Pangram’s Max Spero on why AI detection is harder than ‘Real or Fake’

The interview with Pangram’s Max Spero is a useful window into a quiet arms race. The premise is simple: as generative AI becomes cheaper and more fluent, the cost of producing fake text and images approaches zero. The result is a flood of synthetic content in places where authenticity carries real economic weight—job applications, product reviews, insurance claims. The platforms that host this content are now forced to police it, and they are discovering that detection is not a binary game.

Spero’s point is that the problem is not 'real or fake' but 'how much of this is machine-generated, and does it matter?' A resume written with AI assistance is not the same as a fabricated reference. A product review that is AI-drafted but factually accurate is not the same as a fraudulent claim. The nuance is where the difficulty lies, and it is also where the market opportunity sits. Pangram is betting that platforms will pay for granular, probabilistic signals rather than simple yes/no flags.

The timing is not accidental. The same week that a major tech company announces another round of AI features, the trust infrastructure around those features is still embryonic. Detection startups are cropping up because the incumbents—social networks, job boards, insurance portals—have not built the internal tools to handle synthetic content at scale. They are outsourcing the problem, and that is a signal of both desperation and opportunity.

For the remote work angle, the connection is direct but not manufactured. Job applications are increasingly digital and asynchronous, which makes them easier to automate. A hiring manager reviewing a stack of resumes cannot tell which ones were written by a human with ten years of experience and which were generated by a model in ten seconds. The asymmetry is real, and it is pushing employers toward verification tools that go beyond the resume itself. Pangram’s approach—focusing on the statistical fingerprints of AI text—is one answer, but it is not the only one.

The broader lesson is that AI detection is a moving target. Models improve, and detection methods lag. Spero’s interview does not pretend to solve that problem; it simply names it. That honesty is rare in a space where startups often overpromise. The market will reward the companies that can deliver reliable, explainable signals without crying wolf. The rest will be noise in the same feed they are trying to clean up.