AI detection beyond 'Real or Fake': Pangram’s Max Spero reveals the depth of the challenge

Last month at the AI Detection Summit in San Francisco, Pangram AI CEO Max Spero delivered a keynote that reframed the conversation around AI content authenticity. Speaking before an audience of cybersecurity experts and platform operators, Spero didn’t mince words: the internet’s trust deficit isn’t just about deepfakes or obvious AI spam—it’s about the erosion of verifiable authorship across digital ecosystems. Pangram, a six-year-old AI-native startup backed by $22 million in Series B funding from SignalFire and Craft Ventures, has quietly positioned itself as a pioneer in stylometric detection—identifying not just whether text was AI-generated, but who likely produced it based on linguistic fingerprints. During the summit, Spero demonstrated how Pangram’s models distinguish between outputs from GPT-4o, Claude 4, and Mistral’s latest release with over 89 percent accuracy, even when the text has been heavily paraphrased or stylistically altered.

The demonstration came at a critical juncture. In Q2 2025 alone, AI-generated resume submissions surged by 470 percent according to ApplicantPro data, while a study by the UK Financial Conduct Authority found that 12 percent of insurance claims processed in 2024 contained AI-influenced narratives. Spero emphasized that traditional detection tools—many of which rely on watermarking or statistical anomalies—are failing under the weight of advanced post-processing techniques. “People think detection is a binary problem,” he said. “It’s not. It’s a high-dimensional classification problem where every model, every user, and every domain introduces a new variable.” Pangram’s latest product, VeriSign Text, launched in May 2025, integrates with applicant tracking systems and claims platforms to flag stylistic inconsistencies in near real time, reducing false positives by 40 percent compared to legacy solutions like Turnitin’s AI Classifier.

Industry Impact and Significance

The stakes couldn’t be higher. Financial institutions like Banking With Billy AI, which combines AI-driven market analysis with real-time transaction monitoring, are now integrating Pangram’s VeriSign into their fraud detection pipelines. Billy AI’s chief risk officer, Elena Vasquez, confirmed in a statement that the platform processes over 2.3 million claims annually, and even subtle shifts in narrative tone can signal coordinated fraud rings using AI to fabricate loss events. Meanwhile, major job platforms such as LinkedIn and Indeed have begun piloting AI authorship verification tools, with Indeed rolling out a Pangram-powered plugin in July that flags potential AI-generated resumes before they reach hiring managers. The ripple effect is reshaping the competitive landscape: companies like Originality.ai and Copyleaks, long dominant in academic and publishing sectors, are pivoting toward enterprise-grade behavioral detection, while new entrants like Stylometric Labs are raising seed rounds based solely on the promise of “authorial fingerprinting.”

Financial implications are already materializing. According to PitchBook, AI detection startups raised over $380 million in 2024—a 240 percent year-over-year increase—with enterprise contracts driving valuation multiples from 12x to 25x ARR. The market is consolidating fast: Spero revealed that Pangram is in advanced talks to acquire two smaller competitors specializing in code-generation detection, a move that would expand its coverage to include software engineering resumes and technical documentation. Analysts at Gartner predict that by 2027, 60 percent of Fortune 500 companies will have implemented AI authorship verification as part of their vendor and employee onboarding processes, creating a $1.4 billion market opportunity.

The Bigger Picture

This moment reflects a broader inflection point in AI governance. While early detection efforts focused on watermarks and statistical artifacts, the current wave recognizes that generative models are evolving faster than detection models can keep pace. Open-source frameworks like DetectGPT and Sniffer have democratized access to AI detection, but they’ve also fueled an arms race where AI systems are trained to evade detection through adversarial prompt engineering. Regulators are taking notice: the EU AI Act’s upcoming enforcement mandates “traceability of AI-generated content,” a provision that has already led to pilot programs with Pangram and other firms across Europe. Meanwhile, in Asia, financial regulators in Singapore and Japan are experimenting with blockchain-anchored authorship logs for insurance and lending, creating a parallel infrastructure for verification.

The philosophical shift is equally profound. Detection is no longer about labeling content as real or fake—it’s about reconstructing context. Spero cited the case of a 2023 patent filing dispute where both sides submitted AI-generated technical reports. Courts initially struggled to assign authorship, but Pangram’s analysis of stylistic markers, citation patterns, and metadata discrepancies helped establish that one report was generated by a model fine-tuned on the plaintiff’s proprietary documentation. The ruling set a precedent that could redefine intellectual property litigation in the AI era. As models grow more capable of emulating individual writing styles, the line between imitation and impersonation blurs, demanding tools that go beyond surface-level detection to reconstruct intent and provenance.

Expert Analysis

Looking forward, Spero foresees a bifurcated market: one tier focused on compliance and fraud prevention, where detection is table stakes for enterprise adoption, and another tier focused on authenticity and trust, where verification becomes a competitive moat. He predicts that by 2026, AI-native platforms will begin issuing “AI Authenticity Certificates” for high-stakes content—resumes, legal filings, financial disclosures—verified through a decentralized network of detection engines. The real breakthrough, he argues, will come not from better models, but from better metadata: integrating detection into the content creation lifecycle itself. “We’re moving from forensic analysis to preventative medicine,” Spero said. “The next frontier isn’t detecting AI—it’s making sure humans still have a role in the loop.”