetation or known events fit the claimed date.