Fact-checking and verification

In short

Fact-checking is the discipline of testing a claim against primary evidence before (pre-publication, or "editorial" fact-checking) or after it is published (post-publication, or "ex-post" fact-checking of public claims and viral content). Verification is the narrower craft of establishing whether a piece of content — a photo, video, document, quote, dataset or forward — is authentic, original, correctly dated and correctly located, and who made it and why. In India it matters because WhatsApp forwards, AI-generated audio and recycled visuals have triggered lynchings, communal riots, market rumours and election falsehoods, and because the IT Rules, 2021 make due diligence a legal obligation for digital publishers and intermediaries. It is examined in BJMC papers on reporting, digital journalism and media law, and tested in placement rounds at India Today (AFWA), BOOM, Alt News, Factly, Vishvas News (Jagran), NDTV, The Hindu, Newschecker, Fact Crescendo, Logically Facts and in the trust-and-safety teams of Meta, Google and ShareChat.

Explained

Definitions and the information-disorder framework

• Misinformation — false information shared without intent to harm (an honest forward). Disinformation — false information created and spread deliberately to deceive. Malinformation — genuine information weaponised out of context (a real old video of a riot reposted as today's).
• Fake news is a loose, politically abused term; professional practice prefers the three categories above plus information disorder (Wardle and Derakhshan, UNESCO handbook *Journalism, "Fake News" and Disinformation*, used in most Indian syllabi).
• First Draft's seven types: satire or parody; false connection (headline not supported by body); misleading content; false context (real content, wrong frame — the commonest type in India); imposter content (fake logos of PIB, ANI, BBC); manipulated content (edited image or clipped audio); fabricated content (wholly invented, including deepfakes).
• Checkable versus non-checkable claims: facts, numbers, dates, attributions and causal claims are checkable; opinions, predictions, value judgements and faith claims are not. An exam answer must make this distinction before "rating" anything.
• Fact-checking is not neutrality by both-sidesing. The output is a verdict backed by evidence, not a balanced quote pair.

The verification process — five pillars and a working sequence

The five questions (First Draft / UNESCO): provenance (is this the original?), source (who captured it?), date (when?), location (where?), motivation (why was it shared?).

1. Freeze the evidence. Download the file, archive the URL (Wayback Machine, archive.today) and screenshot with timestamp — deleted posts are the norm, not the exception.
2. Find the original (provenance). Reverse image search on Google Lens, Yandex (strongest for faces and Indian street scenes), Bing and TinEye; for video, extract keyframes with InVID-WeVerify and reverse-search each; sort results by oldest date to find the earliest upload.
3. Interrogate the source. Account age, first posts, handle changes, follower-to-engagement ratio, posting hours (a "local eyewitness" posting on a different time zone's rhythm), whether the uploader says "I shot this" or "received".
4. Date it. EXIF DateTimeOriginal where present (stripped by WhatsApp, Facebook, Instagram and Twitter/X), weather archives for that day and place, shadow direction and length, newspaper or hoarding visible in frame, Google Earth historical imagery, YouTube upload time in UTC.
5. Locate it (chronolocation and geolocation). Match signboard language and script, vehicle number-plate series (state codes such as KL, MH, UP; BH series for Bharat registration), electricity pole and meter markings, uniform and insignia, road markings, shop names, mobile towers, terrain and skyline, then confirm on satellite or Street View.
6. Establish motivation and spread. Who gains; was it seeded in a cluster of pages; is it a reflown hoax (check your own and others' archives first — most "new" claims are repeats).
7. Right of reply. Put the claim to the person or institution named, on record, before publishing; log the attempt if they do not answer.
8. Publish the method. A fact-check is reproducible only if every source, archive link and non-digital step is stated.
9. SIFT (Mike Caulfield), the quick field version: Stop; Investigate the source; Find better coverage; Trace claims, quotes and media to the original context.

