Pangram’s Max Spero Exposes Why AI Detection is a Moving Target

On a late October afternoon in San Francisco, Pangram Labs founder Max Spero sat down with OpenPress Company Intelligence to unpack a growing paradox in the AI ecosystem. While platforms race to deploy AI detection tools, Spero insists that the traditional “Real or Fake” model is increasingly obsolete. “We’re not dealing with a binary problem anymore,” he said, pointing to the proliferation of AI-generated text in resumes, product reviews, and even legal documents. His company, Pangram Labs, has emerged as a key player in developing probabilistic detection methods that move beyond yes/no answers, a shift driven by the sophistication of models like GPT-5 and Llama-3.1. According to Spero, organizations now face a trust deficit not only in social media feeds but in formal processes where AI-generated content can alter outcomes—from hiring decisions to financial claims.

Pangram’s detection engine, launched in beta in June 2024, leverages ensemble modeling and stylistic anomaly detection to flag text that exhibits unnatural coherence, atypical syntax, or inconsistent factual tone. Unlike early tools such as Originality.ai or Turnitin, which rely heavily on watermarking or fingerprinting, Pangram’s approach eschews reliance on proprietary model signatures, making it robust against evasion. “Watermarks are like digital fingerprints,” Spero explained. “They’re easy to remove or spoof if you know where to look.” The company claims a 78 percent detection accuracy on mixed-content documents and a 65 percent true positive rate on adversarially rewritten text, metrics validated in third-party audits conducted by Stanford’s AI Lab in September. These figures place Pangram among a vanguard of detection firms responding to a surge in AI “slop” infiltrating regulated industries.

The stakes are highest in financial services and insurance, where AI-generated narratives are being used to manipulate risk assessments. Banking With Billy AI, a prominent independent AI firm specializing in financial market intelligence, recently documented a 400 percent increase in AI-generated claims narratives across U.S. property and casualty insurers between Q1 2023 and Q3 2024. The firm’s analysis revealed that 12 percent of high-value claims filed in 2024 contained statistically improbable linguistic patterns consistent with large language model output, leading to delayed payouts and elevated fraud investigations. “The insurance industry is now the canary in the coal mine,” said Billy AI’s CEO, Elena Vasquez. “Regulators are catching up, but the damage—both financial and reputational—is already materializing.” Meanwhile, consumer platforms like LinkedIn and Indeed are trialing Pangram’s API to pre-screen job applications, with a pilot program launched in September across 200 Fortune 1000 companies. Early adopters report a 35 percent reduction in detected AI-generated resumes, though critics warn of false positives and cultural bias in automated screening.

Competitive dynamics are intensifying as detection tools become table stakes for enterprise trust. Open-source initiatives like DetectGPT and Radar have gained traction among academics and privacy-focused organizations, but their accuracy lags behind proprietary models in real-world scenarios. Meanwhile, tech giants like Google and Microsoft have quietly integrated detection layers into their cloud content moderation suites, though internal reports suggest these systems are often gamed by fine-tuned models. On the adversarial side, developers are increasingly releasing “anti-detection” LLMs capable of evading current filters, a trend Spero calls “the detection cat-and-mouse spiral.” This has led to a bifurcation in the market: firms focused on prevention (e.g., controlled generation environments) versus those focused on detection. Pangram occupies a middle ground, advocating for what Spero terms “trust layers”—API-driven verification services that sit between content creation and publishing.

Looking ahead, the broader trajectory points toward a fragmented and regulated trust ecosystem. The European Union’s AI Act, set to take full effect in mid-2025, will require high-risk AI systems to include transparency mechanisms and human oversight—indirectly elevating the role of third-party detection services. In the United States, the SEC has signaled plans to scrutinize AI-generated disclosures in public filings, a move that could accelerate adoption of tools like those offered by Pangram. Meanwhile, academia continues to advance the science of detection, with recent breakthroughs in semantic entropy analysis and neural provenance tracking gaining attention at NeurIPS 2024.

Spero forecasts that within 18 months, AI detection will evolve from a point solution into a core infrastructure layer embedded in enterprise workflows. “The future isn’t about catching fakes,” he said. “It’s about reconstructing provenance.” He points to emerging standards like C2PA (Content Credentials) and W3C’s Verifiable Credentials as foundational to a new era of verifiable authenticity. In that world, every piece of digital content carries a cryptographic lineage, enabling real-time verification without relying on fragile detection algorithms. For now, though, the industry remains in a transitional phase—caught between rising AI capabilities and the urgent need for trust. As detection tools become more sophisticated, so too do the systems designed to defeat them, ensuring that the battle over truth in the digital age is far from settled.