DooDooLamb News
The only way to fight deepfakes is by making deepfakes
Brief published April 18, 2026 ยท Original source published April 16, 2026
Original reporting by Gaby Del Valle at theverge.com.
Automated brief. Verify important details at the original source.
AI Detection Companies Finally Discover the Only Business Model Worse Than NFTs
In a stunning display of circular logic that would make Ouroboros jealous, a cottage industry of deepfake detection startups has emerged with a foolproof business plan: create the very problem they're being paid to solve, then charge premium rates to detect it. It's like hiring an arsonist to run the fire department, except the arsonist has Series A funding and a really compelling pitch deck.
The premise is beautifully simple in its absurdity. These companies train AI models to generate increasingly sophisticated deepfakes, then use those same techniques to build detection systems. They're essentially playing both sides of an arms race they started, which would be unethical if it weren't so perfectly emblematic of the modern tech economy. Why solve problems when you can monetize the endless cycle of creating and solving the same problem forever?
The Verge's Gaby Del Valle dove into this rabbit hole by testing her own synthetic voice on her unsuspecting parents, because nothing says "quality journalism" like emotionally manipulating your family with AI-generated audio for content. The experiment worked exactly as expected: her father couldn't immediately tell the difference between his daughter's real voice and a digital simulacrum trained on enough audio samples to fool a parent's ear. This is either deeply concerning for the future of human communication or the most elaborate way anyone has ever avoided calling their parents back.
The detection companies positioning themselves as the heroes in this narrative have a slight credibility problem, namely that they're simultaneously the villains. They've built business models that depend on deepfakes becoming more prevalent and sophisticated, which is convenient since they're also the ones making them more prevalent and sophisticated. It's the ultimate hedge: if deepfakes become a massive problem, their detection services become invaluable. If deepfakes remain niche, well, they'll just have to work harder to make them mainstream. Market forces at their finest.
The technical approach these companies employ is called "adversarial training," which sounds like something you'd find in a corporate team-building retreat from hell. The concept involves two neural networks locked in eternal combat: one generates fake content while the other tries to detect it. Every time the detector gets better at spotting fakes, the generator adapts to fool it. Every time the generator creates more convincing fakes, the detector evolves to catch them. It's like watching two AIs play an infinite game of cat and mouse, except the mouse keeps getting better at looking like cheese and the cat keeps getting better at recognizing fake cheese, and somehow this process is supposed to make everyone safer.
This adversarial dance has produced detection systems that can identify subtle artifacts in synthetic media, things like unnatural eye movements, inconsistent lighting, or the digital equivalent of uncanny valley phenomena. The detectors look for telltale signs that human perception might miss: microsecond timing inconsistencies in speech patterns, pixel-level anomalies in facial reconstruction, metadata fingerprints left by generation algorithms. It's forensics for the deepfake era, except the criminals and the crime lab are run by the same parent company.
The underlying technology relies heavily on what researchers euphemistically call "generative adversarial networks" (GANs), which is academic speak for "we taught two AIs to lie to each other until they got really good at it." One network generates increasingly convincing fakes while its adversary develops increasingly sophisticated detection methods. The result is an endless feedback loop of improvement that benefits everyone involved in the detection business and absolutely no one else.
Meanwhile, the actual use cases for this technology paint a picture of a society that has collectively decided the solution to AI-generated deception is more AI. Companies market these detection tools to newsrooms trying to verify sources, social media platforms attempting to moderate content, and law enforcement agencies investigating fraud. The sales pitch writes itself: "In a world where anyone can fake anything, trust us to tell you what's real." The irony that these same companies could generate the fakes they're being paid to detect apparently gets lost somewhere between the venture capital pitch and the product roadmap.
The detection accuracy rates these companies tout would be impressive if they weren't essentially grading their own homework. They claim success rates in the high 90th percentile, which sounds reassuring until you realize they're primarily testing against their own generated content. It's like a locksmith bragging about how quickly they can pick locks they designed, then selling you both the lock and the lock-picking service. The whole ecosystem has the feel of a protection racket with better marketing and fewer broken kneecaps.
Here's what this actually means for anyone who doesn't have a venture capital fund to invest: the technology to create convincing deepfakes is getting democratized faster than the ability to detect them. While detection startups play their expensive cat-and-mouse games, freely available tools are making it easier for anyone with a decent GPU and questionable ethics to generate synthetic media. The detection systems being built primarily serve institutions with budgets, not individuals trying to figure out if their ex is really sending them apology videos or if their grandmother actually endorsed that cryptocurrency.
The ultimate outcome of this arms race isn't better security for everyone; it's a two-tiered system where those who can afford enterprise-grade detection services get protection, and everyone else gets to play Russian roulette with reality. The same market dynamics that made cybersecurity a luxury good are being applied to truth itself.
The beautiful irony is that these companies have created the most honest business model in AI: they've admitted that the solution to artificial intelligence is more artificial intelligence, forever and ever, amen. They've turned the philosophical problem of distinguishing reality from simulation into a subscription service, which is perhaps the most Silicon Valley thing anyone has ever done.