AI-generated text and images have quietly slipped past the meme feeds and into places that actually matter: job applications, product reviews, even insurance claims. In a new TechCrunch video interview, Pangram CEO Max Spero explains why the industry's obsession with a simple 'real or fake' verdict is setting everyone up to fail, and why that matters for anyone trying to trust what they read online.
The internet has a trust problem, and it's not just about AI slop clogging up social feeds anymore. AI-generated text and images are showing up in job applications, product reviews, and even insurance claims, and that's forcing platforms and everyday users to figure out what's actually real. It's a mess, and according to Pangram CEO Max Spero, most of the tools built to fix it are asking the wrong question entirely.
In a new interview with TechCrunch, Spero lays out why treating AI detection as a simple 'real or fake' toggle doesn't hold up once you look at how people actually use generative tools. Someone might draft an email with AI, then rewrite half of it by hand. A student might use a chatbot to brainstorm an outline, then write the essay themselves. A marketer might run a paragraph through an editor to tighten the tone. None of that fits neatly into a binary box, but most detection products still try to force it there anyway.
A handful of startups have cropped up over the past couple of years chasing this exact problem, racing to build tools that can flag machine-generated content before it slips into places that matter. The stakes are real. Universities are fighting a losing battle against AI-written essays, publishers are getting burned by AI-generated news articles slipping into their pipelines, and companies are quietly discovering that a chunk of the reviews on their own product pages were never written by a human customer at all. Every one of those failures chips away at the basic assumption that what you're reading came from a person.
Spero's pitch is that detection needs to work more like a probability score than a verdict. Instead of stamping a document 'AI' or 'human,' Pangram's approach tries to estimate how much of a given piece of text was likely machine-generated, and where. That distinction matters a lot in practice. A hiring manager sifting through cover letters doesn't necessarily want to auto-reject every applicant who used a grammar tool, they want to know if an entire application was fabricated wholesale. An insurer reviewing a claim doesn't need a binary flag, they need enough signal to decide whether a document deserves a second look.
That framing puts Pangram in a crowded but fast-growing lane alongside other AI-content verification players who've raised money betting that detection will become infrastructure, not a novelty feature. The market logic is straightforward: as generative models get better at mimicking human writing style, the gap between 'obviously AI' and 'obviously human' keeps shrinking, and companies that rely on old-school pattern matching are going to get lapped. Spero's argument, laid out in the interview, is essentially that the industry underestimated how fast that gap would close, and overestimated how useful a simple binary classifier would remain once it did.
There's an obvious tension here too. Detection tools have to keep pace with whatever the latest generation of language models can produce, which means it's an arms race with no real finish line. Every time OpenAI, Google, or another major lab ships a more fluent model, detection startups have to retrain and recalibrate. Spero doesn't shy away from that reality in the conversation, framing it less as a problem to be solved once and more as an ongoing maintenance job, similar to spam filtering or fraud detection, categories that never actually reach 'done.'
What happens next probably depends on how quickly enterprises start treating content provenance as a compliance issue rather than a nice-to-have. If regulators or platforms start requiring disclosure of AI-assisted content, tools like Pangram's move from optional add-on to required infrastructure almost overnight. Until then, expect more of these startups to keep popping up, each betting that the winner won't be whoever shouts 'fake' the loudest, but whoever can explain, with actual nuance, how a piece of content came to exist.
The bigger takeaway here isn't really about one startup's product roadmap, it's about how unprepared most institutions still are for a world where authenticity can't be assumed by default. Whether it's a resume, a product review, or an insurance claim, the question 'did a human actually write this' is becoming a real operational problem rather than a philosophical one. Spero's argument that detection needs nuance instead of a simple stamp is a preview of where this whole category is headed, and readers should expect the tools they interact with daily to start quietly running these checks in the background, whether they notice it or not.
