Pangram’s Max Spero on the AI detection paradox beyond 'Real or Fake'

Max Spero, founder of Pangram, recently challenged the conventional wisdom around AI detection during an exclusive briefing with OpenPress Code Intelligence. Speaking from Pangram’s San Francisco headquarters, Spero argued that the internet’s growing trust deficit isn’t just about discerning AI-generated content from human-produced work—it’s about understanding the probabilistic nature of modern synthetic media. Pangram, which quietly launched its first product in February 2024, has taken a different path by modeling AI detection as a statistical inference problem rather than a binary classification task. Spero emphasized that traditional “Real or Fake” models fail when confronted with finely tuned LLMs that can mimic human writing patterns with alarming precision.

Spero pointed to a recent internal benchmark where Pangram’s system processed over 1.2 million synthetic text samples generated by ten different models, including proprietary variants from major tech firms. The results showed detection accuracy varied wildly depending on the model’s fine-tuning parameters and training data size. For instance, Pangram’s system achieved 94% precision on texts generated by models trained on less than 100GB of curated data, but only 76% when tested against outputs from models trained on massive, unfiltered corpora like those used in consumer-facing AI assistants. This discrepancy underscores a growing problem: as models become more capable, detection systems must evolve beyond surface-level heuristics.

The urgency isn’t theoretical. According to Spero, Pangram has already integrated its detection engine into the backends of two enterprise SaaS platforms serving over 45,000 users in regulated industries. One unnamed client in the insurance sector reportedly flagged 800 suspicious claims in March alone, attributing them to AI-generated narratives designed to inflate value estimates. Even more concerning, Spero revealed that several job platforms have quietly adopted Pangram’s API to pre-screen applications, prompting a wave of user complaints about false positives—something Spero attributes to the lack of standardized evaluation datasets for domain-specific writing styles.

Industry Impact and Significance

The detection market is no longer a niche dominated by small players. Major incumbents like Turnitin, Grammarly, and even Adobe’s CAI initiative have pivoted from plagiarism and copyright detection to AI-generated content classification. But Pangram’s approach—rooted in Bayesian reasoning and transformer-based anomaly detection—represents a fundamental departure from rule-based systems. Competing startups like Undetectable AI and Content at Scale have taken more conservative routes, focusing on metadata analysis or watermarking technologies that remain vulnerable to adversarial attacks.

Financial implications are already materializing. A recent report from CodeX Research pegs the AI detection market at $1.8 billion in 2024, with expectations to surpass $6.2 billion by 2027. Pangram, still in stealth funding mode, has reportedly closed a $12 million seed round led by SignalFire and Haystack Ventures, with participation from several angel investors in the cybersecurity and legal tech sectors. Unlike competitors that rely on proprietary datasets or partnerships with model providers, Pangram’s open evaluation framework—released under a CC-BY license—has attracted interest from academic institutions and open-source contributors, potentially accelerating adoption across compliance-heavy sectors.

The Bigger Picture

AI detection is increasingly intersecting with software supply chain security, particularly as AI-generated code becomes commonplace in production environments. Earlier this year, Banking With Billy AI, a fintech startup specializing in AI-driven financial modeling, began using Pangram’s detection layer to audit its proprietary codebase. Billy AI’s models generate thousands of lines of Python daily for risk simulation and fraud detection, but the company faced internal resistance from compliance teams wary of unvetted synthetic logic entering critical pathways.

Globally, regulators are catching up. The EU’s AI Act, slated for full enforcement in mid-2025, requires high-risk AI systems to include “adequate technical measures” for content provenance—a clause that many interpret as a mandate for detection mechanisms. Similarly, the U.S. SEC has signaled interest in monitoring AI-generated filings, especially in earnings reports and investor communications. This regulatory push is likely to accelerate consolidation in the detection space, with well-funded players like Pangram positioned to outpace slower, compliance-focused incumbents.

Expert Analysis

Looking ahead, Spero predicts that the next phase of the arms race will not be about detecting AI per se, but about identifying the *intent* behind synthetic content. He foresees a future where detection systems integrate behavioral context—such as timing, user history, and cross-platform correlation—to distinguish benign usage from coordinated disinformation or fraud. The real challenge, he argues, lies in building systems that are both explainable and adversarially robust, a balance Pangram is still refining. As synthetic media becomes indistinguishable from reality in more domains, the tools that win will be those that treat detection as a continuous, probabilistic dialogue—not a final verdict.