A Practical Verification Workflow for Newsrooms

The most common mistake in visual verification is starting with the pixels. Pixel forensics is powerful but slow and ambiguous; provenance work is fast and frequently decisive. Order matters.

Step 1 — Establish provenance first

Before touching any tool, answer: where did this file come from? Ask the source directly for the original, unmodified file, not a forward. If the answer is “someone sent it to me,” you are already dealing with a copy of a copy, and every forensic signal has been degraded.

The single highest-value action in verification is obtaining the original file.

Step 2 — Reverse image search

Run the image through multiple reverse-search engines. You are looking for earlier appearances. An image presented as breaking news that appeared three years ago is resolved in under a minute — no forensics required. This step alone catches a large share of recirculated misinformation.

Step 3 — Check the claim against the frame

Does the content match the caption? Verify:

• Geography — architecture, road markings, signage, vegetation, licence plate formats.
• Weather — cross-reference against historical records for the claimed date and location.
• Time — shadow direction and length against sun position for the claimed place and time.
• Language — scripts and dialects visible on signage.

Most false imagery is authentic imagery with a false caption. Forensics cannot detect that. This step can.

Step 4 — Now run forensics

With provenance established and context checked, forensic analysis tells you whether the file itself has been altered. Run it on the original file. Read the layer breakdown, not just the score. Look at the ELA map for regional anomalies rather than overall brightness.

Step 5 — Document everything

Record what you did, in order, with results:

• Source and how the file was obtained
• Hash of the original file
• Reverse-search results with dates
• Contextual checks performed
• Forensic output, including the layer breakdown
• Your conclusion and confidence level

If a correction becomes necessary later, this record is what protects you.

Step 6 — Publish with calibrated language

Match your language to your evidence:

• “We verified” — provenance established, source confirmed, multiple checks agree.
• “We could not verify” — checks were inconclusive. Say so plainly.
• “This image shows signs of manipulation” — forensics flagged specific anomalies, described specifically.

Never write “AI detected this image is fake.” No tool detects fakeness. Tools detect artefacts, and artefacts have innocent explanations.

The organisational part

Make one person accountable for visual verification on any story that turns on an image. Diffuse responsibility produces unverified publication. Build a fifteen-minute check into the workflow and treat it as non-negotiable — because a correction costs far more than fifteen minutes.