Skip to content
The Future Explorerby Benjamin Talin
Enablertechnological

Generative Media

Software that makes pictures, voices and video on request, so convincing that people can no longer tell them from the real thing. In tests, people picked out computer-made faces about as well as flipping a coin, and one system copied a stranger's voice from three seconds of recording. In Hong Kong, an employee made 15 transfers worth HK$200 million after a video call where every colleague on screen was fake. Faking used to be costly; now proving the truth is.

Maturity

4/ 5
measured
BasisGeneration is mature and everywhere; people can no longer tell a synthetic face from a real one. What lags is the layer that would let a genuine recording prove itself - signing at capture, and platforms that keep the signature.SourceNightingale & Farid, PNAS 119(8) (Feb 2022); Hong Kong Police via Hong Kong Free Press (5 Feb 2024); C2PA Technical Specification 2.3 (5 Jan 2026); Center for Democracy & Technology (12 Dec 2025); Regulation (EU) 2024/1689, Art. 50Vintage2022–2026
ReadinessCapability 5Deployment 5Absorption 4Bound bystandardsTrendflat

Where it feeds in

4 megatrends in one force

Where you have already met this

In January 2024 a finance employee in Hong Kong joined a video call with his chief financial officer and several colleagues. None of them were real. He made 15 transfers, HK$200 million in total. The only real person on the call was the one paying.

If this moved

If generation got better still, nothing on the map would move: the four megatrends it feeds sit at 2, held by regulatory clarity, epistemic trust and generational turnover. If provenance were signed at capture and kept on upload, regulation would at least have something to enforce.

What to watch for

Phones that sign photos with Content Credentials by default, at the moment of capture. Platforms that keep those credentials on upload instead of stripping them.

Not to be confused with

Large language models. Both generate; language models produce text, generative media produces pictures and sound - the kind of evidence people were trained to trust on sight.

What it is

Systems that produce images, audio, video and speech on demand, at a cost approaching zero and a quality approaching indistinguishability.

How it works

  1. 01Learn from examplesA model is trained on very large numbers of images, recordings or clips, usually paired with descriptions of what they show.
  2. 02Start from noiseA diffusion model begins with random static and removes it step by step, steered by the prompt, until a picture is left.
  3. 03Work small, then enlargeMost systems do this in a compressed version of the image and decode it to full resolution at the end. That is what makes it fast and cheap.
  4. 04Copy a voice from a sampleVoice systems learn a speaker from a short recording. Microsoft's VALL-E research system needed three seconds.
  5. 05Label it - or notThe provider can attach a signed provenance record or an invisible watermark. Nothing makes the file keep it once it leaves.

State of play

  • Participants told AI-synthesised faces from real ones with 48.2% accuracy - about chance - and training raised that only to 59.0% (Nightingale & Farid, PNAS 119(8), Feb 2022).
  • Microsoft's VALL-E research system imitated an unseen speaker from a 3-second recording (Wang et al., arXiv 2301.02111, 5 Jan 2023).
  • A Hong Kong employee made 15 transfers totalling HK$200 million after a video call in which the other participants were deepfakes; Arup later confirmed it was the company (Hong Kong Free Press, quoting Hong Kong Police, 5 Feb 2024; CNN, 16 May 2024).
  • The Leica M11-P was the first camera to sign photos with Content Credentials (Content Authenticity Initiative, 26 Oct 2023); Nikon suspended its Z6III signing service and revoked every certificate it had issued after a researcher got it to sign an AI-generated image (Digital Camera World, Sep 2025).
  • Most social platforms strip file metadata on upload, which removes provenance records with it (Center for Democracy & Technology, 12 Dec 2025).
  • Article 50 of the EU AI Act has applied since 2 August 2026: providers must mark AI-generated audio, images, video and text in a machine-readable way, with a four-month transition for systems already on the market (Regulation (EU) 2024/1689; Morgan Lewis, Aug 2026).

Why it matters

It collapses the cost of producing convincing evidence. That is a capability change with a direct institutional consequence: verification, which was previously implicit in the cost of faking something, now has to be built explicitly.

  • It feeds four megatrends, all at 2, and holds none of them. Regulation, epistemic trust and generational turnover hold them. None of those is a model.
  • It removed the cost that used to make faked evidence rare. That cost was the verification system, and nothing has replaced it yet.
  • The damage runs both ways. Fakes get believed, and real reports get waved away as fakes - the liar's dividend, which experiments find works mainly against text reports, not video.

What holds it back

Provenance at capture

Signing at the sensor exists on a handful of cameras, and one maker had to revoke every certificate after its signing was fooled.

Metadata stripped on upload

Platforms remove file metadata when a picture is posted, and the provenance record goes with it.

Copyright and liability

Lawsuits over training data - The New York Times sued OpenAI and Microsoft in December 2023 - keep cautious institutions at arm's length.

Detection is an arms race

Every detector is a training signal for the next generator. People are already at chance on faces.

Terms

Diffusion model
A model that generates an image by removing noise from random static, step by step, guided by a prompt.
Deepfake
Synthetic video or audio that shows a real person saying or doing something they did not.
Content Credentials (C2PA)
An open standard for attaching a signed, tamper-evident record of where a file came from and how it was changed.
Watermark
A signal hidden in the content itself that marks it as AI-generated, meant to survive when metadata is stripped.
Liar's dividend
The advantage a liar gains once fakes are common: genuine evidence can be dismissed as fabricated. Experiments find it works mainly against reports in text, not on video.

Steepens

The gradient behaves like a sticky rubber band: it is pulled hardest toward the weakest enabler. Marginal links are shown but do not pull.

If this moved one step

Computed from the map as it stands: this score raised from 4 to 5, every other score unchanged. It is arithmetic, not a forecast.

Raised one step on its own, it lifts no megatrend's weakest link; each gradient it feeds still rises a little.

The gradient also rises in

What if it moved?

What would change the gradient

Conditional, not predictive. No dates, no probabilities.

  1. 01Provenance standards implemented at the camera and microphone, not in post
  2. 02Platform liability for unlabelled synthetic media
  3. 03Nothing on the detection side - that race is structurally unwinnable

Common misconception

That the danger is deception at scale. The more corrosive effect is the liar's dividend - once anything can be fabricated, anything real can be dismissed as fabricated, and that costs nothing to deploy, though experiments find it works mainly against text reports.

Talin, B. (2026). Generative Media. The Future Explorer, dataset v0.4.0 (2026-11-16). https://explorer.talin.digital/e/generative-media/

Generative Media: maturity 4/5 · The Future Explorer