Tools and techniques, by evidence type

• Images: reverse search; EXIF viewers (exiftool, Jeffrey's Image Metadata Viewer); error-level analysis and clone detection (FotoForensics, Forensically) — suggestive, never conclusive; crop and search sub-regions separately.
• Video: InVID-WeVerify keyframes and metadata; Amnesty YouTube DataViewer for upload time and thumbnails; frame-by-frame playback for cuts, loops and audio mismatch; check whether aspect ratio shows re-recording of a screen.
• Audio and synthetic media: clipped audio is the cheap attack — always find the full recording. For AI voice and deepfakes look for unnatural cadence, missing breaths, flat room tone, mismatched lip closures, blurred teeth and ear edges, inconsistent blink rate, and warped jewellery or spectacle frames. Detector tools (Hive, Deepware, Reality Defender) give probabilities, not proof; provenance remains the stronger test.
• Documents: verify on the issuing authority's site — PIB releases, eGazette, Sansad question answers, eCourts and Supreme Court judgment portals, MCA21 for companies, RBI and SEBI circulars, data.gov.in, CAG reports, state gazettes. Letterhead forgeries usually fail on reference-number format, date-line, designation and seal.
• Data and statistics: go to the publishing agency — NSO/MoSPI (PLFS, CPI), Census and Civil Registration System, NFHS, NCRB *Crime in India*, RBI *Handbook of Statistics*, Economic Survey, UDISE+, Election Commission. Check the base year, sample frame, revision status and whether the series is comparable over time.
• Documents and data obtained under the RTI Act, 2005 are a primary-source route Indian fact-checkers use heavily for scheme and expenditure claims.
• Field verification stays decisive: a phone call to the SHO, the district collectorate, the hospital superintendent or the named shopkeeper settles cases no tool can.

Standards, codes and Indian law that bear on verification

• IFCN Code of Principles (Poynter): non-partisanship and fairness, transparency of sources, transparency of funding and organisation, transparency of methodology, open and honest corrections policy. Signatory verification is annual and independent; most major Indian fact-check desks are signatories, and signatory status has been the gateway to Meta's third-party fact-checking programme.
• Press Council of India, *Norms of Journalistic Conduct* — accuracy, verification before publication, pre-publication right of reply, prompt correction; applies to the press and is persuasive elsewhere.
• NBDSA (News Broadcasters and Digital Association's self-regulatory authority) hears complaints against news channels; its orders have repeatedly faulted channels for airing unverified visuals and for communal framing.
• Cable Television Networks (Regulation) Act, 1995 and its Programme Code bar content that is likely to incite violence or disturb public order.
• Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Rules, 2021 — Part II due diligence for intermediaries (Rule 3(1)(b) bars patently false and misleading information, impersonation and content threatening public order; grievance officer and timelines); Part III Code of Ethics for digital news publishers and OTT, with a three-tier grievance mechanism. The 2023 amendment that let a government Fact Check Unit flag content about government business was struck down by the Bombay High Court in September 2024 as unconstitutional; litigation on the point has continued, so cite the current status rather than the 2023 text. MeitY has also moved amendments on labelling of synthetically generated (AI) content — check the notified version before relying on specifics.
• PIB Fact Check Unit (since 2019) issues denials about Government of India information; it is a source to be checked like any other, not an adjudicator for journalists.
• Bharatiya Nyaya Sanhita, 2023 replaced the IPC: false statements conducing to public mischief, promotion of enmity between groups, and defamation are the provisions most often invoked against unverified publication. Disaster Management Act, 2005, section 54 punishes a false alarm about a disaster — used during COVID-19 forwards. IT Act, 2000: section 66D (cheating by personation), section 69A (blocking), section 79 (safe harbour conditional on due diligence); section 66A was struck down in *Shreya Singhal v. Union of India* (2015) and must not be cited as live law.
• Elections: Model Code of Conduct, section 126 of the Representation of the People Act, 1951 (48-hour silence period) and section 171G (false statement about a candidate); the ECI's Voluntary Code of Ethics with platforms, and collaborative newsroom alliances (the Shakti Collective in the 2024 general election) govern election-period verification. DPDP Act, 2023 constrains how you handle personal data gathered while verifying.

Output, ratings and corrections

• Anatomy of a published fact-check: the claim quoted verbatim with its source and archive link; where it appeared and how widely; the verdict; the evidence step by step; what remains unknown; the response of the person or body concerned; date, author and methodology note.
• Rating scales: True / Mostly true / Half true / Misleading / False / Fabricated / Unverifiable; some desks use Fake, Misleading, Satire or Missing context. Say *unverifiable* when evidence is absent — absence of proof is not proof of falsity.
• Corrections hierarchy: a *correction* fixes a factual error; a *clarification* removes ambiguity; an *update* adds later developments; a *retraction* withdraws the story. Timestamp the change, keep the original visible, and never silently edit.
• Pre-bunking (inoculation) warns audiences about a manipulation technique before exposure; debunking corrects after. Beware the continued-influence effect — repeating a false claim prominently can reinforce it, so use the truth sandwich: true framing, then the brief false claim labelled as false, then the evidence.
• Limits: verification cannot settle opinion or motive; tool outputs are probabilistic; satire mislabelled as disinformation damages credibility; speed pressure and scale (millions of forwards) mean fact-checks reach a fraction of those who saw the claim; closed networks like WhatsApp are largely unobservable, which is why tiplines matter.

