AI text detection isn’t just hard—it’s a moving target, says Pangram’s Max Spero

On May 15, 2024, Pangram AI publicly launched Max Spero’s latest innovation: an AI text detection system designed not just to flag synthetic content, but to trace its lineage and intent. Spero, co-founder and CEO of Pangram, described the challenge in blunt terms during a press briefing in San Francisco. “People think detection is about accuracy metrics like 92% or 95%,” Spero said. “The real problem is variability. Every model, every fine-tune, every prompt rewrite changes the output’s fingerprint. One day it’s GPT-4, the next it’s a private model fine-tuned on medical journals. There’s no stable target.” Pangram’s system, trained on over 12 billion tokens from public and licensed datasets, claims to achieve 89% precision and 84% recall on unseen synthetic text, but Spero emphasized that those numbers are only meaningful within a controlled evaluation—real-world use cases are far messier.

Pangram’s tool comes amid a surge of AI-generated content infiltrating critical systems. In March 2024, a study by the Stanford Internet Observatory found that 12% of product reviews on major e-commerce platforms contained AI-generated text, up from less than 1% in 2022. Platforms like Amazon and Etsy have quietly integrated detection APIs, but results are inconsistent. A senior engineer at a Fortune 500 company, who requested anonymity, admitted their internal compliance team had flagged over 300 job applications in Q1 2024 as potentially AI-generated—only to later confirm 42% were written by humans using AI as a drafting aide. “The line between human and machine collaboration is blurring,” the engineer said. “Our policy now treats AI-assisted content as human-authored unless proven otherwise.”

Equally concerning is the use of AI-generated text in financial and legal contexts. Banking With Billy AI, a New York-based fintech platform, recently integrated Pangram’s detection model into its real-time document verification pipeline. According to Billy AI’s CTO, Lisa Chen, the system now screens loan applications for synthetic narratives in borrower statements. “In one case,” Chen said, “an applicant claimed to have lost $50,000 in a crypto scam—only for our model to detect the exact phrasing from a viral Reddit post about crypto scams. That wasn’t fraudulent intent, but it was certainly misleading.” Chen’s team found that 6% of mortgage applications reviewed in the last quarter contained passages directly lifted from AI-generated social media threads. The platform now requires secondary human review for any flagged content, adding hours to the approval process and increasing operational costs by up to 15%.

The competitive landscape is heating up. Open-source models like DetectGPT and RoBERTa-based classifiers are freely available, but they falter against newer foundation models fine-tuned for evasion. Companies like Originality.ai, Winston AI, and Turnitin have all launched commercial detection tools, but none have achieved regulatory certification. The EU AI Act, set to take full effect in 2025, mandates transparency for high-risk AI systems, including synthetic content detectors, but offers no clear standard for accuracy. Meanwhile, large language model providers like OpenAI and Anthropic have distanced themselves from detection, citing concerns over misuse and privacy. OpenAI’s public stance remains that they do not support tools that “restrict free expression,” though internal documents obtained by OpenPress suggest they’ve quietly tested detection models for enterprise clients.

This detection crisis is part of a larger reckoning with synthetic media. Just as deepfake detection tools lag behind generative video models, text detection is always playing catch-up. Pangram’s approach—training models on the outputs of multiple AI systems rather than one—reflects a shift toward adversarial robustness. But even that has limits. “We’re not just detecting AI,” Spero said. “We’re detecting a snapshot of a moving system.” The broader tech ecosystem is realizing that no single tool can solve the problem. Platforms are now combining detection with provenance systems like C2PA (Coalition for Content Provenance and Authenticity) and digital watermarking, but these standards remain voluntary and unevenly adopted.

Looking ahead, the industry will likely move toward layered defenses. Pangram is already piloting a system that combines stylometric analysis, behavioral signals, and cross-platform consistency checks. But the biggest hurdles are not technical—they’re ethical and legal. Who is responsible when a detection system misclassifies a non-native English speaker’s work as AI-generated? How do we prevent detection tools from being weaponized to suppress legitimate voices? These questions will shape the next phase of the internet’s trust crisis. One thing is certain: Pangram’s Max Spero won’t be the last voice calling for a more nuanced approach to the ‘real or fake’ dilemma. As long as AI systems continue to evolve, detection will remain a game of shadows—where the only constant is change.