Deepfake detection and synthetic media: verification beats guessing
A source-backed guide to recognising and verifying synthetic media, covering provenance, labels, contextual checks, detection limits and the growing gap between “looks real” and “is verified”.
Last reviewed 2026-07-25
Can people reliably spot a deepfake by eye?
Not reliably. Ofcom research found only a small minority of UK adults were confident in their ability to identify deepfakes, while substantial numbers reported encountering them. Visual oddities can still be clues, but modern synthetic media can survive casual inspection. Verification should therefore focus on source, provenance, independent confirmation and context rather than a hunt for strange fingers or blinking.
What is the best way to verify suspected synthetic media?
Start by preserving the original link or file, then identify the earliest credible source and compare the claim with independent reporting or official channels. Check whether reliable provenance metadata exists, examine frames or audio only as supporting evidence, and use reverse search where practical. For high-stakes content, treat automated detector scores as one signal rather than a verdict and document what was checked.
Do AI labels and watermarks solve the problem?
They help, but they are not a universal truth layer. Ofcom’s attribution work examines watermarking, provenance metadata, AI labels and contextual annotations, each with different strengths and weaknesses. Labels can improve disclosure when platforms and generators cooperate, yet content may be re-encoded, stripped of metadata or created outside compliant systems. Absence of a label is not proof that media is authentic.
Why is deepfake detection still difficult?
Detection is an adversarial problem: generation improves, content gets compressed or edited, and detectors can be trained on yesterday’s artefacts. The UK government described the deepfake detection market as still nascent in March 2026. That makes process more durable than product worship. Organisations need escalation rules and corroboration methods that still work when a particular detector becomes unreliable.
What does the current evidence say?
- Ofcom reported that 43% of people aged 16+ said they had seen at least one deepfake online in the previous six months, while 9% said they were confident identifying one. Source
- Ofcom reported in July 2025 that 85% of adults supported AI labels on content, while 34% said they had seen such a label. Source
Limitations
Encounter surveys measure what people believe was a deepfake, not a forensic confirmation of every item. Detector performance also varies by medium, model, compression and manipulation type.
A counterpoint worth keeping
A “deepfake detector” can create false confidence. In high-stakes cases, provenance and corroboration may be more useful than a single probability score, especially when the media has been repeatedly edited or reposted.
Sources and provenance
- Ofcom · A deep dive into deepfakes that demean, defraud and disinform · 2024-07-23 · Primary source
- Ofcom · Deepfake Defences 2 - The Attribution Toolkit · 2025-07-11 · Primary source
- Department for Science, Innovation and Technology · Deepfake detection technology · 2026-03-26 · Primary source