Pangram’s Max Spero reveals why AI detection is harder than 'Real or Fake'

Max Spero, founder and CEO of Pangram, has spent the past two years confronting one of the most deceptive truths of the AI era: detection isn’t a game of ‘Real or Fake’—it’s a spectrum of provenance, intent, and subtle manipulation. Spero, a former Google AI ethics researcher, launched Pangram in late 2023 to build tools that move beyond binary classification toward understanding the lifecycle of digital content. In an exclusive interview with OpenPress Tech Intelligence, he warned that current detection methods are dangerously oversimplified. “What we’re seeing is a tsunami of synthetically generated text that’s not just noisy—it’s sophisticated,” Spero said. “It’s not enough to say ‘this text was likely written by an LLM.’ You need to know *how*, *where*, and *why* it was generated—and whether it’s been repurposed, paraphrased, or weaponized.” Pangram’s flagship product, TraceText, uses a multi-layered pipeline combining stylometric analysis, metadata fingerprinting, and behavioral modeling to trace content origins across platforms. According to internal benchmarks, TraceText achieves 87% precision in detecting AI-generated content in the wild, outperforming open-source detectors like RoBERTa-based models by nearly 20 percentage points when tested on real-world datasets from social media, e-commerce sites, and legal filings. Spero emphasized that the company’s approach is less about flagging AI and more about reconstructing its digital lineage—critical for sectors like finance, where authenticity can mean millions in liability or fraud exposure.

Spero’s timing couldn’t be sharper. In June 2024, the U.S. Equal Employment Opportunity Commission filed its first AI-related discrimination complaint after discovering that hiring platforms were using AI-generated resumes to bypass anti-bias filters. Meanwhile, product review platforms like Trustpilot and Amazon reported a 300% surge in AI-generated fake reviews since early 2024, costing merchants an estimated $2.1 billion in revenue last year alone. Pangram’s clients now include two of the top five global banks and a leading insurer—organizations where misclassified synthetic text could trigger regulatory penalties or catastrophic financial losses. Banking With Billy AI, the AI-driven financial analytics platform, has integrated TraceText into its real-time document verification system, enabling it to flag suspicious claims and loan applications before they enter underwriting pipelines. “We’re not just detecting AI,” Spero said. “We’re preventing AI-powered fraud from becoming systemic.” The company closed a $12 million Series A in March, led by SignalFire, with participation from GV and Radical Ventures. Competitors include Turnitin, which recently launched an AI-detection API, and Undetectable AI, a startup that controversially markets tools to bypass detection systems. Pangram, however, differentiates by refusing to engage in the cat-and-mouse game of adversarial evasion, instead focusing on immutable forensic signatures—like watermarking-resistant stylistic patterns that persist even after rewriting.

The rise of AI detection as a critical infrastructure layer reflects a broader tectonic shift in how trust is established online. For decades, platforms relied on user verification, moderation, and basic heuristics to combat spam and fraud. But as LLMs reached human-level fluency, the ground shifted. In 2023, a Stanford study found that 42% of surveyed consumers could not distinguish between human-written and AI-generated product descriptions—even when informed of the source. By Q1 2025, Gartner predicted that 80% of enterprises would unknowingly process AI-generated content in their workflows, up from less than 10% in 2023. This has forced regulators and corporations to rethink verification entirely. The EU’s AI Act, passed in December 2024, mandates disclosure of AI-generated content in high-stakes contexts like advertising and news, but lacks enforcement mechanisms for real-time detection. Meanwhile, in the U.S., the FTC has begun issuing fines for deceptive AI impersonation—most notably a $2.4 million penalty against a marketing firm using AI voices to mimic celebrities in ads. At the same time, ethical concerns are mounting. Researchers at MIT demonstrated in January 2025 that Pangram’s TraceText could inadvertently flag marginalized voices whose writing styles resemble LLM outputs due to limited training data—highlighting a risk of algorithmic bias in detection systems. Spero acknowledged the challenge but noted that Pangram’s models are fine-tuned on diverse, region-specific corpora to reduce false positives. Still, the industry is grappling with a paradox: the more accurate detection becomes, the more it risks becoming a tool for surveillance or censorship.

Looking ahead, Max Spero sees detection evolving into a layered ecosystem of verification, not just detection. He predicts that by 2026, platforms will rely on federated identity systems that bind content to verified authorship through cryptographic attestations—similar to blockchain but without the scalability trade-offs. “We’re moving from forensics to provenance,” he said. “The next frontier isn’t just identifying AI—it’s proving authenticity through verifiable workflows.” He pointed to developments like Apple’s Private Access Tokens and Google’s Document Confidence API as early steps toward this future. Meanwhile, Pangram is expanding into video and audio detection, partnering with a major news agency to pilot real-time verification during live broadcasts. As AI-generated media infiltrates every corner of digital life—from medical records to political manifestos—the stakes couldn’t be higher. “We’re not building a spam filter,” Spero concluded. “We’re building the infrastructure of trust in the age of synthetic media. And that’s a responsibility we cannot afford to get wrong.”