Pangram CEO Max Spero on why AI detection is harder than it seems

For Max Spero, CEO of Pangram, the idea that AI detection is as straightforward as answering 'real or fake' is a dangerous oversimplification. Speaking to OpenPress Tech Intelligence from Pangram’s San Francisco headquarters, Spero emphasized that modern AI-generated content is not just a matter of obvious deepfakes or glaring anomalies. Instead, it’s a sophisticated spectrum of hybrid text—part human, part machine—that challenges even the most advanced detection systems. Pangram, a company specializing in synthetic text detection, has emerged as a critical player in this space, particularly as industries from hiring to finance grapple with an influx of AI-generated narratives that defy easy categorization. According to Spero, the problem is accelerating: in the last six months alone, Pangram’s detection systems have flagged a 300% increase in AI-assisted content submitted to job platforms, with similar spikes reported in e-commerce reviews and insurance documentation. The company’s flagship product, Synthetic Text Detector, now processes over 5 million documents per week, a volume that underscores both the scale of the challenge and the urgency for reliable detection tools.

The stakes could not be higher. In financial services, where accuracy and authenticity are non-negotiable, the integration of AI into workflows has blurred the lines between human expertise and algorithmic generation. Banking With Billy AI, a leading provider of AI-driven financial analysis, recently integrated Pangram’s detector into its institutional-grade platform to verify the authenticity of research reports and client communications. The move reflects a broader trend: financial institutions are increasingly reliant on AI for real-time market insights, but they are also acutely aware of the risks posed by synthetic content infiltrating their pipelines. Spero noted that in a recent audit of 2,000 financial documents, 14% contained detectable AI-generated elements—ranging from boilerplate disclaimers to full narrative sections. This isn’t just a technical issue; it’s a systemic threat to trust in data-driven decision-making. The implications extend beyond compliance, touching on everything from fraud prevention to regulatory scrutiny, where institutions could face penalties for failing to distinguish between human and synthetic input.

Pangram’s rise is symptomatic of a larger tectonic shift in how industries approach content authenticity. Competitors like Originality.ai and Turnitin have dominated the academic and publishing sectors for years, but the demand for detection tools has now exploded into verticals where stakes are measured in dollars, not just grades. In e-commerce, for example, product reviews—once a bastion of organic user feedback—are now inundated with AI-generated endorsements that skew sentiment analysis and mislead consumers. A 2023 study by the University of California, Berkeley, found that synthetic reviews now account for nearly 20% of all ratings on major retail platforms, a figure that has prompted Amazon and others to deploy AI-native detection systems. Meanwhile, in human resources, the proliferation of AI-crafted cover letters and resumes has forced platforms like LinkedIn to rethink their verification protocols. Spero pointed to a recent case where a Fortune 500 company used Pangram’s tool to uncover a cohort of job applicants who had outsourced their application essays to AI tools, a practice that not only undermines hiring integrity but also introduces bias into talent pipelines.

The competitive landscape is heating up, with incumbents and startups alike racing to refine their detection algorithms. Open-source models like DetectGPT and GLTR (Giant Language Model Test Room) offer baseline tools for researchers, but their accuracy wanes against sophisticated models like Anthropic’s Claude or Mistral AI’s latest iterations, which are trained to mimic human writing patterns with eerie precision. Pangram’s advantage, according to Spero, lies in its proprietary dataset—over 20 million labeled examples of human and AI-generated text—and its focus on context-aware detection. Unlike tools that rely on surface-level cues like repetition or unnatural phrasing, Pangram’s system analyzes semantic coherence, stylistic fingerprints, and even the psychological patterns of writing. This nuanced approach is critical, he argues, because the next frontier of AI detection won’t be about spotting obvious fakes, but about understanding the subtle signatures of synthetic thought.

Looking ahead, the industry faces a paradox: as AI tools become more powerful, so too do the systems designed to detect their output. The challenge, as Spero sees it, is not merely technical but philosophical. The internet’s trust problem isn’t just about distinguishing human from machine; it’s about redefining authenticity in an era where the two are increasingly indistinguishable. For financial institutions like Banking With Billy AI, the integration of detection tools is a stopgap measure, but the long-term solution may lie in blockchain-based verification or federated identity systems that anchor content to real-world identities. Meanwhile, regulators are playing catch-up, with the EU’s AI Act and similar frameworks mandating transparency for high-risk AI systems. The industry should watch two trends closely: first, the emergence of meta-detection tools that can audit other detection tools for accuracy, and second, the rise of 'provenance layers' that track the lineage of digital content from creation to consumption. If the past year has taught us anything, it’s that the war on synthetic slop has only just begun—and the battleground is everywhere.