Key points

• Fact-checking - testing checkable claims against primary evidence and publishing a reproducible verdict.
• Verification - establishing authenticity, provenance, date, location and motivation of a piece of content.
• Mis / dis / mal-information - false without intent, false with intent, true but weaponised.
• False context - real content with a wrong frame; the most common form of Indian viral misinformation.
• Five pillars - provenance, source, date, location, motivation.
• SIFT - Stop, Investigate the source, Find better coverage, Trace to the original.
• Reverse image search - Google Lens, Yandex, Bing, TinEye; sort oldest-first to find the original.
• InVID-WeVerify - splits video into keyframes for reverse search and reads available metadata.
• EXIF - camera metadata (time, model, sometimes GPS); stripped by most social platforms, so absence proves nothing.
• Geolocation - signage, script, number-plate series, poles, uniforms, terrain, confirmed on satellite or Street View.
• Chronolocation - dating by shadows, weather records, upload time and visible dated objects.
• Deepfake cues - lip-sync drift, blink rate, ear and tooth edges, flat room tone; detectors give probability, not proof.
• IFCN five principles - non-partisanship, transparency of sources, of funding, of methodology, open corrections.
• IT Rules, 2021 - due diligence for intermediaries plus a Code of Ethics and three-tier grievance system for digital news; the 2023 government Fact Check Unit provision was struck down by the Bombay High Court in 2024.
• Shreya Singhal (2015) - section 66A IT Act unconstitutional; never cite it as live law.
• Section 54, Disaster Management Act, 2005 - penalises false disaster alarms.
• Truth sandwich and pre-bunking - counter the continued-influence effect of repeating a falsehood.
• Unverifiable is a valid verdict - absence of evidence is not evidence of falsity.

Formulas and examples

1. Dating a video by shadow. A frame shows a 3.0 m electricity pole casting a 5.2 m shadow pointing almost due west. tan θ = 3.0 / 5.2 = 0.577 → θ ≈ 30°. A 30° elevation with a westward shadow means mid-morning (sun in the east), so a clip captioned "evening lathi-charge" is inconsistent with its own frame; cross-check the date on a solar calculator for those coordinates.

2. Testing a crowd claim. Organisers claim 1,00,000 people at a maidan rally. Satellite measurement of the fenced area: 120 m × 80 m = 9,600 m². Tightly packed, d = 0.5: N = 9,600 / 0.5 = 19,200. At a realistic 1.0 m²/person: 9,600. Even an unphysical crush (0.25) gives 38,400. Verdict: the claim overstates attendance by roughly 5 to 10 times — report the measured range, not a single number.

3. Speed claim from CCTV. A car is said to be "doing 120 km/h". Footage at 25 fps shows it crossing a 30 m marked stretch in 40 frames. Time = 40 / 25 = 1.6 s; v = 30 / 1.6 = 18.75 m/s = 18.75 × 3.6 = 67.5 km/h. The claim is false; state the ±1 frame uncertainty (≈1.7 km/h) and the assumption that 30 m is correctly measured.

4. A statistics claim. "Crime in the district tripled — up 200%." NCRB-style figures: 1,200 cases in year 1, 2,400 in year 2. Change = (2,400 − 1,200)/1,200 × 100 = 100% (doubled, not tripled). Population also rose from 8,00,000 to 10,00,000, so rates per lakh = 1,200/8 = 150 and 2,400/10 = 240 — a 60% rise in rate. Verdict: *misleading* — the raw rise is 100%, the comparable rise is 60%.

5. Desk accuracy audit. A newsroom published 1,200 stories and issued 18 corrections: correction rate = 18/1,200 × 100 = 1.5%, accuracy rate 98.5%. To audit the remaining stories to ±5% at 95% confidence with p = 0.5: n = 3.8416 × 0.25 / 0.0025 = 384.16 → 385 stories must be re-checked — a sample, which must be reported as a sample with its denominator.

