A routine you can actually run

Why a written routine

Everything in this module is knowable and most of it is known by people who nonetheless get it wrong. The failure is not ignorance. It is that verification happens under pressure — a story is breaking, a client is waiting, a post is spreading — and under pressure people skip steps and trust their eyes, which is exactly what does not work.

A checklist removes the judgement about whether this case needs checking. Aviation, surgery and newsrooms all reached the same conclusion for the same reason.

The routine

Before anything: preserve. Save the file. Record the URL, the account, the time you found it, and the claim being made. Screenshot the surrounding context. Content vanishes.

Step 1 — What is the claim? Write it in one sentence. "This shows X happening at Y on Z date." If you cannot write it, there is nothing to verify and you are reacting to a feeling.

Step 2 — Where has this appeared before? Reverse image search on at least two services. Sort by oldest where available. An earlier appearance ends the process.

Step 3 — Who is the source? Named or anonymous, account age, posting history, whether anyone accountable stands behind it. Find the original poster, not the account you saw it on.

Step 4 — Does the place check out? Satellite and street imagery against the visible geography. Signage, language, vehicles, side of the road, architecture, vegetation.

Step 5 — Does the time check out? Shadows, weather records, seasonal cues, anything dated in frame.

Step 6 — Is there corroboration? Independent recordings, credible reporting, official statements. Absence of corroboration for a major claim is itself a finding.

Step 7 — Provenance and metadata. Check for content credentials. Check metadata. Remember both are informative when present and meaningless when absent.

Step 8 — Only now, look at the image. Any artefacts consistent with the mechanisms in module three. Treat this as the weakest evidence you have collected, because it is.

Step 9 — Record the verdict as one of three. Verified, debunked, or unverified — with the reasons written down.

The rules that go with it

Unverified is a legitimate stopping point and the most common one. Do not upgrade it to either certainty because a decision is needed.

Never accuse in public on inspection alone. The cost of a false accusation falls on a real person and is not recoverable.

Time-box it. Decide in advance how long you will spend. Thirty minutes settles most cases; the ones it does not settle are rarely settled by three hours.

Say what you did. When you publish a conclusion, publish the method: "the image first appears in 2019 in an archive of a different event". This is far more persuasive than authority, and it lets others check you.

Adapting it

For personal use, steps 1 to 3 catch the overwhelming majority of misleading material circulating in group chats, and take under two minutes. That short version is the one worth teaching to family, and it is more useful than any amount of explanation about how the models work.

For a newsroom or an organisation, the full routine with a named person responsible and a written log is the professional standard, and the log matters as much as the checks — it is what lets you defend a decision later.

For a moderation or trust-and-safety context, the same routine plus scale: automated triage that routes to this process rather than deciding, for exactly the base-rate reasons from the first lesson.

The last word on this module

Nothing here requires you to identify a generated image, and that is the point. The methods that work are about origin, source and corroboration; the methods that expire are about pixels.

That is a more optimistic conclusion than most writing on this subject reaches. The tools for establishing where something came from are free, they are the same ones journalists have used for years, and they do not become obsolete when a better model is released. What they require is the willingness to spend two minutes and to accept "unverified" as an answer, which is a matter of discipline rather than of technology.