How ForwardCheck checks a forward

1. We read the forward itself

• Videos up to 15 minutes and files up to 200 MB.
• We go through the video scene by scene, picking the moments where the picture changes.
• We listen to what is said, in Hindi or English. In long videos and voice notes we listen to the first 10 minutes of speech.
• We read the text on screen, and any message that came with the forward.
• Every report has a “What we checked” card: the frames, the words we heard and the text we read.
• Some apps don't let us open their videos. Then we read the post's title, description and cover picture, say so at the top of the report, and show you how to share the video file for a full check.

2. We list the claims

We write down each claim that can be checked: a fact, a number, a quote, a date or a place. Opinions (“this leader is bad”) are labelled as opinions, not checked.

3. We look for evidence

• First, earlier fact-checks of the same claim, from the public database that fact-checkers publish to.
• Then our research step searches the web and reads the pages it finds.
• It may cite only pages it actually retrieved. A quote counts only if it is in the text we read.

4. Rules in our code decide how sure we can be

These rules run after the AI writes its draft. The AI can't override them.

• No source, no verdict. TRUE needs a strong source that supports the claim; FALSE needs one that refutes it. Otherwise the answer is NOT PROVEN.
• Sources are ranked by the kind of organisation they are (official records, independent fact-checkers, news organisations, everything else), never by their politics.
• “Very sure” needs an official record or an independent fact-checker, or two separate news organisations, and at least one of them read in full.
• If a claim is about a government, party or company and the only strong sources are its own, we are at most “fairly sure”, and the report says why.
• Our own look at a video counts only for what the video shows. It never proves that a real person said or did something, because voices and faces can be faked.

Source standards: how we rank sources

5. A second review, for fairness

A separate review reads the draft as a sceptical outsider. It replaces loaded words, checks that the sources aren't one-sided, and checks that we described the forward in its own words. It can make an answer more careful, never more certain. The one exception: if we misquoted the forward, it restates the claim faithfully, and the evidence rules run again.

6. The dial: how sure we are

Every report has a 0–100 dial.

• 80–100: Very sure
• 55–79: Fairly sure
• Below 55: Not very sure

Strong, independent sources raise it. Weak or missing sources, or a verdict we had to correct, lower it.

7. What the verdicts mean

TRUE

This is true.

MOSTLY TRUE

This is mostly correct.

MISSING CONTEXT

It is real, but the full story is different.

MISLEADING

Some of it is true, but the way it's told gives the wrong idea.

MIXED

Some parts are true and some are not.

FALSE

This is not true.

NOT PROVEN

We couldn't find a source that confirms or rules this out.

SATIRE

Made as a joke, by a known satire publisher, or confirmed by a fact-checker.

OPINION

Someone's view, not a fact.

CAN'T CHECK

There is nothing here we can check.

8. What we won't do

• We don't check accusations about private people. Their names are left out of reports and never searched for.
• We hide phone numbers, email addresses, UPI IDs and account numbers in the message you send.
• If a video or photo may show the sexual abuse of a child, we don't check it. We delete it.
• We don't rate people, parties or religions. We check claims.
• We don't show ads, and we don't sell your data.

9. Safety first

• A scam message: “Don't click links, don't pay any fee, and never share an OTP. Report it at 1930 or cybercrime.gov.in.”
• Someone unwell or in danger: “Call 112 (ambulance 108).”
• Election claims: “Check with the Election Commission: call 1950 or visit voters.eci.gov.in.”

10. Our limits

• AI writes the reports, and AI can make mistakes. Open the sources and judge for yourself.
• Brand-new news may not be checkable yet. Then we say NOT PROVEN.
• Hindi and English only, for now.
• A check is true as of its date. After 7 days, members can ask for a fresh check.

11. When we get it wrong

Members can tell us from any report: under “Was this check right?”, tap “No”. Anyone can write to us from the Support page. When we find a mistake, we check the forward again, and the old report points to the new one.

12. Changes to the method

• 17 Oct 2026: A verdict now needs a source that points its way. “Very sure” needs a source we read in full. A government vouching for itself gives at most “fairly sure”. Safety advice for scams, emergencies and elections.
• 13 Oct 2026: Every summary and reply is now checked against the forward's own words, after an early report described a claim the video never made.
• 13 Oct 2026: First version.

Who is behind ForwardCheck

ForwardCheck is built and run by Vidit Chopra (GlobalInno Services), an independent maker in Meerut, India. He decides what ForwardCheck checks and how, and sets the evidence rules. In each check, AI does the reading and research, inside rules enforced in code.

We don't show ads, and we don't sell your data. Write to us at contact.techfusionlab@gmail.com.