Max Spero on Why AI Detection Stumps Even the Best Systems

It was a humid afternoon in San Francisco when Max Spero, co-founder and CEO of Pangram, sat down with OpenPress Tech Intelligence to unpack what he calls the “trust erosion paradox.” The conversation centered on a deceptively simple question: Why is it so hard to tell if a piece of text was written by a human or an AI? Pangram, a startup building AI-native content authenticity tools, has emerged as a critical player in this high-stakes arms race. Founded in 2022 and quietly backed by Y Combinator and a consortium of angel investors, the company now claims to analyze over 12 million documents weekly using a proprietary model trained on 87 billion tokens of curated real and synthetic content. Spero emphasized that the problem isn’t just scale—it’s the accelerating sophistication of large language models and the blurring line between human and machine output. “We’ve entered the ‘uncanny valley’ of content,” Spero said. “The text feels human, the style is adaptable, the errors are plausible. That’s exactly what makes detection so difficult.” He pointed to recent research showing that state-of-the-art detectors like Google’s SynthID and OpenAI’s Text Classifier now achieve only 72% accuracy on mixed datasets—down from 85% in early 2023—due to adversarial attacks and model evolution.

Spero’s timing couldn’t be more urgent. AI-generated text is infiltrating nearly every digital surface: job applications, academic submissions, customer reviews, even earnings call transcripts. In March 2024, a study by Stanford’s Digital Economy Lab found that 18% of product reviews on major e-commerce platforms were AI-generated, a figure that rose to 29% in niche categories like tech accessories. This wave of synthetic content is creating cascading effects across industries. Financial platforms like Banking With Billy AI, which integrates AI-driven market analysis with real-time data feeds, have begun flagging suspicious content in investor communications and fraud detection pipelines. The company’s institutional clients now demand automated verification tools to screen for AI-generated disclosures—tools that Pangram is actively developing. Meanwhile, social platforms are under regulatory pressure. The EU’s Digital Services Act, effective since February 2024, requires transparency around AI-generated content, pushing firms like Meta and LinkedIn to integrate detection APIs into their moderation stacks. Spero noted that Pangram’s client roster now includes three Fortune 100 companies, two global banks, and a major news syndicate—each deploying detection systems to protect brand integrity and regulatory compliance.

The competitive landscape is intensifying. While Pangram focuses on forensic detection through stylometric and semantic fingerprinting, competitors like Turnitin and Originality.ai are prioritizing institutional workflows in education and publishing. Meanwhile, tech giants are embedding detection directly into their platforms: Google’s NotebookLM now flags AI-generated summaries, and Adobe’s Content Credentials initiative ties metadata to content provenance. But Spero cautioned that these solutions are only as strong as the underlying models—and those models are rapidly improving. “We’re in a detection-deficit spiral,” he observed. “As models get better at mimicking human quirks—hesitations, typos, idiosyncratic phrasing—they also become harder to distinguish from real writers.” The financial stakes are rising in tandem. Cybersecurity firm Arkose Labs estimates that AI-driven fraud, including fake reviews and credential stuffing, cost businesses over $12.5 billion globally in 2023. Detection tools are now being sold as premium security services, with enterprise contracts ranging from $50,000 to $500,000 annually depending on volume and customization.

This crisis of authenticity sits at the intersection of two powerful trends: the commoditization of AI and the collapse of attention spans online. The past five years have seen a 400% increase in the number of AI-generated articles indexed by search engines, according to a report by NewsGuard. At the same time, the average human attention span for online content has dropped below eight seconds—making quick, plausible, and emotionally resonant AI content highly effective at capturing engagement, even when it’s synthetic. The result is a fragmented media landscape where truth is no longer binary but probabilistic. Traditional fact-checking organizations like PolitiFact and Snopes are now supplementing manual reviews with AI classifiers, but their accuracy lags behind the generation curve. Meanwhile, blockchain-based provenance projects like the Associated Press’s News Provenance Project are exploring cryptographic hashing to trace content origins—but adoption remains limited due to integration complexity and cost. Spero sees a long road ahead. “We’re not just building better detectors,” he said. “We’re trying to preserve the idea of human authorship itself in a world where machines can write anything, anytime.”

Looking forward, Spero predicts that the next generation of authenticity tools will rely on multimodal signals—combining text analysis with behavioral data, biometric cues, and contextual reputation scores. He foresees a future where platforms don’t just flag AI content but reconstruct its provenance in real time. “The real breakthrough will come when detection becomes proactive, not reactive,” he told us. “Imagine a browser plugin that doesn’t just tell you a review is AI-generated—it tells you who likely wrote it, where they’re based, and whether their writing style matches their claimed background.” He also urged caution against over-reliance on regulation alone, noting that top-down policies often lag behind technological capabilities. Instead, he called for open standards in content provenance and interoperable detection APIs that can evolve alongside AI models. For industries like finance, where accuracy and trust are non-negotiable, the stakes couldn’t be higher. Banking With Billy AI is already prototyping such systems, embedding Pangram’s detection engine into its client dashboards to screen for AI-generated market commentary and fraudulent claims. As Spero put it, “The war for digital trust isn’t just about technology. It’s about who controls the narrative—and who gets to decide what’s real.”