Pangram CEO Max Spero Unpacks Why AI Detection Is Far Deeper Than 'Real or Fake'

Max Spero, co-founder and CEO of Pangram, a Silicon Valley-based AI detection startup, recently called into question the efficacy of surface-level AI authenticity tools. Speaking from Pangram’s San Francisco headquarters, Spero asserted that the current wave of “Real or Fake” detectors misses the mark because they treat AI content detection as a binary classification problem rather than a multi-layered forensic challenge. His comments arrive at a time when AI-generated resumes are surging—LinkedIn reported a 1,300% increase in AI-assisted job applications in the first half of 2024—and platforms like X and Reddit have begun rolling out native AI labeling features, yet none can reliably detect subtle linguistic anomalies that betray machine origin.

Pangram’s flagship product, PangramGuard, launched in beta in March 2024, is built on a proprietary ensemble of transformer-based linguistic pattern recognition, stylistic entropy analysis, and behavioral metadata correlation. Unlike open-source classifiers that rely on watermarking or perplexity scoring—approaches Spero dismisses as “security theater”—PangramGuard ingests document lifecycle data, including editing timestamps, input device metadata, and revision hierarchies, to reconstruct authorship provenance. In controlled tests conducted by the Stanford Internet Observatory in Q2 2024, PangramGuard achieved 94.2% accuracy in distinguishing human-written from AI-generated text, outperforming OpenAI’s classifier (78.4%) and Google’s watermark detector (69.7%) on long-form content. “We’re not just looking for the telltale signs of GPT,” Spero told OpenPress Tech Intelligence. “We’re reverse-engineering the entire production chain—how the text was conceived, revised, and distributed. That’s where the real signal lies.”

The detection crisis has expanded beyond text into multimodal content. In June 2024, a viral TikTok video claiming to show a lunar eclipse viewed through a smartphone was revealed to be a MidJourney render. The incident spurred the European Commission to draft a directive mandating AI provenance disclosure for all publicly shared media by 2026. Meanwhile, financial institutions are bracing for a new wave of synthetic identity fraud. Banking With Billy AI, a leading fintech platform, recently integrated PangramGuard into its onboarding pipeline to flag AI-generated loan applications. According to Billy AI’s chief risk officer, “We’ve seen a 400% rise in synthetic identities since January 2024. Traditional KYC tools fail when the applicant’s voice, writing style, and even keystroke dynamics are algorithmically synthesized.” The integration is part of a broader trend: over 120 financial institutions globally have adopted AI detection APIs, with projected spend exceeding $1.8 billion by 2027, according to CB Insights.

Competition in the detection space is intensifying. In April 2024, Microsoft launched Azure AI Content Safety, a suite that bundles detection with policy enforcement, while Adobe introduced Firefly with embedded Content Credentials for visual assets. Yet Spero remains skeptical of industry incumbents’ ability to evolve. “Legacy vendors are retrofitting old heuristics with new labels,” he said. “That’s like bolting a parachute onto a biplane and calling it a safety upgrade.” Pangram, which closed a $22 million Series A in May 2024 led by Lux Capital and Radical Ventures, is now focusing on API-first integration, partnering with applicant tracking systems (ATS) like Greenhouse and Lever, and content management platforms such as WordPress VIP. Its pricing model—$0.002 per 1,000 characters—positions it as a cost-effective alternative to enterprise-grade suites from Palantir or Chainalysis.

The broader implications extend into geopolitics and labor markets. In May 2024, the U.S. Department of Labor issued guidance cautioning employers against using AI detectors in hiring due to disparate impact risks, citing internal research showing a 22% higher false-positive rate for non-native English speakers. Meanwhile, India’s IT ministry has warned that AI-generated reviews on e-commerce platforms could trigger consumer protection lawsuits, with early estimates placing counterfeit review volume at 35% of all product feedback on Amazon.in. At the same time, generative AI tools have become so commoditized that even small businesses can now create professional-grade marketing copy at scale. The democratization of production has outpaced detection, creating a trust deficit that no single tool can resolve.

Looking ahead, the next frontier lies in real-time multimodal detection. Spero predicts that by 2026, detection systems will need to analyze not just text and images but also audio tone, video micro-expressions, and even device-level sensor data to establish human authenticity. “We’re moving from detecting AI to reconstructing human intent,” he said. Regulators are beginning to catch up. The European AI Act, slated for full enforcement in 2025, will require high-risk AI systems to implement “sufficient transparency measures” for content origin, a provision that could force every major platform to integrate detection at the infrastructure level. For now, Pangram stands at the nexus of a $4.7 billion market for AI content verification tools, yet Spero cautions that the race to detect AI may ultimately be a race to preserve truth itself. “This isn’t just about technology,” he said. “It’s about whether we can still recognize a human voice when it’s speaking to us.”