What Happens When AI Gets Better at Creating Fakes Than We Get at Detecting Them?

Honestly? In a lot of cases, we're already there.
Deepfake detection has gotten better over the past few years - no argument there. But the tools used to generate fakes have gotten better faster. And that's not a gap that's going to close with the next model update or the next funding round. It's baked into how this whole race works, and it's worth understanding why if you're counting on detection to protect you.
Why generation keeps winning
Think about what each side actually has to do. A fake only needs to work once — one convincing clip, one viral moment, and it's already been seen by however many people see it before anyone raises a flag. Detection doesn't get that luxury. It has to catch nearly everything, every time, or the fakes that slip through are the ones that do the damage.
There's a second problem, and it's a bit uglier: generation tools can be trained directly against detectors. Someone builds a detector, someone else runs their fake generator against it, sees what gets flagged, and tweaks the model until it doesn't. Every improvement on the detection side quietly becomes a training signal for the other side. And detectors, by their nature, are trained on fakes that already exist — they can't really prepare for a technique nobody's invented yet.
So it's not that detection teams are doing bad work. It's that the whole setup asks them to defend against an opponent that gets to see the playbook first.
So what happens if detection just... can't keep up?
A few things, and none of them are great. 
"Probably real" stops meaning much. A confidence score is only reassuring up to a point, and that point keeps moving as fakes get more convincing.
The burden of proof also ends up in the wrong hands. Right now, when a fake surfaces, it's usually the person or brand being faked who has to scramble to prove it isn't true — often after most of the damage is already done. That's backwards, if you think about it.
And trust doesn't disappear evenly. People don't stop believing things altogether — they just get pickier in inconsistent, gut-feeling ways, which arguably makes things worse, since existing bias ends up doing the job verification should be doing. Meanwhile, whole systems — journalism, courts, personal reputation — were built assuming a photo or recording meant something. Take that away, and they're all standing on ground that isn't quite there anymore.
Maybe the fix isn't a better detector
If detection is always going to be a step behind, throwing more resources at it just buys time — it doesn't close the gap. That's because the real issue isn't a technology problem, it's a timing problem. Detection only ever gets to act after content already exists. The only way around that is to act before it spreads in the first place.
That's the thinking behind Vastavik, an independent platform for validating and verifying digital media authenticity. Rather than trying to win a race that's rigged against it, Vastavik steps out of it entirely — creators validate their content (photo, video, audio, whatever it is) right at the point they create or publish it, which creates a public, independently checkable record from that moment on.
It genuinely doesn't matter how good fake generation gets after that, because verification was never trying to catch anything. It just gives real content a way to speak for itself.
Put simply: detection asks "does this look fake?" — a question that gets harder every year. Verification asks "can this be proven real?" — a question that doesn't get any harder, no matter what generation tools can do next.
What to actually do with this
The detection-vs-generation race probably isn't going to resolve itself in detection's favor anytime soon, so it's worth adjusting how you think about this now rather than later.
Stop trusting your gut on "this looks real" or "this looks fake" — that instinct is getting less reliable, not more, every year. If content is part of how you make a living, treat verification the way you'd treat backing up your files: set it up before you need it, not scrambling afterward. And if something viral crosses your feed, look for an actual independent source rather than judging it on how convincing it seems.
Frequently asked questions
Is deepfake detection getting worse?
Not really worse on its own - detection keeps improving. It's just that generation is improving faster, and since a fake only needs to work once while detection needs to work almost every time, the practical gap keeps widening.
Will AI detection ever fully catch up to AI generation?
Most people working in this space treat it as an ongoing race rather than something with a clean finish line. Detection can close the gap temporarily, but each improvement tends to become training data the other side learns from.
If detection can't be trusted, how do I verify something is real?
Look for content carrying an independent, publicly checkable authenticity record from when it was created - rather than relying on how convincing it looks or trusting a probability score from a detection tool.
Is there a free way to start verifying my own content?
Yes - Vastavik offers free content verification and free blockchain-backed registration for your own content, up to 5GB.
The bottom line
Detection was probably never going to win this race outright - not for lack of trying, but because it's stuck playing defense against an opponent that only needs one good move to get through. A smarter detector isn't really the fix. Giving real content a way to prove itself, one that doesn't depend on staying ahead of whatever comes next, is.
Try it free at vastavik.org — [ Protect . Verify . Build Trust]