Pathway: AI in policing and public safety Module 3 of 4

Deepfakes and digital evidence

How fake audio, video and images can appear in complaints and evidence and how to preserve, verify and refer digital material to forensic experts.

Last reviewed: . Reviewer: Muhammad Asadullah.

Why this matters

Members of the public increasingly bring recordings, screenshots, photos and videos to the police. Many are genuine and important. Some may be faked. AI tools can now produce realistic deepfakes: fake video of a person saying or doing something, fake images and voice cloning that imitates a real person's voice from a short sample.

This matters in two directions. A fake recording can be used to accuse an innocent person, harass someone or support a false complaint. And a genuine recording can be dismissed by someone who simply claims "it is a deepfake". Officers and staff who handle complaints and evidence need to treat all digital material with care: neither accepting it at face value nor rejecting it without proper checks. This module gives general guidance only. Always follow your force's procedures and advice from forensic and legal colleagues.

Where you might meet fakes

  • Complaints that include audio or video of a threat, an insult or an admission.
  • Blackmail and harassment using fake intimate or compromising images.
  • Fraud and scams using a cloned voice of a relative, an official or a business contact. The deepfake and voice scams guide explains these for the public.
  • Misinformation shared online during an incident, including old or altered footage presented as new.
  • Material submitted as evidence in an investigation.

Warning signs and their limits

Some fakes show signs such as unnatural blinking or lip movement, odd lighting or shadows, blurred edges around a face, mismatched audio or a voice with strange rhythm. Context can also help: does the date, place, weather or clothing match what else is known?

But these signs are not reliable. Good fakes may show none of them and genuine recordings can look odd because of compression or poor equipment. Do not decide that material is real or fake from how it looks or sounds. Online "deepfake detectors" are also not a reliable basis for any decision and uploading material to them can breach confidentiality.

Preserving digital material

How material is handled from the start can decide whether it can later be verified. General good practice:

  • Get the original where possible, not a forwarded copy or screenshot. Forwarding through messaging apps often strips information and reduces quality.
  • Record where it came from: who provided it, when, how they got it and on what device.
  • Do not edit, crop, enhance or convert it. Work on copies, never the original.
  • Keep a record of everyone who handles it, following your force's procedures for exhibits and continuity.
  • Do not upload it to public AI tools or online services to "check" or "improve" it. That can alter it, leak it and harm the people in it.

Information about where material came from and how it has changed is sometimes called content provenance. Some tools add watermarks to AI-generated content, but their absence proves nothing.

Verification and referral to experts

Deciding whether material is genuine is a specialist task. Refer to qualified digital forensic experts through your force's procedures when:

  • the material may be used as evidence
  • someone disputes that it is genuine
  • it is central to an allegation against a person
  • there are signs of editing or doubts about its source

While waiting, treat the material as unverified. Look for other evidence that supports or contradicts it, such as witnesses, other recordings or records. An allegation should not rest on an unverified recording alone.

Treating people fairly

People shown in fake material, especially intimate images, may be victims of serious harm. Handle such material sensitively, limit who sees it and support the person through the right channels. People who report a fake should be taken seriously, not dismissed.

Try it

Find a published fact-check from a reputable news organisation about a fake or altered image or video. Note how the fact-checkers established it was not genuine. Which of those methods relied on context and other sources rather than how the material looked?

Check yourself

  1. Why should you not decide whether a recording is fake from how it looks or sounds?
  2. Name three good practices for preserving digital material.
  3. When should digital material be referred to forensic experts?

Answers: (1) Good fakes may show no signs and genuine material can look odd because of compression or poor equipment. (2) Any three of: get the original, record where it came from, do not edit it, work on copies, keep a handling record, do not upload it to public tools. (3) When it may be used as evidence, its authenticity is disputed, it is central to an allegation or there are signs of editing or doubts about its source.