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, and from the websites of India's fact-checkers, such as BOOM, Alt News, Factly, Vishvas News and Newschecker.
• For a photo or video, if you agreed to it, up to 4 still frames go to Google's image search, to find where the same picture was online before. Frames that show private people, documents, or sexual or graphic content, and frames that are mostly text, are never sent. The report shows an earlier copy only if our research opened that page and used it as evidence, with the date the page gives. Finding none is a normal answer, never a verdict.
• 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 strong the evidence is

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.
• “Strong evidence” needs an official record or an independent fact-checker, or two separate news organisations, and at least one of them read in full.
• A claim about what a politician or party said or did needs an official record or a fact-check for “Strong evidence”; news reports alone give at most “Some evidence, not complete”. A quote we can't trace to an official record, the full original or a fact-check is NOT PROVEN.
• If a claim is about a government, party or company and the only strong sources are its own, the evidence is at most “Some evidence, not complete”, 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.
• When a forward misleads, the report says how it misleads in one line from a fixed list (for example “Old video, new caption”), and only when the evidence shows it. “A cut or edited clip”, “Words or a screenshot not from the real source” and “A picture made or changed by computer” need an official record or a fact-checker.
• A health claim can't be TRUE without a health authority (WHO, ICMR, the health ministry) or a published review, and health reports say: “This is not medical advice. Ask a doctor before you change any medicine or treatment.”

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. How strong the evidence is

Reports say how strong the evidence is, in words. There is no number: nobody has measured our accuracy yet, so we don't show one.

• Strong evidence
• Some evidence, not complete
• Weak evidence

NOT PROVEN, CAN'T CHECK, OPINION and SATIRE have no evidence line, because there is nothing to be more or less sure of. NOT PROVEN says “We couldn't settle this either way.”

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 claim 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 judge matters of faith. ForwardCheck checks facts about events, not beliefs.
• We don't rule on questions a court is still deciding.
• We don't publish lists of verdicts.
• We don't show ads, and we don't sell your data.

9. Safety first

• A scam message: “Don't click the link, don't pay anything, and never share an OTP. Report the message on Chakshu at sancharsaathi.gov.in. If you lost money, call 1930 or go to cybercrime.gov.in now.”
• Someone unwell or in danger: “Call 112 (ambulance 108).”
• Election claims: “Check with the Election Commission: call 1950 or visit voters.eci.gov.in.”
• Links: every link in a forward, and every page a report cites, is looked up in Google Web Risk, Google's list of dangerous websites. A page on the list is never used as a source, and a link on it gets a red warning. A link that isn't on the list may still be a scam, and we say so. Only the link is sent to Google.

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, and the report suggests independent fact-checkers you can ask yourself on WhatsApp. They aren't connected to ForwardCheck, and they may not reply.
• 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

Who reviews a report: no person reads a report before it is published. Vidit reads every report people flag, and he decides every correction.

Anyone reading a report can tell us it's wrong, harmful or shows private information, without an account: tap “Something wrong with this check?”. Vidit reads every problem report; harm and private-information reports are looked at within 24 hours, and private information shown in a report is removed within 24 hours.

Right of reply: if a check is about you or your organisation, choose “I'm the person or organisation this check is about”. A person reviews it within 72 hours. If your account checks out, the correction quotes your position.

When we find a mistake, we check the forward again. Before a correction is published, Vidit reads the sources of both the old and the new check; a correction is published only when the new answer rests on a source the old one missed or misread. The old report then leads with the new answer, and everyone who looked at it is told once. We never remove a report because someone disagrees with it.

12. Changes to the method

• 24 Oct 2026: No number on how sure we are; reports say how strong the evidence is, in words. Matters of faith and questions before a court are not judged. A higher bar for claims about what politicians said. Pictures from a link are not shown when they show private people. We no longer publish a list of corrections.
• 23 Oct 2026: Links are looked up in Google Web Risk; a page Google lists as dangerous is never a source. Anyone can report a problem with a check.
• 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.