Can independent fact-checkers do this instead?

Not at scale. Meta ended third-party fact-checking on its platforms in January 2025, moving to Community Notes, which its own Oversight Board called an inadequate global substitute in March 2026. The IFCN's 2025 report found 45.3% of fact-checking organisations reporting revenue declines and only 22.6% calling themselves financially sustainable. One person with a Mac mini can check a reel in about two minutes; Instagram serves billions of reels.

103,478 likes on a quote that was never said

Q: Why should Instagram run this check itself?

Instagram already has the video, audio, caption, overlay, and audience graph in one place, before a clip goes viral, with the compute to check it in seconds. A ministry needed half a day to label one post; a Mac mini needed two minutes. A platform would need neither. What's missing isn't capability, it's policy, particularly after Meta ended third-party fact-checking in January 2025.

Q: Would this solve fake news?

No. The check only verifies whether a post's text matches its video content, not whether that content is true, and models still make mistakes, especially with satire. But a real clip wearing a false caption is one of the most shared shapes of misinformation, and flagging that mismatch before the share button would limit its spread. Limited is not solved, but it is still fewer people misled.

103,478 likes. A quote that was never said.A viral reel put words in a government spokesperson's mouth. A browser extension backed by a pipeline of local AI models on one Mac mini caught it in about two minutes: it listened to the video, read the caption and the on-screen text, and checked the claim against the web. FactLens is a working system built on local AI models: an open-source downloader, a local speech model for transcription and translation, a local vision model for on-screen and image text, and a local language model for the judgment, with free web search for evidence. Its sibling tool, Nudger, judges the diet instead of the dish: it classifies every post that crosses the screen into a private on-disk ledger and quietly flags when one category dominates. Both run entirely on one Mac mini.Vindicated by the ministry itselfAt 12:10 am IST on Sunday 26 July 2026, FactLens flagged a viral reel that put a fabricated quote in the mouth of the Ministry of External Affairs spokesperson: verdict UNRELATED, the claim appears nowhere in the speech. Around midday the same Sunday, MEA FactCheck, the ministry's official fact-checking handle, labelled the same claim misleading and a misrepresentation of the Spokesperson. The Agent was roughly twelve and a half hours ahead of the official denial, with no newsroom and no privileged access: a Mac mini, reading the post itself.The evidence

Screenshot: FactLens's amber verdict on the reel, before any official word

Screenshot: MEA FactCheck's official post the same day, calling the claim misleading

Frame from the reel: the overlaid claim the speaker never actually says

Screenshot: FactLens running inside Instagram, the verdict panel rendered on the post itself

How it worksFetch: an open-source downloader pulls the public post. Listen: a local speech model transcribes and a local language model translates the audio. Read: video frames are sampled and OCR reads the overlaid text; standalone images and photo posts also get a local vision model for stylised type. Compare: a local model judges caption and overlay against the speech, SUPPORTED, CONTRADICTED, or UNRELATED, no internet needed. Check: extracted claims are searched on the open web free of charge and judged strictly against the evidence with a 0 to 100 confidence score; weak evidence is flagged, never dressed up. Show: a verdict banner states the outcome in one sentence, green, red, or amber, with the evidence one click away. Everything runs on one Mac mini. Honest limits: models mislabel edges, the compare step judges match not truth, thin search evidence returns INSUFFICIENT rather than a guess, and every verdict ships with its evidence.And clearing an honest oneThe contrast case: a public Instagram reel by Akash Rajawat (5,735 likes at the time of analysis) fact-checking two viral Sonam Wangchuk clips. The system transcribed the Hindi speech, translated it, read the overlay text, and judged both the caption and the overlay SUPPORTED: that reel is honest. The check works in both directions.Why should Instagram run this check itself?Only the platform has the video, the caption, the overlay and the audience graph in one place, at marginal cost near zero, before a clip goes viral, with no scraping walls. Independent fact-checkers cannot do this at scale: Meta ended its third-party fact-checking program in the US on 7 January 2025 and moved to Community Notes; Meta's own Oversight Board warned in March 2026 that Community Notes are not a proper global substitute; and the IFCN's State of the Fact-Checkers 2025 report found 45.3% of fact-checking organisations reporting revenue declines, with only 22.6% considering themselves financially sustainable.Would this solve fake news? No. It checks whether the text matches the content; it does not establish ground truth. But a real clip wearing a false caption is one of the most shared shapes of misinformation, and flagging it before the share button would limit it. Limited is not solved. Limited is still fewer people misled.