1. Quick Summary
Synthetic audio and video can now be produced cheaply enough that forgery is no longer a specialist activity. Detection attempts to tell generated media from real recordings.
Detection works, but it degrades: as generation improves, the artefacts that detectors rely on disappear, so any detector has a limited shelf life.
2. What It Means
Statistical detection looks for inconsistencies that generation processes leave behind, such as unnatural blinking, odd lighting or audio-visual mismatch.
Provenance is a different strategy: recording where media came from and whether it was modified, rather than trying to infer it afterwards.
Both are imperfect. Detection gives a probability, not a verdict, and provenance only works where it was applied from the start.
3. Why It Happens
Generative systems are trained to minimise exactly the artefacts detectors look for, so each improvement in generation tends to reduce detector accuracy.
Compression and reshares destroy the fine signal details that detectors use, so performance in the wild is far worse than in laboratory tests.
Distribution shift is the practical killer: a detector trained on one generator often fails on another.
Provenance standards attach signed capture information to media at creation, which is more durable but requires cameras, platforms and publishers to adopt it.
Watermarking helps but is fragile, since it can be removed by re-encoding and does not survive every editing workflow.
Human judgement is unreliable and gets worse as quality improves, which is why the emphasis has shifted toward context: who shared it, when, and whether independent sources corroborate.
4. Real Examples
Artefact-based detectors: models trained to spot inconsistencies left by generation.
Audio-visual consistency checks: testing whether lip movement and speech really match.
Content credentials: signed records of capture and edit history attached to a file.
Watermarking: embedded markers that can be damaged by normal processing.
Provenance chains: platforms displaying where an image came from and what was changed.
5. How It Affects Us
Journalism: verification workflows now routinely include provenance checks.
Platforms: labelling and provenance display are becoming expected features.
Law and policy: rules are emerging, though definitions and enforcement remain difficult.
Public behaviour: healthy scepticism toward unverified media is now a basic skill.
6. Key Takeaways
- Detection is a moving target because generation improves against it.
- Real-world performance is much lower than laboratory results.
- Provenance recorded at capture time is more durable than detection after the fact.
- Corroboration and source context remain the most reliable checks.