How to verify a conflict claim before publication

Confidence in a conflict report can be strong in one respect and weak in another: authentic footage may be old, a genuine record may be misread, and a witness may know one street but not events elsewhere. Record the checks separately, then consult the international events conflict guide for the broader reporting context.

Claim origin: Weak version and Better version

Weak version: A desk picks up the most widely shared post and treats its reach as evidence that the claim is established.

That loses the trail back to the first available account and blurs the difference between an original assertion and later copies, each of which has its own publication time.

Better version: Save the earliest available post, file, document or statement with its URL, account, timestamp and surrounding context. Record reposts separately. A recaptioned clip may show real footage from a different conflict, country or year, so tracing its origin matters more than choosing the most dramatic copy. Use the breaking news verification checklist for general source fields, then add checks specific to conflict claims.

Caption evidence: Weak version and Better version

Weak version: A caption names a place, date and military activity, and readers assume the picture confirms all three. A genuine image can be made to support an unrelated claim when its context is removed or replaced, even if the photograph itself has not been altered.

Better version: The photograph shows a military relief convoy passing damaged buildings in Pago Pago, American Samoa, on October 1, 2009, after the Samoa tsunami. It is unrelated to armed conflict. A caption that identifies the event, location, date and mission keeps a real photograph from being mistaken for evidence of something it never showed.

Media dating: Weak version and Better version

Weak version: Treat the file's metadata as a clock that cannot lie, or assume missing information proves the file is fake. Metadata may be absent, changed, or remain at a device's default setting; one timestamp cannot settle where or when footage was made.

Better version: Inspect file history and metadata when opening the file is safe. ExifTool reads embedded capture time, camera model and GPS coordinates without altering the original. Search key frames with Google Images, TinEye or InVID-WeVerify, then test landmarks, road geometry, terrain, weather records, shadows, spoken language, insignia and earlier appearances.

Compare satellite or map imagery only when its date is known, using dated satellite imagery. Copernicus Sentinel imagery can help check a location, while SunCalc can test shadows against the claimed date and solar position. The Berkeley Protocol, written with the UN human rights office, sets minimum professional standards for open-source investigation standards and evidence preservation. A newsroom is not a legal investigation, but a careful record keeps findings usable when accounts conflict later. Read the Berkeley Protocol.

Timing: Weak version and Better version

Weak version: Publish a precise capture hour because the upload or statement has one. Those times refer to different events, and visible conditions or supporting records may only establish a range. A single unsupported time gives readers more precision than the available evidence can defend.

Better version: Keep five clocks apart, then publish the earliest and latest possible capture time supported by visible conditions and corroborating records. The upload time, the event time, the statement time and the time you checked the claim are separate from capture. A defensible window is more informative than a guessed hour.

Death counts: Weak version and Better version

Weak version: Combine every number attached to a conflict into one total, even when the reports count different people or use different evidence. A total without a closing date or counting method can sound definitive while concealing unresolved duplicates and incompatible categories.

Better version: Ask four questions of every figure: who is included, what evidence qualifies, how duplicates are removed, and when the count closes. Hospital admissions, named deaths, missing-person reports and party claims describe unlike categories, so do not merge them into one total.

Brown University's human costs overview estimates more than 940,000 people killed by direct post-9/11 war violence through 2023, counts civilians separately, and notes that indirect deaths are far larger and harder to establish. Treat this as a human-cost estimate with a defined method. For related newsroom checks, use the conflict reporting checklist.

Independent checks: Weak version and Better version

Weak version: Count two outlets repeating the same official statement as two confirmations. If both accounts rely on one source chain, the apparent agreement adds no independent access and may simply carry the original claim farther.

Better version: Seek sources with different access and incentives: local authorities, medical facilities, aid agencies, monitors, residents, remote sensing and official records. Check conflict event databases such as ACLED, UCDP and Airwars, which covers airstrikes and civilian harm. Courts and treaty bodies can provide direct findings, including ICC charges and arrest warrants or OPCW material concerning chemical-weapons allegations. The same-source trap also appears in live update problems.

Risk of exposure: Weak version and Better version

Weak version: Publish every identifying detail because it makes an image or account easier to verify. Names, faces, landmarks, routes, radio details and file metadata can expose people, while graphic evidence may cause harm without being necessary to explain what is known.

Better version: Remove identifying details when they could put people at risk, and consider consent alongside the power imbalance between the newsroom and the person providing information. Describe graphic evidence when showing it is unnecessary. Apply the public safety alert checklist to details that could reveal a person's identity or location.

Publication wording: Weak version and Better version

Weak version: Use one confidence label for every part of a claim, or declare responsibility when the available evidence establishes only that an event occurred. That wording hides which part was checked and turns an unresolved question into an apparent conclusion.

Better version: Choose among said, reported, verified, corroborated, disputed and could not independently confirm according to the work completed. State what was checked rather than implying total certainty. If responsibility remains unclear, describe the verified event, attribute competing claims to their sources and say plainly that responsibility is unconfirmed. Keep status language consistent with the breaking updates and timelines guide.