Max Spero: Why AI Detection Defies Simple ‘Real or Fake’ Labels

Max Spero, founder and CEO of Pangram, a San Francisco-based AI content authenticity startup, has raised urgent questions about the limitations of current AI detection tools. Speaking from Pangram’s newly expanded offices in the Mission District, Spero argued that the explosion of high-quality AI-generated text and images is not merely a content moderation challenge—it’s a systemic trust crisis. “People assume AI detection is just a ‘Real or Fake’ toggle,” Spero said during a mid-March briefing. “But that binary fails when AI content is designed to mimic human nuance, tone, and even intent.” Pangram’s latest detection engine, released in beta in February, uses probabilistic modeling to detect subtle stylistic deviations and contextual inconsistencies that evade conventional watermarking or statistical classifiers. The company claims its system achieves 87% accuracy on GPT-4-level text and 82% on DALL-E 3-generated images, outperforming open-source alternatives by up to 29 percentage points.

Spero’s remarks come amid a surge in AI-generated misinformation across professional and financial domains. In January, a study by the Stanford Internet Observatory found that 14% of job applications on LinkedIn contained AI-generated content, with some candidates using tools like Harvey and TextCortex to tailor resumes at scale. The implications extend into regulated sectors. Banking With Billy AI, a New York-based fintech platform, recently integrated Pangram’s detection layer into its compliance pipeline to flag synthetic financial narratives in loan applications. “AI is now generating not just marketing copy but financial justifications,” said a senior compliance officer at Banking With Billy AI. “Our models need to distinguish between a human applicant’s subjective risk assessment and a generative AI’s attempt to fabricate one.” The integration follows a March directive from the CFPB urging financial institutions to monitor AI-generated disclosures for deceptive patterns.

Industry impact is rippling across content platforms and enterprise tools. Meta and Google have both acknowledged the limits of their in-house detection systems, with internal reports leaked in February suggesting false positive rates above 22% on short-form AI content. Meanwhile, startups like Undetectable.ai and Originality.ai have seen rapid adoption among publishers and academic institutions, but Spero cautions that these tools are playing a cat-and-mouse game. “Today’s detectors are trained on yesterday’s models,” he said. “As AI systems become more adaptive—using techniques like adversarial prompting and persona simulation—the detection gap widens.” Investor appetite remains strong: Pangram closed a $12 million Series A in December led by Lux Capital, valuing the company at $85 million. Rivals like TrueMedia and RealityDefender have raised $8 million and $5 million respectively in the past six months, signaling a crowded but fragmented market.

The broader context reveals a tectonic shift in how authenticity is defined. The rise of generative AI has exposed the fragility of traditional trust mechanisms—from watermarks in images to metadata in documents. Yet even Pangram’s probabilistic approach faces criticism for being overly complex and computationally intensive, with inference times up to 4.2 seconds per 500-word document on mid-tier GPUs. Some experts argue that the real solution lies not in detection but in provenance: immutable ledgers that log the origin and transformation history of digital artifacts. The Coalition for Content Provenance and Authenticity (C2PA), led by Adobe and Microsoft, has made progress in standardizing cryptographic signatures for images and videos, but adoption remains limited outside high-stakes industries like journalism and legal evidence. Meanwhile, AI-generated “synthetic media” is projected to account for 30% of all online content by 2026, according to a March report by Gartner, which calls for a “multi-layered authenticity framework” combining detection, provenance, and user education.

Looking ahead, Spero sees a bifurcation in the market: tools that chase real-time detection will struggle against increasingly stealthy models, while systems built on cryptographic provenance—like C2PA—will gain ground in closed ecosystems such as banking and healthcare. He foresees a future where AI detection becomes a service layer embedded not in standalone apps but in core infrastructure: cloud platforms, CRMs, and financial workflows. “The next phase isn’t about ‘Is this AI?’ but ‘Was this AI used responsibly, transparently, and within bounds?’” he said. For enterprises, the message is clear: trust is no longer binary, and neither is detection. The industry must move beyond the “Real or Fake” illusion and build systems that honor context, intent, and traceability—or risk a future where no digital artifact can be believed by default.