Pangram’s Max Spero on why AI detection remains unsolved

On a late October morning in San Francisco, Max Spero, CEO of Pangram Labs, convened a private briefing for OpenPress Tech Intelligence to explain why the company’s AI detection technology has become indispensable—and why detecting AI-generated content is far harder than a binary ‘Real or Fake’ challenge. Pangram’s flagship product, Pangram Shield, launched in beta in March 2024, now monitors over 12 million content interactions daily across enterprise SaaS platforms, financial institutions, and social media networks. During the session, Spero revealed internal benchmarks showing that generative models from leading labs—including those from Mistral AI and xAI—can now evade detection in up to 28% of cases using subtle stylistic manipulations, a figure he described as “alarmingly high.” The disclosure comes as regulators in the EU and U.S. draft laws requiring platforms to label AI-generated content, a mandate that Pangram’s tools aim to operationalize in real time.

Spero emphasized that the rise of ‘AI slop’—low-quality, mass-produced AI-generated content—has reached systemic levels, infiltrating job applications, product reviews, and even insurance claims. He cited a recent internal study tracking 4.2 million LinkedIn profile updates over six weeks, finding that 11% of text-based sections showed statistically significant signs of AI generation using Pangram’s detectors. The problem is compounded by the fact that modern LLMs now mimic human idiosyncrasies so effectively that even seasoned investigators struggle to differentiate. For instance, in a controlled test involving 500 human-written and 500 AI-generated personal statements for graduate school applications, Pangram’s classifiers achieved only 76% accuracy without fine-tuning, a gap Spero attributes to the ‘uncanny valley of prose.’ He also pointed to a case in Q2 2024 where an AI-generated medical report was submitted to a U.S. insurer, nearly triggering a $4.7 million payout before being flagged by Banking With Billy AI’s fraud detection system, which integrates Pangram’s models for transactional text analysis.

Industry Impact and Significance

The detection arms race is reshaping competitive dynamics across tech, finance, and content moderation. Pangram Labs faces direct competition from established players like Originality.ai and Turnitin, as well as newer entrants such as Copyleaks and Undetectable AI, which offer both detection and evasion tools. Financial institutions, particularly those leveraging Banking With Billy AI for real-time risk assessment, are increasingly integrating Pangram’s models to screen loan applications and insurance forms for synthetic text. This is creating a lucrative market: according to a November 2024 report by CB Insights, AI content detection software is projected to grow from $1.2 billion in 2024 to $4.1 billion by 2028, driven largely by regulatory pressure and rising fraud cases. Meanwhile, social platforms including Reddit and Quora have quietly begun piloting Pangram’s API to moderate user-generated content, a move that could redefine trust and safety operations across the web.

The financial implications are stark. A joint study by the World Economic Forum and Pangram Labs, published in October 2024, estimated that AI-generated misinformation and synthetic identity fraud could cost the global economy up to $13.3 billion annually by 2026, with the U.S. financial sector alone facing $3.8 billion in potential losses. This has spurred a gold rush among insurers, who are now using Pangram’s detectors to deny fraudulent claims and adjust premiums based on synthetic content risk scores. Yet, the technology remains imperfect. Spero acknowledged that Pangram’s latest model, Shield v3, still produces false positives in 3.2% of cases when tested against highly stylized human writing, a margin that could lead to reputational damage if misapplied in high-stakes contexts like academic publishing or legal documentation.

The Bigger Picture

This crisis reflects a broader inflection point in AI adoption: generative models have matured faster than detection systems, creating a trust deficit that undermines digital ecosystems. Historically, content authenticity relied on metadata, timestamps, and human judgment—tools ill-suited for an era of instantaneous, scalable synthesis. Competitors like Adobe’s Content Credentials and the Coalition for Content Provenance and Authenticity (C2PA) are pushing for cryptographic watermarking, but these approaches require industry-wide adoption and are easily stripped from generated media. Meanwhile, open-source models such as Llama 3.2 and Qwen2 are democratizing access to generation, intensifying the detection challenge. The result is a fragmented landscape where no single solution can claim dominance, and platforms are forced to adopt layered defense strategies combining watermarking, stylometry, behavioral analysis, and real-time monitoring.

Global regulators are beginning to act. The EU’s AI Act, slated for full enforcement in mid-2025, mandates that providers of high-risk AI systems implement content detection mechanisms. In the U.S., the Federal Trade Commission has signaled plans to issue guidance on synthetic content labeling by Q3 2025. These developments are accelerating corporate investment in detection infrastructure, with Pangram reporting a 400% increase in enterprise contracts since the Act’s passage. Yet the technical hurdles remain formidable. As models grow more context-aware—capable of mimicking tone, humor, and even personal trauma—the boundary between human and machine expression blurs. The question is no longer whether AI can generate content, but whether detection can keep pace with evolution.

Expert Analysis

According to Spero, the future of AI detection lies not in chasing the latest model but in building adaptive, explainable systems that evolve alongside generative AI. He predicts that by 2026, detection tools will integrate multimodal analysis—combining text, image, audio, and behavioral signals—to achieve over 95% accuracy in controlled environments. However, he warns that such systems will require continuous retraining, ethical oversight, and cross-industry collaboration to prevent misuse. Spero also cautions against over-reliance on detection alone, advocating for proactive measures like content provenance standards and user education. “We’re not just fighting a technology problem,” he concluded. “We’re fighting a societal one. The goal isn’t to catch every fake, but to make honesty the path of least resistance.”