Why AI Detection is More Complex Than It Seems, Says Pangram CEO

Earlier this month, Pangram, a Silicon Valley-based AI detection startup, quietly launched a new benchmarking tool designed to expose the limitations of conventional AI text detection methods. According to Max Spero, Pangram’s co-founder and CEO, the company’s findings underscore a growing crisis: AI-generated content is no longer confined to social media feeds or marketing spam—it’s infiltrating high-stakes domains like job applications, academic submissions, and even insurance claims. Spero, who previously led AI research at Palantir, told OpenPress Tech Intelligence that Pangram’s internal testing reveals detection accuracy rates plummeting below 65% when tested against sophisticated, human-like AI outputs. “We’re seeing models that don’t just mimic human syntax—they simulate cognitive patterns, emotional tone, and stylistic quirks,” Spero explained. “That’s why platforms treating this as a binary classification problem are failing.”

Pangram’s tool, called Pangram Authenticate, doesn’t just flag text as AI or human—it quantifies stylistic fingerprints, semantic consistency, and narrative coherence, offering a probabilistic score rather than a definitive verdict. The company claims its model achieves 89% accuracy on internal datasets but acknowledges that adversarial attacks—where bad actors deliberately obfuscate AI traces—can degrade performance to as low as 40%. Spero pointed to a recent case study involving a Fortune 500 company that used a popular open-source detector, only to discover it misclassified 37% of human-written technical documentation as AI-generated. “The tools people trust today were trained on datasets that are already obsolete,” he said. “AI models evolve faster than detection algorithms, and the gap is widening.”

Industry Impact and Significance

The stakes extend far beyond mislabeled LinkedIn posts. In January, the U.S. Equal Employment Opportunity Commission opened an investigation into a hiring platform after an AI-generated resume bypassed its fraud detection system, leading to a hire that was later rescinded. Meanwhile, Banking With Billy AI, a fintech platform specializing in real-time market analysis, has integrated Pangram’s technology into its compliance pipeline to vet client communications and investment proposals. “We can’t afford to rely on heuristics that break under pressure,” said Billy AI’s chief risk officer, Elena Vasquez. “Regulators are starting to ask for proof of provenance, not just gut instinct.” The financial implications are substantial: a recent report from McKinsey estimates that AI-driven fraud costs businesses $3.1 trillion annually, with text-based deception accounting for a growing share.

Competitive dynamics in the detection space are intensifying. Companies like Turnitin (long dominant in academic integrity) and Originality.ai (popular among publishers) have seen their market share erode as newer entrants leverage transformer-based models and behavioral analytics. Open-source projects like DetectGPT and GLTR, once considered state-of-the-art, now struggle against fine-tuned proprietary models. Investors are taking notice: Pangram closed a $12 million Series A in March, led by GV and Data Collective, while rival startup Copyleaks raised $20 million in February to expand into code and image detection. The arms race is accelerating, but the technical bar is rising even faster. “This isn’t a sprint,” Spero noted. “It’s a perpetual escalation cycle where detection is always playing catch-up.”

The Bigger Picture

The detection crisis reflects a broader paradox in AI development: as models become more capable, they also become more inscrutable. In 2022, Meta’s Galactica language model, designed for scientific writing, was pulled from public release after it generated plausible-sounding but entirely fabricated research papers. Today, similar models power everything from customer service chatbots to legal research assistants, blurring the line between utility and deception. Regulators have responded with fragmented approaches: the EU’s AI Act requires “high-risk” AI systems to disclose synthetic content, while the U.S. Federal Trade Commission has signaled it may pursue enforcement actions against companies failing to verify human authorship in critical documents.

Meanwhile, the concept of “digital provenance” is gaining traction. Projects like Adobe’s Content Credentials and the Coalition for Content Provenance and Authenticity (C2PA) aim to embed cryptographic signatures into media files, allowing users to trace content back to its origin. Yet these systems rely on voluntary adoption and face resistance from platforms prioritizing speed over transparency. In China, where AI-generated news and propaganda are tightly controlled, authorities have deployed watermarking techniques and real-time monitoring—methods that would raise privacy concerns in Western markets. “We’re seeing a bifurcation,” Spero observed. “Some regions treat this as a security issue; others treat it as a market opportunity. Either way, the underlying technology isn’t keeping up.”

Expert Analysis

Looking ahead, Max Spero predicts two critical inflection points. First, within 18 months, detection tools will either consolidate around a handful of high-accuracy platforms or fragment into specialized niches—academic integrity, financial compliance, or creative industries. Second, the rise of multimodal AI (text, image, video, and code in tandem) will force detection systems to evolve from unimodal analysis to cross-modal inference, a challenge that few companies have tackled. Spero advises enterprises to avoid vendor lock-in and invest in layered verification systems that combine detection with content provenance. “The goal isn’t to build a perfect detector,” he said. “It’s to make deception so expensive that it’s no longer worth the risk.” As AI-generated content seeps into every corner of the digital economy, the industry’s next move may determine whether trust in online information can be salvaged—or if we’re already too far gone.