Method: how an analysis is made
Fact AI takes the claims in a social media video or image one by one, with AI assistance, and checks each of them against sources found by a live web search. This page explains the steps an analysis goes through and where the result can be wrong.
1. The content comes in
You send a post from Instagram, TikTok, YouTube, X or LinkedIn to Fact AI from the share menu. For a claim without a link you can send a screenshot; the text in the image is read too.
2. Claims are separated
The factual, checkable claims in the content are extracted one by one. Instead of one overall impression, each claim is assessed on its own, so you can see which sentence in a video holds up and which does not.
If the content contains no checkable factual claim, the analysis says so and shows no numeric score. "No verifiable claims" is not a verdict.
3. Sources come from a live search
We have no fixed pool of sources. For each claim a live web search runs during the analysis; whatever it returns is attached to that claim, and the links appear in the report. We don't guarantee any particular outlet. You can open the sources and read them yourself.
If no source turns up for a claim, we don't hide it: the claim's card says "No sources found".
4. Each claim gets a verdict
Each claim is rated on a scale from true to false based on the sources found (True, Mostly True, Mixed, Misleading, False). The sources behind the verdict appear in the report with their links.
Claims we cannot source are marked Unverified. That does not mean the claim is false, and we don't comment on it.
Each claim has its own confidence rating, and the analysis as a whole gets a confidence score from 0 to 100. The real information is not in the score but in the claim-by-claim verdicts and their sources.
Verdicts are about claims, not people. An analysis is not a judgement of the person who posted the video or appears in it.
5. AI-generation likelihood is read separately
Fact AI looks for signs of AI generation in the video or image and shows a likelihood percentage, with the signs that led to it. It does not give a definitive "AI or not" verdict. In compilation videos it separates which parts look synthetic and which look real. When it cannot assess the footage it says so plainly; that does not mean "real". This reading is a separate question from the accuracy score of the claims. Full guide: How can you tell if a video was made with AI?
Limits
Fact AI is an AI-assisted tool and can make mistakes. A search can return wrong or incomplete sources, a claim can be misread, and for a very recent event there may be no sources yet. For important decisions, open the sources yourself.
Reporting errors and disputes
- In the app: every analysis has a flag button. Pick a reason (wrong score, wrong source, missing information) and the report is saved. That record is what we look at when we fix the system. We don't promise automatic corrections.
- For an urgent error: write to dev@fact-ai.app. With Pro you can re-run the analysis.
- If you own the analysed content: use the dispute and takedown request form; no account needed. Requests are reviewed by hand; the analysis may be corrected, taken down, or left as is with a stated reason.
Public analyses and search engines
Analysis results enter a public pool without any identifier of the requester. Analyses that contain no sensitive content may be listed by search engines; analyses involving private individuals, children, victims, or crime, suicide or health contexts, and analyses of uploaded screenshots, are kept out of search engines. Details: Privacy Policy.