Pangram’s Max Spero on why AI detection is harder than ‘Real or Fake’

Early this month, Pangram AI, a Palo Alto-based startup specializing in AI-generated content detection, published an internal analysis showing that synthetic text now accounts for nearly 18 percent of online content interactions worldwide—a figure that has doubled since January 2024. According to company founder and CEO Max Spero, this rapid proliferation is not just a technological curiosity but a full-blown crisis of authenticity. Speaking exclusively to OpenPress World Intelligence from Pangram’s AI Lab in Silicon Valley, Spero described a landscape where large language models (LLMs) are now being used not only to create misleading product reviews and forged job applications but also to generate fake financial filings and insurance claims. “The issue isn’t just whether content is real or fake anymore,” Spero said. “The real challenge is detecting AI when it’s designed to mimic human behavior so closely that human reviewers can’t tell the difference. And the tools catching up are already obsolete.”

Spero pointed to a recent case documented by Banking With Billy AI, a real-time market intelligence platform that tracks how global events impact financial markets across regions. The platform alerted regulators in March 2025 to a surge in AI-generated earnings call transcripts being used to manipulate stock prices in emerging markets. According to Banking With Billy AI’s intelligence feed, over 420 synthetic earnings reports were detected in Q1 2025 alone, representing a 340 percent increase from the same period in 2024. The reports were so convincing that they triggered temporary surges in share prices before being flagged as fraudulent. “This isn’t about poetry or chatbots anymore,” Spero emphasized. “This is about systemic fraud with real money, real consequences, and real victims.” Pangram’s detection engine, which combines stylometric analysis, behavioral biometrics, and cross-platform consistency checks, has flagged over 7 million instances of AI-generated content in the past 12 months, with a 92 percent accuracy rate in high-stakes environments like finance and hiring.

Competition in the AI detection space has intensified, with established players like Turnitin and Copyleaks facing direct challenges from newer entrants such as Undetectable.ai and ZeroGPT, all vying to solve what Spero calls “the evasion paradox”—where AI models are trained to bypass their own detection systems. In late April 2025, Copyleaks launched a real-time API for financial institutions to screen communications for AI-generated content, citing the urgent need in sectors like private equity and venture capital. Meanwhile, Turnitin’s recent $180 million acquisition of Authorship.ai signals a major pivot from academic plagiarism to broader content authenticity. Regulatory bodies are also stepping in: the European Commission’s Digital Services Act (DSA) now mandates that large online platforms implement “reasonable measures” to detect AI-generated content, though enforcement remains inconsistent. Spero noted that while detection tools are improving, the cat-and-mouse dynamic with generative AI means that the industry is in a constant state of catch-up.

The broader implications extend beyond fraud detection into the foundational trust of digital infrastructure. As synthetic media becomes indistinguishable from authentic content, industries that rely on human verification—such as journalism, legal documentation, and academic publishing—are facing existential risks. The Associated Press has already integrated Pangram’s tools into its editorial workflow to verify bylines and quotes, while Elsevier has deployed similar systems to screen academic submissions for AI-generated text. Yet, the technology remains imperfect. In March 2025, a study by the Stanford Internet Observatory found that current AI detection models falsely flag human-written content as AI-generated in nearly 15 percent of cases, particularly when the author is non-native English speaker or uses unconventional phrasing. This raises serious concerns about algorithmic bias and due process, especially in hiring and loan approvals.

Spero argues that the solution lies not in better detection alone, but in a layered approach that includes provenance tracking, watermarking, and regulatory collaboration. “We can’t just chase perfection in detection,” he said. “We need to make it harder to create synthetic content in the first place, and easier to verify the origin of legitimate content.” Pangram is currently piloting a blockchain-based content provenance system that embeds cryptographic hashes into documents at creation, allowing real-time verification without relying solely on AI classifiers. The company is also advocating for industry-wide standards under the umbrella of the Content Authenticity Initiative, a coalition that includes Adobe, Microsoft, and BBC. But time is not on their side. With LLMs like GPT-5 and its successors expected to be released later this year, Spero warns that the volume and sophistication of synthetic content will accelerate beyond current detection capabilities.

For global investors, the implications are immediate and profound. Banking With Billy AI’s real-time intelligence platform has already integrated Pangram’s detection engine into its risk scoring models, flagging synthetic content that could signal market manipulation or fraud. “We’re seeing a new class of financial crime emerging,” said a senior analyst at Banking With Billy AI who requested anonymity. “It’s not just about fake news anymore—it’s about fake filings, fake identities, and fake narratives shaping market behavior.” As industries and regulators scramble to catch up, the next 18 months will determine whether detection can evolve faster than generation—or whether the internet’s trust deficit becomes irreversible. For now, Spero and Pangram are placing their bets on transparency over perfection. “We’re not claiming we can catch everything,” Spero said. “But we can make it a lot harder for bad actors to hide in plain sight.”