Hany Farid spent twenty-five years teaching machines — and people — how to detect manipulation. Now he’s warning that detection is no longer a reliable foundation.
In a recent profile by The New York Times, the UC Berkeley professor and digital forensics pioneer admitted he’s struggling to distinguish real media from AI-generated content.
“I feel like I’m going blind,” he said.
Taken literally, that sounds like an individual limitation. It isn’t. It’s a signal that the assumptions underlying modern verification are breaking.
When Detection Stops Scaling
For years, Farid’s work showed that humans are poor judges of authenticity. That was manageable as long as expert analysis could compensate.
What’s changed is not just quality — it’s scaling dynamics.
Tools like ChatGPT, Gemini, ElevenLabs, and Veo 3 have shifted synthetic media from a specialized capability to a default one. The constraint is no longer skill. Its intent.
Detection systems, by contrast, scale reactively. They require known patterns, training data, and time to adapt. Each improvement in generation resets the baseline that the detectors are trying to learn.
This is not a temporary gap. It’s a structural asymmetry:
- Generation improves with usage and capital
- Detection improves with delay and hindsight
That difference compounds.
The Speed Mismatch
Even if detection were perfect, it would still fail in practice.
Farid’s long-standing argument is that information online stabilizes within minutes, not hours. The first version of a piece of content — true or false — defines its reach, framing, and perceived legitimacy.
By the time verification happens, distribution is already complete.
This creates a hard constraint: truth that arrives late has near-zero corrective power.
That’s why synthetic media doesn’t need to be flawless. It only needs to be credible long enough to spread.
Where the Model Breaks
The implications are not theoretical.
Audio clones have been used to authorize financial transfers. Fabricated footage has circulated through media pipelines before verification. Synthetic content is beginning to interact directly with systems that were designed under the assumption that “someone will catch it.”
That assumption is now unreliable.
The failure point isn’t that detection never works. It’s that it doesn’t work fast enough, consistently enough, at scale to support the systems built on top of it.
Newsrooms, courts, financial institutions, and election infrastructure all depend — implicitly or explicitly — on delayed verification. That delay used to be tolerable. It no longer is.
Why This Was Inevitable
The deeper issue is incentive alignment.
Platforms optimize for speed, engagement, and volume — conditions that amplify unverified content. Strong verification introduces friction, and friction directly reduces those metrics.
At the same time, the cost of generating synthetic media is collapsing, while the cost of verifying it remains high and labor-intensive.
That creates a predictable outcome:
- More content
- Less friction
- Weaker verification relative to volume
The system doesn’t fail suddenly. It degrades until reliability can’t be assumed.
After Detection
Farid is leaving Berkeley and returning to Dartmouth College. The move itself is routine. The timing, given his public admission, underscores something less routine: the field is confronting limits it can’t engineer around quickly.
If detection can’t carry the load, the model has to shift.
That means moving from “analyze after the fact” to “establish authenticity at the point of creation” — through cryptographic provenance, controlled capture environments, and systems where verification is embedded, not appended.
The technology for this exists. The adoption does not.
And adoption is not a technical problem. It’s a coordination problem across platforms, governments, and institutions that don’t share incentives.
What Changes Now
Farid’s warning is not that reality is unknowable. It’s that unauthenticated reality is no longer trustworthy by default.
That’s a different claim, and a more actionable one.
In that world, the burden shifts:
- From proving something is fake
- To prove something is real
Most systems in use today are not built for that shift.
They will be forced to adapt — not because the technology has matured, but because the old assumption failed under pressure.
The transition won’t happen cleanly. It will happen through errors, losses, and institutional misjudgments that expose the gap.
Detection didn’t disappear. It just stopped being enough.
Related: Think for Yourself in the Age of AI: The Cognitive Sovereignty Survival Guide
