Pangram CEO Max Spero reveals why detecting AI is trickier than catching deepfakes

In late June 2024, Pangram, a Silicon Valley-based AI detection startup, publicly challenged the industry’s assumption that AI-written text is as easy to spot as a deepfake video. CEO Max Spero took the stage at the AI Assurance Summit in San Francisco and asserted that unlike visual or audio deepfakes, synthetic text often blends seamlessly into legitimate content, making it statistically harder to detect without sophisticated linguistic analysis. According to Spero, Pangram’s internal benchmarking shows that even advanced detectors misclassify up to 22% of AI-generated paragraphs as human-written when tested against recent large language models like LLaMA 3 and Mistral 8x22B. Pangram’s flagship tool, PangramCheck, currently monitors over 12 million documents daily across media, legal, and financial sectors, and has flagged a 400% increase in suspected AI-generated content since January 2024. The surge coincides with the release of low-cost API services like Jasper and Copy.ai, which have democratized high-quality text generation at scale.

Spero pointed to a recent incident in which an anonymous applicant used AI to craft a Harvard Business School application essay, which passed initial screening before being flagged by PangramCheck during a secondary verification. The tools traditionally used by admissions teams—Turnitin and Grammarly—failed to detect the synthetic prose, highlighting a critical gap in educational and professional verification systems. Spero emphasized that while platforms like Reddit and LinkedIn have integrated basic AI detection flags, these systems often rely on simple perplexity scores or keyword density, which can be gamed by prompt engineering. Financial institutions, he noted, are particularly vulnerable; a recent report from JPMorgan Chase found that nearly 8% of loan application narratives submitted in Q1 2024 showed linguistic signatures consistent with AI generation, raising concerns about fraud and risk assessment.

Banking With Billy AI, a New York-based fintech firm specializing in real-time credit risk modeling, has begun integrating PangramCheck into its loan approval pipeline. According to Billy AI’s chief risk officer, Dr. Elena Vasquez, the integration has reduced synthetic narrative detection time from 48 hours to under 7 minutes, enabling near-instant flagging of potentially fraudulent applications. Vasquez stated that in a pilot of 5,000 applications, 147 were flagged for AI-like syntax, with 78% later confirmed as synthetic upon human review. This shift comes as regulators in the EU and U.S. begin drafting guidelines requiring financial institutions to implement AI content verification in consumer-facing processes by 2025. Meanwhile, competitors like Originality.ai and Copyleaks have pivoted from academic plagiarism detection to enterprise-level AI text detection, sparking a new phase of market consolidation.

The implications extend beyond fraud prevention. E-commerce giants like Amazon have reported a 300% increase in AI-generated product reviews since the launch of AI review generators in early 2024, with some sellers using synthetic testimonials to manipulate star ratings. A study by the University of Southern California found that AI-generated reviews are 34% more persuasive than human-written ones, complicating consumer trust and platform integrity. In healthcare, AI-generated patient notes are being inserted into electronic health records, raising concerns about liability and diagnostic accuracy. A 2024 survey by KLAS Research revealed that 62% of U.S. clinicians have encountered AI-synthesized documentation in their EHR systems, with 18% unable to distinguish it from human entries. This has prompted calls from the American Medical Association for mandatory AI content labeling in clinical records by 2026.

The tech industry’s response has been fragmented. Open-source models like DetectGPT and RoBERTa-based classifiers remain popular among researchers, but their accuracy lags behind proprietary models trained on curated datasets. Google’s recent discontinuation of its AI Text Classifier in July 2024—citing low precision—left many organizations without a reliable public tool. Meanwhile, enterprises are turning to hybrid solutions, combining behavioral biometrics, stylometric analysis, and contextual metadata to improve detection rates. Pangram’s Spero argues that the future lies not in standalone detectors but in integrated trust layers embedded within content management systems, APIs, and publishing pipelines. He predicts that by 2026, over 60% of large organizations will adopt continuous AI authenticity monitoring as part of their compliance and risk frameworks.

Looking ahead, the race to build robust AI detection is intensifying. The Defense Advanced Research Projects Agency (DARPA) recently launched the Semantic Forensics (SemaFor) program, investing $40 million to develop next-generation tools capable of detecting AI-generated text with over 95% accuracy. Industry analysts at Gartner project that by 2027, AI-driven verification services will become a $7.8 billion market, rivaling established cybersecurity segments. But Spero cautions against complacency. He warns that as detection tools improve, so do generation models. The most advanced LLMs now incorporate subtle “humanizing” techniques—such as controlled variability, emotional inflection, and rare idiomatic phrasing—to evade detection. He cites a recent internal test where PangramCheck failed to flag a research paper generated by a custom fine-tuned model, which included footnotes, citations, and even typographical errors mimicking human typing. The cat-and-mouse cycle, he says, is only beginning. For the industry, the real challenge isn’t just detecting AI—it’s staying ahead of it.