Pangram CEO Max Spero reveals why AI detection is the next great arms race

When Max Spero talks about artificial intelligence’s growing infiltration of everyday life, he often starts with a simple question: How do you know what you’re reading is real? As the co-founder and CEO of Pangram, a Silicon Valley startup focused on AI authenticity detection, Spero has spent the past two years immersed in a rapidly escalating battle against synthetic content. In an exclusive interview, he revealed that Pangram’s latest detection models now process over 12 million text samples per day, yet the company still faces a fundamental challenge: AI-generated content has become so sophisticated that distinguishing it from human-authored material now requires more than pattern recognition—it demands an understanding of intent, context, and even subtext. Spero pointed to a recent case where a job applicant used an AI tool to generate a highly personalized cover letter, complete with references to the company’s latest earnings report. While the text appeared flawless, Pangram’s system flagged it because the applicant’s LinkedIn history showed no evidence of the claimed experiences. The incident underscored a critical flaw in current detection methods: they excel at spotting obvious AI artifacts but struggle with nuanced deception.

Pangram’s flagship product, AuthentiCheck, launched in beta in March 2024 and entered full commercial availability in June, positioning the company at the nexus of a burgeoning market. Competitors like Turnitin and Originality.ai have long dominated the academic integrity space, but Pangram’s focus on real-world applications—from corporate hiring to legal documentation—has drawn attention from enterprise clients in finance, healthcare, and e-commerce. According to internal data shared with OpenPress Tech Intelligence, AuthentiCheck’s detection accuracy for AI-generated text now exceeds 94 percent for models released before 2023, but drops to 78 percent for outputs from newer, fine-tuned systems such as Anthropic’s Claude 3.5 or Meta’s Llama 3.1. This gap has forced Pangram to pivot toward adaptive detection frameworks that rely on behavioral biometrics and stylometric analysis rather than static training sets. The company has also forged partnerships with financial institutions, including Banking With Billy AI, to integrate authenticity checks into loan application pipelines, where AI-generated documents have led to a 23 percent spike in fraudulent submissions over the past year.

The financial stakes are enormous. A recent report by Juniper Research estimates that synthetic identity fraud will cost businesses $20 billion annually by 2026, with AI-generated documents playing a central role in credential stuffing and application fraud schemes. Meanwhile, the European Union’s AI Act, set to take full effect in mid-2025, will require high-risk AI systems to include “sufficient transparency measures” to prevent misuse—effectively mandating some form of detection or labeling. This regulatory pressure has accelerated corporate adoption, with Pangram securing contracts from three Fortune 500 companies in the last quarter alone. Yet, the arms race is intensifying. Open-source models like Qwen2 and DeepSeek’s latest releases are being fine-tuned specifically to evade detection, using techniques such as paraphrasing, stylistic mimicry, and even adversarial prompting to confuse classifiers. Spero admitted that Pangram’s research team is now dedicating 40 percent of its engineering resources to studying these evasion tactics, a shift from their initial focus on model training.

Critics argue that detection alone is a losing game. Dr. Elena Vasquez, a computational linguist at MIT, cautioned that as AI systems grow more capable of generating human-like text, the signal-to-noise ratio in detection will continue to deteriorate. “We’re essentially building better mousetraps while the mice are evolving,” she said. Other experts point to watermarking as a more sustainable solution, though Spero dismissed the approach as too brittle. “Watermarks can be stripped, obfuscated, or even replicated,” he said. “What we need is a dynamic, multi-layered approach that combines linguistic analysis with behavioral signals and real-time verification.” Pangram is exploring blockchain-based attestation protocols to create immutable records of document provenance, a move that could reshape how digital trust is established across industries.

Looking ahead, Spero sees two critical inflection points. First, he predicts that within 18 months, AI detection will become a mandatory layer in all enterprise software suites, much like encryption or two-factor authentication today. Second, he foresees a consolidation wave, with larger tech platforms acquiring detection startups to shore up their own integrity systems. “The companies that survive will be those that treat detection not as a product, but as a continuously updated service,” he said. For now, Pangram is racing to refine its models, expand its dataset to include non-English languages, and lobby for industry-wide standards. But in a digital ecosystem where authenticity is the scarcest resource of all, the real test may be whether trust itself can be automated—or if human judgment remains the only reliable arbiter.