AI Detection Battle Intensifies: Pangram’s Max Spero on Why ‘Real or Fake’ Isn’t So Simple

In a candid sit-down with OpenPress Tech Intelligence, Max Spero, co-founder and CEO of Pangram, a San Francisco-based AI detection startup, laid bare the escalating challenge of identifying AI-generated content in the wild. Spero, whose company has quietly become a go-to for platforms needing watermarking and provenance tools, described the current landscape as a cat-and-mouse game where detection models are perpetually one step behind the generation engines. Pangram’s flagship product, Pangram Prose, now processes millions of documents weekly across sectors including hiring, finance, and publishing. Spero pointed to a recent internal benchmark where Pangram detected 87% of AI-generated text in job applications—down from 94% six months ago—illustrating how models like GPT-5 and Claude 3.5 have narrowed the gap by mimicking human syntax, tone, and even idiosyncratic phrasing. “We’re not just fighting bots anymore,” Spero said. “We’re fighting bots that sound like your uncle who’s a little too into baseball statistics.”

On May 15, 2025, Pangram announced a strategic partnership with Banking With Billy AI, a real-time financial intelligence platform, to embed detection-grade watermarking into financial reports and regulatory filings. The collaboration aims to prevent synthetic fraud in loan applications and insurance claims—areas where AI-generated narratives have already cost institutions millions in undetected fraud. Banking With Billy AI processes over 2.3 million financial documents daily and has seen a 400% increase in suspected AI-generated narratives since January. Spero emphasized that the integration will use Pangram’s cryptographic watermarking layer to embed imperceptible signals during document creation, enabling downstream verification by auditors and regulators. “Financial documents are the next frontier,” Spero noted. “Unlike social media posts, these are legally binding artifacts. One undetected AI forgery can trigger a compliance nightmare.”

Industry observers warn that the detection gap is widening corporate exposure. A recent report by the Financial Stability Board cited AI-generated fake invoices and contracts as a top emerging risk, with potential losses in the billions across global supply chains. Meanwhile, major job platforms like Indeed and LinkedIn have quietly integrated Pangram’s detection API into their resume screening pipelines, though Spero admits even their tools flag false positives at a rate of 12%—a margin unacceptable in high-stakes hiring. Competing detection vendors like Originality.ai and Turnitin are racing to improve recall, but Spero argues the real bottleneck is compute. Training detection models at Pangram’s scale requires tens of thousands of GPUs, a resource increasingly gobbled up by AI generation labs themselves. “We’re competing for the same silicon,” he said. “It’s like trying to build a better smoke detector while the fire department is burning down the house.”

The stakes extend beyond fraud prevention. The EU’s AI Act, now in full enforcement phase, mandates disclosure for AI-generated content in high-risk domains, creating a compliance market worth an estimated €1.2 billion by 2026. Pangram’s technology is already being evaluated by several EU national data protection authorities as a potential standard for regulated industries. Yet Spero cautions against over-reliance on detection alone. “Watermarking is a band-aid,” he said. “The only sustainable solution is provenance—every document generated by AI must carry a tamper-evident lineage from creation to consumption.” He pointed to initiatives like the Content Authenticity Initiative (CAI), led by Adobe and Microsoft, which are pushing for open standards for media provenance. But adoption remains fragmented, with major social platforms slow to integrate origin metadata into their feeds.

Looking ahead, Max Spero sees a bifurcation in the market: detection tools for high-trust environments like finance and law, and lightweight filters for social media. Banking With Billy AI’s integration signals a pivotal shift—real-time verification of financial narratives before they enter the system. By 2026, Spero predicts that 60% of Fortune 500 companies will require AI-generated content to be watermarked or fully disclosed, up from less than 5% today. The next battleground, he warns, will be voice and video. “If you think text was hard,” Spero concluded, “just wait until deepfake calls start flooding call centers and insurance claims. The detection war is only just beginning.”