Verifying synthetic media: a working protocol for legal teams

A photograph, a voice note or a short video can now be generated or altered with tools that are cheap and widely available. The results can pass a casual look and, in some cases, a careful one. For legal teams this creates two risks. Fabricated material may be relied on as genuine. Genuine material may be dismissed as fabricated by someone who has reason to deny it. Both risks call for the same response, a structured and documented check before the file is used.

The protocol in outline

The order matters. Provenance and preservation come first, because later checks are worth little if the file has changed or its origin is unknown.

• Provenance and custody. Record who supplied the file, when, how it reached you and who has handled it since.
• Preservation. Keep the original file unaltered, calculate its cryptographic hash and work only on copies.
• Metadata and credentials. Examine embedded metadata and any content credentials, knowing that either may be absent, stripped or altered.
• Earlier versions. Run reverse image searches on the file, and on keyframes from video, to find earlier or different versions.
• Consistency and corroboration. Test the content against itself and against independent records of the time and place it claims to show.
• Graded conclusion. Treat automated detection as one weak signal, and report a conclusion with a stated level of confidence.

Provenance, custody and preservation

The first step is to establish who provided the file, whether they created it or received it from someone else, and when and through what channel it arrived. A file forwarded through a messaging application has usually been recompressed and stripped of much of its metadata. A file taken directly from the device that recorded it carries more information, and the device itself may hold related material.

Preserve the file exactly as received. Calculate a hash (a digital fingerprint produced by an algorithm such as SHA-256) at the point of receipt and record it, so that any later copy can be shown to be identical to the original. Carry out every subsequent step on a working copy, and log every transfer. This is ordinary chain of custody discipline, and it allows the analysis to be explained and defended later.

Confidentiality applies throughout. Uploading a client's file to a public search or analysis service may disclose it to a third party. Decide in advance which tools may be used on which material, and take advice where privilege or data protection is in question.

What the file says about itself

Metadata can record the device, software, time and sometimes location of capture. It is also easy to remove or edit, and many platforms strip it on upload. Its presence is useful, its absence proves nothing, and metadata that looks consistent can itself be fabricated. It should be treated as a claim to be tested.

Some cameras, editing tools and generators now attach content credentials based on the C2PA standard. These record a file's origin and the edits made to it in a cryptographically signed form. Where credentials are present and valid, they are useful evidence of the file's history as recorded by the tools that signed it. A valid signature shows the record has not been altered. It does not show that what the file depicts is true. Where they are absent, that tells you very little, because attaching them is optional and credentials are often lost when a file is shared.

Searching, comparing and corroborating

Reverse image searches, run on more than one search engine and on keyframes from video, can reveal earlier versions of an image or show that it first appeared in another context. An earlier, different version of a file is often the most decisive finding available.

Internal consistency comes next. Light, shadows and reflections should agree with each other and with the scene. Text, signage and fine detail should hold up under magnification. In audio, background noise, room acoustics and breathing should be continuous, and any cuts should have an explanation. The artefacts that betray generated material change as the tools improve, so no checklist of flaws stays current for long.

External corroboration tests the file against the world. Weather records, the position of the sun, landmarks, building works and other recordings of the same event can confirm or contradict the time and place a file claims to show.

Detection tools and the conclusion

Automated detection tools estimate whether a file was generated or manipulated. They are trained on particular techniques, perform unevenly on new ones and can be confused by compression or editing. Their output is one weak signal among several.

A detection score is one weak signal, and it should never decide the question alone.

The conclusion should be graded and explained. It should state what was checked, what was found, what could not be tested and what level of confidence the evidence supports. Often the honest result is that a file is consistent with being genuine and that nothing found suggests manipulation. That falls short of proof, and the report should say so.

We approach media verification as an evidential exercise with a technical component. Provenance, preservation and corroboration do most of the work, and we record each step so that a court, a regulator or opposing counsel can follow it. Where a file cannot be verified, we say so and explain why, because an unsupported conclusion in either direction can cause real harm.