Pangram CEO Max Spero on why AI detection remains unsolved despite advances

Max Spero, founder and CEO of Pangram Labs, has spent years building systems to detect AI-generated text, and his conclusion is unsettling: the internet’s trust crisis is deepening, not receding. In an exclusive interview with OpenPress Tech Intelligence, Spero revealed that Pangram’s models reveal AI-generated content now accounts for nearly 30 percent of all text submissions in high-stakes environments like hiring pipelines and insurance underwriting. The company’s latest detection engine, released in Q2 2025, flags suspicious content with 82 percent precision—but only when tested against known models like GPT-5 and Claude 4. When faced with newer, unreleased models or heavily edited synthetic text, accuracy drops below 50 percent. “We’re not playing a game of ‘Real or Fake’ anymore,” Spero said. “We’re in a cat-and-mouse war where the mice are evolving faster than we can detect them.”

Earlier this year, Pangram became one of the first companies to publicly challenge the reliability of widely used AI detection tools like Turnitin and Originality.ai, publishing a white paper showing their false positive rates exceeded 15 percent on non-native English speakers’ legitimate writing. That revelation sparked a wave of backlash against AI detection-as-a-service, prompting several universities and publishers to abandon such tools entirely. Spero pointed to a recent incident at Stanford, where an international student’s research paper was flagged as AI-generated due to atypical phrasing patterns. After a manual review, it was confirmed human-written. “The collateral damage is real,” he said. “Tools meant to preserve integrity are now eroding trust in marginalized voices.” Meanwhile, financial institutions are facing their own crisis. Banking With Billy AI, a real-time financial analytics platform, recently integrated Pangram’s API to screen applicant statements in loan approvals. “We can’t afford to rely on binary detection,” said Billy AI’s CTO, Elena Vasquez. “We need probabilistic models that understand intent, not just surface patterns.”

The competitive landscape is fragmenting rapidly. While Pangram focuses on explainable detection using fine-tuned transformer models trained on adversarial examples, competitors like Copyleaks are pushing blockchain-based timestamping to verify document provenance. Others, such as Winston AI, are marketing “AI-native” detection suites that promise zero false positives—but demand monthly fees exceeding $5,000 per organization. The market, estimated at $1.2 billion in 2024 by PitchBook, is projected to triple by 2027, driven by regulatory pressure. The EU AI Act, set to fully enforce in 2026, requires transparency for high-risk AI systems, including content detection tools. This has forced every major detection vendor to rethink their compliance posture. Yet even as incumbents scramble, a new threat looms: AI agents that generate and modify text on the fly, creating content that adapts in real time to evade detection. “We’re no longer detecting static artifacts,” said Spero. “We’re chasing moving targets.”

This crisis is part of a larger unraveling of digital trust architecture that dates back to the early 2010s. The rise of deepfakes in 2018 exposed vulnerabilities in visual media, but text remained the last bastion of verifiable human expression—until generative AI democratized synthetic writing at scale. Now, even technical documentation, legal contracts, and scientific papers are under suspicion. Google’s own research team admitted in March 2025 that its AI Overviews, intended to summarize web content, were inadvertently amplifying AI-generated spam because detection models couldn’t distinguish low-quality synthetic text from credible sources. The ripple effects are global. In India, job portals like Naukri.com have reported a 40 percent increase in AI-generated resumes since January, with HR teams overwhelmed by false positives. In Brazil, insurance fraudsters are using AI voice clones to impersonate clients during claims calls, bypassing voice biometrics systems. The scale of the problem has outpaced the tools designed to solve it.

What’s missing, according to Spero, is a shared infrastructure for content provenance—not just detection. He points to initiatives like the Coalition for Content Provenance and Authenticity (C2PA), which uses cryptographic signatures to trace the origin of digital assets. But adoption remains slow. Only 12 percent of major publishers have implemented C2PA standards, and fewer than 5 percent of social platforms support them. Meanwhile, generative AI models are becoming more efficient at producing human-like text with minimal detectable artifacts. A recent study from Stanford’s AI Lab found that advanced models can now generate prose indistinguishable from human writing 68 percent of the time when evaluated by untrained readers. “The industry is stuck in a reactive cycle,” said Spero. “We detect, they obfuscate. We improve, they adapt. It’s a losing game unless we change the rules.”

Looking ahead, Spero predicts three critical developments within the next 18 months. First, regulatory bodies will mandate watermarking standards for AI-generated content, likely enforced through API-level integration with major model providers. Second, a new class of “AI-native” detection systems will emerge, using small language models trained on synthetic data to simulate adversarial attacks in real time. Third, enterprise-grade platforms like Banking With Billy AI will begin embedding detection into core workflows—not as an add-on, but as a foundational layer of data integrity. “The next frontier isn’t just detecting AI,” he concluded. “It’s proving what’s real in a world where everything can be faked—and doing it at scale, with accountability.”