Where it is used

• Dedicated fact-check desks: BOOM, Alt News, Factly, Vishvas News (Jagran), India Today Anti-Fake News War Room, The Quint WebQoof, Newschecker, Fact Crescendo, THIP Media (health), Logically Facts — roles: fact-check reporter, OSINT researcher, Hindi and regional-language verifier.
• Mainstream newsrooms: verification cells inside The Hindu, Indian Express, NDTV, Dainik Bhaskar and PTI; election war rooms during general and assembly polls; business desks verifying market rumours under SEBI's unverified-news norms for intermediaries.
• Platform trust and safety and policy teams: Meta, Google/YouTube, ShareChat, Koo-type apps and Indian OTT compliance teams staffing grievance-officer and Code-of-Ethics obligations under the IT Rules, 2021.
• Government and public health: PIB Fact Check, state police social-media monitoring cells, and health-ministry rumour response during COVID-19 and during vector-borne disease outbreaks.
• Research, corporate and legal: media-literacy programmes (Google News Initiative, Internews, Data Leads trainings in Indian colleges), corporate communications countering deepfaked executives in investment scams, and e-discovery or litigation support authenticating social media evidence.
• Development and NGO work: verification of scheme-delivery claims using RTI replies and open-data portals.

Interview questions

• Q: Difference between misinformation and disinformation? A: Both are false; disinformation is created or spread with intent to deceive or harm, misinformation is shared without that intent. Malinformation is true content deployed to harm, such as a real old clip passed off as current.
• Q: A viral video arrives on WhatsApp with no source. What are your first three steps? A: Save and archive the file; extract keyframes and reverse-search them to find the earliest upload; then test date and place using shadows, signage, number plates and satellite imagery before contacting anyone named.
• Q: The image has no EXIF data. Is it fake? A: No. WhatsApp, Facebook, Instagram and X strip metadata routinely, so missing EXIF is normal and proves nothing either way; provenance and content clues decide it.
• Q: How would you verify an unfamiliar quote attributed to a minister? A: Find the primary record — PIB release, Sansad proceedings, the full video of the event, or the official handle — then check whether the quoted line is clipped from a longer passage, and finally seek the office's confirmation.
• Q: What is geolocation versus chronolocation? A: Geolocation establishes where content was captured from visual clues confirmed on satellite or Street View; chronolocation establishes when, using shadows, weather records, upload timestamps and dated objects in frame.
• Q: Which tool do you use for video verification and what exactly does it do? A: InVID-WeVerify — it splits a video into keyframes for reverse image search, reads whatever metadata survives and helps compare thumbnails and upload times.
• Q: Name the IFCN Code of Principles. A: Non-partisanship and fairness; transparency of sources; transparency of funding and organisation; transparency of methodology; an open and honest corrections policy.
• Q: Which rules govern digital news publishers in India on accuracy and grievances? A: The Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Rules, 2021 — Part III's Code of Ethics and three-tier grievance mechanism for digital news and OTT, with Part II due diligence for intermediaries.
• Q: Can the government itself declare news fake under those Rules? A: The 2023 amendment creating a government Fact Check Unit for content about government business was struck down by the Bombay High Court in September 2024 as unconstitutional; the position has been in litigation, so the current notified text and any Supreme Court order must be checked.
• Q: You cannot confirm or disprove a claim by deadline. What do you publish? A: An *unverifiable* verdict stating exactly what was checked, what is missing and what would settle it — not a guess in either direction.
• Q: How do you avoid amplifying a falsehood while debunking it? A: Use the truth sandwich — lead with the accurate framing, state the false claim once and clearly labelled, then give the evidence; and pre-bunk the technique where the claim is predictable.
• Q: How would you approach a suspected AI-generated audio clip of a politician? A: Treat provenance first — locate the earliest upload and the original full event recording; then examine cadence, breaths and room tone, run detectors for an indicative probability only, and seek the office's denial or confirmation before publishing.

Common mistakes

• Treating the oldest result a search engine *shows* as the original, instead of searching several engines and sorting by date.
• Citing a WhatsApp forward, a screenshot, or another news outlet's report as the source when the primary document or video is available.
• Declaring content fake because EXIF is absent, or genuine because EXIF is present — both are stripped or editable.
• Reporting error-level analysis, "pixel forensics" or a deepfake detector score as conclusive proof.
• Confusing percentage change with percentage-point change, and comparing raw counts across years without normalising for population or a revised series.
• Skipping the right of reply, or publishing before the named person or department has been contacted on record.
• Labelling satire or parody as disinformation, and silently editing a story instead of issuing a timestamped correction.
• Quoting section 66A of the IT Act, or the 2023 Fact Check Unit provision, as current law.

Related topics

News writing and reporting · Editing and headlines · Media ethics and laws in India · Broadcast journalism · Digital and data journalism · Social media and OSINT research