Faces that never existed

Week 5 — deepfakes and synthetic video

USD 25,000,000.

One real person on the call.

Case one: the Hong Kong video call

  • early 2024, an employee of Arup — the firm behind the Sydney Opera House — joins a video call with several “colleagues,” including the CFO
  • instructed to transfer roughly USD 25 million across multiple transactions
  • every participant except the employee was a synthetic reconstruction
  • Arup confirmed the incident to CNN; no research-grade model was required

Case two: the crypto endorsement

  • a deepfake video impersonating a well-known entrepreneur
  • used to promote a cryptocurrency scheme
  • convinced an elderly retiree to invest his life savings
  • the fake didn’t need to fool an expert — it only needed to fool one person, once

Synthesis vs reenactment

  • synthesis — building a face or voice largely from a model’s learned distribution
  • reenactment — puppeting a real, consenting or non-consenting person’s likeness with someone else’s performance
  • different techniques, different artefacts, different defences

What a post-hoc review can and can’t do

  • can: check compression artefacts, audio-visual sync, lighting consistency, metadata trails where they still exist
  • can’t: fully undo what a lossy re-encode already destroyed, or prove a negative from absence of evidence alone

“Just look closely” stopped being useful advice for anyone outside a research lab.

Further reading

CNN’s original report on the Arup case

For discussion

If both cases had happened to your organisation, what single process change — not tool, process — would have stopped the money moving?