Synthetic media verification: a working guide for reporters

What to take away

Synthetic media is a family of things, not one thing. It covers several kinds of synthetic media:

That range is why "is this AI?" is the wrong first question. A newsroom has to establish what file it holds, where the file came from, what it claims to show, and whether the evidence supports that claim.

Define the claim before inspecting pixels

Write the claim as testable fields:

Separate the findings these fields produce. A clip can be authentic and falsely dated. A real voice recording can be laid over unrelated images. A generated picture can illustrate a true event while being passed off as documentary evidence.

Each of those is a different verdict, and each needs different wording in print.

Preserve the evidence package

Save the highest-quality file you can get without re-encoding it. Record the URL, the account name and identifier, the post text, and visible engagement. Also record upload time with its time zone, and your download time.

Capture the whole page, not just the media. Note any edits or platform labels applied since publication. If the publisher hands over an original through a secure channel, keep that copy separate and write down how it reached you.

Hash every file so later copies can be compared. Keep working copies apart from preserved originals. Before uploading sensitive unpublished material to a public detector, read that tool's retention and privacy terms.

Build four evidence lanes

1. Source and distribution

Find the earliest known appearance. Trace reposts backward, compare captions, and ask whether the first visible account had access to the claimed scene or person.

Contact that account through an independently confirmed channel. Ask for the original file, the capture device, an approximate time, and the circumstances. Do not reveal every detail you expect to hear back.

2. Event and context

Check schedules, weather, daylight, and architecture. Check language, clothing, official statements, and live streams. Check other independent recordings. Context often settles a claim without deciding how the pixels were made.

A person shown somewhere else at the claimed time disproves the caption even if the production method stays unknown.

3. Provenance and file history

Inspect embedded metadata and any signed provenance record. Ask whether the credential validates, who signed it, what actions it describes, and whether it belongs to this exact file.

Missing metadata is ordinary after platform processing. It is not evidence of fabrication.

4. Media forensics

Examine these for continuity:

A single odd hand, shadow, or waveform is a lead, not a verdict. When several such leads point the same way, you have something worth reporting. When they conflict, say so.

The National Institute of Standards and Technology's report on technical approaches to synthetic content transparency separates provenance, labeling and watermarking, detection, testing, and auditing. No single method answers every verification question.

Use detectors as measured instruments

Before you rely on a detector, record five things. Record its version, the media types it supports, and its stated training scope. Also record what its output actually means and its known error rates. Run the best available file, not a screenshot, when the original exists.

Compare tools that use different methods. Do not treat several opaque scores as independent confirmation of each other.

A score reading "82 percent AI" may describe a model output, not an 82 percent probability that the claim is false. Cropping, compression, subtitles, noise removal, and platform transcoding all move the number. Test an authentic comparison file from the same device or distribution path where you can.

The field measures itself this way. The NIST Open Media Forensics Challenge scores whether a system can decide an image or video was manipulated, then whether it locates the altered region and names the manipulation type, using benchmark datasets on shared evaluation infrastructure.

A tool worth quoting has been measured on something, and the measurement has conditions attached.

Reach a bounded finding

Useful conclusions include:

Name the decisive evidence. Avoid "looks fake," "an AI detector confirmed it," and "no metadata proves it was generated." Where a finding could harm an identifiable person, seek a second review and give that person a fair chance to respond.

Publish a verification note

Set out the exact claim. Set out the material examined. State whether you had the original or a repost. List the tools and versions. Give independent context. Give source contact. Give limitations. Give your evidence cutoff.

State your revision policy. Keep detector output in the reporting file, but give readers the reasoning rather than a dashboard of unexplained numbers. When a claim survives to this stage, the remaining disputes usually follow a handful of familiar patterns, and it helps to know common visual verification problems before you publish.

Common questions

Is every deepfake fully generated?

No. A deepfake can alter part of authentic media, such as a face or a voice, while leaving the rest untouched. That distinction changes what you can safely write.

Does missing metadata prove a file is fake?

No. Social platforms, editing software, messaging services, and screenshots routinely strip or rewrite metadata. Treat its absence as a gap in the record, not as a finding.

Can one detector settle the question?

No. Treat a detector score as one reproducible signal and test it against provenance, context, source access, and forensic observations. Where those disagree, report the disagreement.

What if the source will not provide the original?

Report that limitation plainly. You may still verify context, but your conclusion should not imply access to evidence you never examined.