Sourcy

Method and limitations

Last updated: September 2026

1. What the analysis does

• An artificial intelligence (Gemini, Google's model) receives the link to the video. Sourcy neither downloads nor stores the video.
• The AI listens to the entire soundtrack and looks at the picture at roughly one frame every 5 seconds.
• It splits the video into time-stamped passages, ideally one per verifiable claim. When the creator has declared chapters in the description, it uses them as landmarks.
• For each passage, it summarises what the creator claims ("According to the creator…"), without giving an opinion, then looks for sources that confirm or contradict that claim. The AI never settles the question itself: each note reports what the cited source establishes.
• For analyses run with Sourcy credits, a second pass checks the sources with Google Search, passage by passage: a proposed source is kept only if the search finds it and it does say what the note credits it with. Otherwise it is replaced by a source that was found, or the passage reverts to grey.

2. The three colours

Rebuttal videos. When a creator quotes someone in order to criticise them, it is the position the creator defends that is assessed, never the thesis they are taking apart. If the sources confirm the rebuttal, the passage is green.

3. Accepted sources

• A source must be an identifiable publication: a recognised publisher and a precise title (scientific journal, news outlet, institution, book).
• Rejected: the video itself, other videos (YouTube, including science or debunking channels) and podcasts, social-media posts, testimonies, and any description without a named publisher.
• A source that merely reports the statements of the person on screen (the record of their own hearing, an interview, their book) does not count as confirmation: it only proves those statements were made.
• A response to or "debunk" of the analysed video is not a source: it is an opinion about the video. The AI must cite the studies and publications it relies on directly.
• Quoting a document accurately is not enough: it is the conclusion the creator draws from it that is assessed. If the document itself qualifies it or warns against it, the passage is orange.
• When the creator cites a document themselves (report, study, official statistics), the AI must cite that document rather than an article about it.
• When the creator lists sources in the video description, the AI receives them and attaches them to the passages they concern. They are checked like any other source: a source cited by the creator is not automatic proof. Links to social networks, videos, donation pages or shops are discarded.
• The AI suggests at most two sources per passage.
• Any link must belong to a list of approved websites (news outlets, journals, institutions); other links are rejected automatically.
• A source cited by its title opens a search to find the publication: check it yourself.

Automatic safeguards. An analysis is rejected if it does not cover at least 75% of the video's length, or if it is too coarse (on average, at least one passage every 7 minutes). The AI gets two attempts, within the server's computing time limit; if they fail, the credits are refunded. A coloured passage with no accepted source reverts to grey.

4. The community's role

• Any signed-in adult user can suggest a source for a passage: a link to an approved website, and the same source only once per video.
• Before publication, an AI reads the source and checks that it actually covers the passage's topic and that it confirms (or contradicts) what the contributor states. Otherwise it is rejected, with the reason. This check can be wrong either way; it is limited to 20 suggestions per account every 24 hours.
• Anyone can vote for or against community sources: one vote per account, and every new source starts at zero votes.
• A community source counts towards the score as soon as it has at least one more "for" vote than "against" votes.
• With three more "against" votes than "for" votes, it becomes disputed: collapsed and excluded from the score, but still visible and open to votes, so it can be reinstated.
• When a video is analysed again (at most once every 24 hours), community sources and their votes are kept and attached to the new passages.

5. How the evidence score is calculated

For each green or orange passage, two sources are expected. We count the supporting sources that qualify (suggested by the AI, or from the community and validated by votes), at most two per passage:

Grey passages are not included. Contradicting sources are counted the same way, separately.

Example. A video has 7 green passages and 2 grey ones: 14 sources are expected. If 9 supporting sources are counted, the score is 9 ÷ 14, i.e. 64%.

6. Known limitations

• The AI can be wrong. It may misunderstand a passage, pick an imprecise or overly general source, or misclassify a relationship, especially in videos that rebut someone else. A dedicated rule reduces this risk without removing it.
• The AI's sources do not go through votes. They count as soon as the analysis is done and cannot, for now, be disputed by vote: only community sources can.
• Two analyses of the same video can differ. An AI is not deterministic: the split, the sources and sometimes the colours can change from one analysis to the next.
• Not every passage is checked. The second pass stops when the server's computing time runs out: the remaining passages keep the source proposed by the first analysis, which was not searched for.
• A timestamp can be off. The AI places each passage itself: on a long video it can put a topic at the wrong moment, especially when the creator outlines their plan at the start. If in doubt, check the video.
• The picture is sampled. At roughly one frame every 5 seconds, a very brief visual detail can be missed.
• Very recent news is tricky. The AI's knowledge has a cut-off date. When Google Search is not available, a claim that is too recent stays grey.
• Votes can be biased. One vote per account, starting at zero, the list of approved websites and the dispute threshold limit manipulation without making it impossible.
• A score is not a verdict. Read the notes and open the sources before drawing conclusions.

7. Independence and funding

• Viewing an analysis is free. Running a new analysis costs credits, which pay for the AI, or is free with your own Gemini key.
• Paying does not change the rules: the score is calculated the same way for every video, and nobody can buy a better score.
• Analyses run with a personal key go through the same server-side checks and are capped.
• Sourcy is developed in France by an independent developer. It is not affiliated with YouTube or Google.

8. Reporting an error

Vote against a community source, suggest a better one, use the report feature in the app, or write to contact@getsourcy.com. Criticism of the method is welcome: it helps the tool improve.