Copy article

Deepfake defence is becoming a VIP feature. That is a trust problem, not a tooling one

ended 13. March 2026

YouTube says it is expanding its likeness detection tool to a pilot group of journalists, politicians and civic leaders. In practical terms, this is Content ID for faces: spot synthetic impersonation and request removal.

That is a sensible step, but it exposes a bigger problem. Deepfake risk is not evenly distributed. Public figures get platform tooling, verification lanes, and escalation paths. Everyone else gets a form and a wait. Yet the most common harm is not a fake prime minister speech. It is harassment, non-consensual edits, and local reputation damage that never makes headlines.

There is also a governance trade-off baked into the fix. The system needs high-quality reference data and identity checks to work. That means more people submitting ID and video to platforms that already struggle with trust. Protection becomes another data-collection pipeline.

The future fight is not just detection. It is provenance, accountability, and speed. When a clip can travel across platforms in minutes, a two-week review cycle is not a safety feature.

We'd like your views:

  • Should 'likeness protection' be a right for all users, not just high-profile ones?
  • What is the least-bad way to verify identity without building a new surveillance asset?
  • Who should be liable when a platform fails to act on obvious synthetic impersonation?
  • Do platforms need shared provenance standards, or will this stay a walled-garden tool?
  • How fast is fast enough for takedowns when the harm is personal and immediate?

1 responses from the Newspage community

Copy all

Star Quote
Copy

Rohit Parmar-Mistry
Founder at Pattrn Data
YouTube's likeness detection pilot is a sensible move, but likeness protection should not be a VIP perk. If a platform can recognise a politician's face at scale, it can recognise a teacher, nurse, or teenager too. The most common harm is not a fake PM speech, it is harassment, non-consensual edits, and local reputation damage.

The hard bit is governance, not detection. To make this work, platforms want high quality reference data plus ID checks. That is another data pipeline from users to companies that already struggle to earn trust. In our AI audits, the failure mode is always the same: no evidence trail you can use in a complaint, slow human review, and nobody accountable when the clip has already spread.

Minimum bar: identity checks that minimise data, a rapid takedown lane for obvious impersonation, and shared provenance standards across platforms. Under the UK Online Safety Act (and the EU DSA), "we are reviewing" cannot mean two weeks. Speed is the product.