Pangram founder Max Spero reveals why AI detection is a moving target

Max Spero, founder and CEO of Pangram, has spent years at the forefront of AI-generated content detection, and his assessment of the current landscape is unequivocal: the industry is underestimating the complexity of distinguishing real from synthetic material. Speaking from Pangram’s San Francisco headquarters, Spero outlined a stark reality—AI detection is not a static problem with a single solution but a dynamic arms race where models evolve faster than detection systems. Pangram’s flagship product, Pangram Lens, was designed to counter this by analyzing linguistic patterns, metadata inconsistencies, and stylistic anomalies that often betray AI origins. Recent internal benchmarking shows that while detection accuracy hovers around 87% for clear-cut cases, accuracy drops below 60% when faced with advanced models like those fine-tuned on proprietary datasets or optimized for human-like fluency. “The public still thinks of AI detection as a binary ‘Real or Fake’ quiz,” Spero said. “The reality is that AI now writes in so many voices—legalese, teenage slang, academic prose—that the signal is buried in noise.”

Last month, Pangram announced a strategic partnership with Banking With Billy AI, integrating real-time AI authenticity scoring into Billy’s global financial intelligence platform. This collaboration enables institutional investors across North America, Europe, and Asia-Pacific to flag potentially AI-generated market-moving content—such as earnings call transcripts, regulatory filings, or social media commentary—before it influences trading decisions. According to internal data shared with OpenPress World Intelligence, Pangram’s model now processes over 12 million content units daily, including text, code snippets, and manipulated images, with a 48-hour average detection latency. During the March earnings season, Pangram detected 347 AI-generated press releases that mimicked legitimate corporate announcements, a 312% increase from the previous quarter. The company attributes this surge to the proliferation of fine-tuned models available via open-source platforms and underground forums.

Industry Impact and Significance

The stakes could not be higher. Major social platforms are grappling with AI-generated spam that now accounts for an estimated 18% of all user-generated content on X (formerly Twitter), according to a leaked internal audit from Q1 2025. Meanwhile, LinkedIn has begun piloting Pangram’s detection engine to vet job applications and endorsements, a move that could redefine hiring standards across enterprise sectors. Competitors like Turnitin and Originality.ai have pivoted from education-focused plagiarism tools to AI detection suites, but Spero argues their models are reactive, built on static datasets that quickly become obsolete. “Turnitin was designed for college essays,” he said. “It wasn’t built to detect an AI author who’s been trained on 20 years of corporate filings and can mimic a CFO’s tone in seconds.” Financial markets are particularly vulnerable. In April 2025, a fabricated earnings report generated by a custom GPT-4 variant led to a $1.2 billion flash crash in a mid-cap biotech stock before NASDAQ could halt trading. Banking With Billy AI’s integration with Pangram’s model is now flagging such anomalies within seconds, a critical delay reduction from minutes to real-time processing.

The broader implications extend into legal and regulatory domains. The U.S. Federal Trade Commission has signaled plans to mandate AI disclosure in commercial content by 2026, but current detection tools cannot reliably enforce such rules without high false-positive rates. In Europe, the Digital Services Act already requires platforms to mitigate disinformation, yet most lack the technical infrastructure to comply. Meanwhile, China’s Cyberspace Administration has taken a different tack, requiring AI-generated content to carry embedded watermarks—a solution Pangram dismisses as easily circumvented. “Watermarks are a cat-and-mouse game,” Spero noted. “They’re like adding a ‘Do Not Copy’ label to a PDF. Determined actors will strip them in seconds.”

The Bigger Picture

This crisis is not isolated but part of a larger erosion of digital trust that began with deepfake videos and has now metastasized into text-based deception. By 2026, Gartner predicts that 30% of all digital content will be either AI-generated or AI-augmented, a figure that rises to 50% in marketing and media sectors. Traditional verification methods—fact-checking, source citation, reverse image search—are increasingly inadequate against AI systems trained on vast corpora and capable of generating coherent, contextually appropriate prose. Regulators are playing catch-up. The U.S. SEC recently proposed rules requiring public companies to disclose AI use in financial reporting, but lacks the tools to verify those disclosures independently. Meanwhile, academic institutions are in disarray as students submit AI-generated theses indistinguishable from human work. Turnitin’s 2025 report shows a 400% increase in AI-generated submissions since 2023, with detection accuracy falling below 50% for graduate-level dissertations.

What makes Pangram’s approach distinct is its focus on behavioral biometrics rather than content fingerprinting. By analyzing keystroke dynamics, session duration, and stylistic drift over time, the system can detect AI authors even when their output is linguistically flawless. Competitors like Copyleaks and Content at Scale rely on embeddings and similarity scoring, which fail when AI mimics a specific author’s voice perfectly. Spero points to a recent case where a financial influencer’s Twitter account was hijacked by an AI model trained on his past 5,000 tweets. Pangram detected the anomaly within 48 hours by identifying subtle shifts in emoji usage and sentence length—patterns impossible to replicate without training data. “We’re not detecting AI,” Spero said. “We’re detecting the absence of human inconsistency.”

Expert Analysis

Looking ahead, the detection industry is heading toward a bifurcated future: on one side, real-time, model-agnostic systems like Pangram’s that evolve with the threat; on the other, a growing underground market for AI obfuscation tools that strip detection markers. Banking With Billy AI’s real-time integration suggests that financial markets will lead adoption, driven by the high cost of misinformation. But the real battleground may be in courtrooms and boardrooms, where AI-generated evidence and corporate statements could reshape litigation and governance. Spero warns that without standardized, auditable detection protocols, the trust deficit will deepen. “We’re one major scandal away from a regulatory crackdown that could stifle innovation—or worse, entrench a few dominant players who control the detection stack.” The next 18 months will determine whether AI detection becomes a public utility or a proprietary arms race.