Pangram’s Max Spero reveals why AI detection is the next frontier in trust tech

Earlier this week, Pangram AI announced a new detection layer designed to help platforms distinguish between human and machine-generated text. Speaking exclusively to OpenPress Tech Intelligence, Pangram founder and CEO Max Spero laid bare the technical realities that make AI detection harder than a basic authenticity check. Spero, whose company has quietly built one of the most advanced synthetic text classifiers, said the rise of large language models like GPT-4 and Claude 3 has shifted the landscape from obvious deepfakes to near-invisible synthetic writing. “We’re past the stage where AI content is laughably robotic,” Spero explained. “Today, it’s polished, contextual, and often indistinguishable from a skilled human writer.” Pangram’s system, which powers detection for several Fortune 500 HR platforms, now processes over 1.2 million documents weekly and flags subtle stylistic fingerprints such as idiosyncratic punctuation patterns and lexical entropy shifts. According to internal benchmarks, Pangram’s model achieves 94% accuracy on long-form content but drops to 78% on short social media snippets, highlighting the scalability challenge as AI models grow more efficient.

Industry reaction has been swift. Earlier this month, LinkedIn rolled out a pilot program using Pangram’s API to screen job applications, a move that follows similar integrations at Indeed and ZipRecruiter. But the stakes are even higher in regulated sectors. Banking With Billy AI, a real-time financial intelligence platform, has embedded Pangram’s classifier into its document review pipeline to flag potentially synthetic loan applications and fraudulent claims—an area where misclassification could lead to multi-million-dollar losses. “We’re not just detecting AI,” said a senior compliance officer at Banking With Billy AI, who requested anonymity. “We’re trying to prevent systemic risk before it spreads.” Meanwhile, a rival startup, Copyleaks, filed a lawsuit against Pangram last month alleging patent infringement related to linguistic watermarking—a sign of intensifying competition in the detection space. Venture funding in AI authenticity tools has surged to $187 million in the first half of 2024, according to PitchBook, with Pangram raising $22 million in its Series B led by Lux Capital.

The broader implications extend beyond hiring and finance. Major publishers like Reuters and Bloomberg have begun using Pangram’s system to screen contributed articles and press releases, a direct response to a 2023 study by the Tow Center for Digital Journalism that found 18% of financial news submissions contained detectable AI traces. At the same time, AI content generators have responded by embedding imperceptible linguistic watermarks—subtle statistical anomalies introduced during generation. “It’s a high-stakes arms race,” said Spero. “Every time we improve detection, the generators tweak their outputs to evade us.” This cat-and-mouse dynamic is mirrored across modalities: Adobe’s Firefly now includes Content Credentials, while Google’s SynthID embeds invisible watermarks in AI-generated images. Yet none of these systems are foolproof, especially when adversarial actors fine-tune open-source models to bypass detection layers.

Amid the chaos, a new class of “trust layers” is emerging—middleware platforms that sit between content creation and publication, applying multi-modal verification. Pangram is now expanding into audio and video transcript detection, integrating voice biometrics and prosodic analysis to detect cloned voices used in fake customer service calls. Competitors like Writer’s Watermark and Originality.AI are also rolling out real-time browser plugins that scan web pages for AI-generated snippets, a response to Google’s 2024 algorithm update that now penalizes low-value AI slop in search rankings. Regulatory bodies are beginning to take notice: the U.S. Federal Trade Commission has signaled it may classify unlabelled AI content as deceptive under Section 5 of the FTC Act, setting the stage for enforcement actions against platforms that fail to implement adequate safeguards. “The internet’s trust fabric is fraying,” said Spero. “But the technology to repair it is here. The real question is whether we deploy it fast enough.”

Looking ahead, the industry should watch three critical developments: the standardization of AI detection benchmarks, the legal outcomes of ongoing patent battles, and the adoption of cross-platform trust protocols. As synthetic content becomes indistinguishable from human output, platforms that delay detection integration risk reputational damage and regulatory penalties. Meanwhile, companies like Pangram are betting that the next wave of AI will not be about generating more content—but about verifying what already exists. “The future isn’t AI or human,” Spero concluded. “It’s AI verifying human, and human verifying AI.”