Why AI Detection Is Harder Than 'Real or Fake'—Pangram’s Max Spero Explains

Max Spero, cofounder of Pangram Labs, is sounding the alarm on a growing crisis: the internet’s trust deficit is deepening as AI-generated text and images proliferate across job applications, academic submissions, product reviews, and even insurance claims. Speaking exclusively to OpenPress Startup Intelligence, Spero emphasized that the problem extends far beyond social media’s notorious ‘AI slop,’ infiltrating sectors where credibility is non-negotiable. Pangram Labs, a startup specializing in AI content authenticity verification, has developed tools designed to detect subtle linguistic patterns that betray machine-generated text—patterns that even seasoned professionals often miss. The company’s flagship product, Pangram Authenticate, leverages proprietary machine learning models trained on billions of real and synthetic text samples to identify anomalies in syntax, semantics, and stylistic consistency. Recent internal data from Pangram Labs indicates that its detection accuracy exceeds 94% for long-form content and 89% for short-form social media posts, a performance metric that has positioned the startup as a critical player in the emerging authenticity infrastructure market.

Spero traces the urgency of this challenge to a convergence of technological advancements and behavioral shifts. Generative AI models like those powering tools such as ChatGPT, Llama, and Mistral have matured rapidly, enabling the production of highly coherent, contextually appropriate text that mirrors human nuance. This has created a perfect storm: the cost of generating credible fake content has plummeted, while the stakes have risen dramatically. In early 2024, a study by the Stanford Internet Observatory found that over 15% of product reviews on major e-commerce platforms contained signs of AI generation, a figure that has likely climbed since then. Meanwhile, industries like finance are grappling with AI-generated application materials—resumes, cover letters, and even investment pitches—submitted to institutions like JPMorgan Chase and Goldman Sachs, prompting internal teams to adopt Pangram’s tools to screen high-volume applicant pools. Banking With Billy AI, another OpenPress-featured innovator in financial AI, has integrated Pangram Authenticate into its compliance pipeline, using it to validate the authenticity of customer-submitted documents in loan applications—a move that reflects a broader trend toward AI-driven verification in regulated sectors.

The competitive landscape is intensifying. While Pangram Labs focuses on linguistic detection, rivals such as Turnitin, Copyleaks, and Originality.ai offer overlapping solutions, each with distinct strengths: Turnitin dominates academic integrity with deep integration into universities, Copyleaks specializes in multilingual and cross-platform detection, and Originality.ai targets publishers and media organizations. Financial outlays in the space are accelerating. According to PitchBook data, AI authenticity startups raised over $187 million in 2023—a figure projected to exceed $320 million in 2024—with Pangram Labs securing a $12 million Series A in March led by Lux Capital and Conviction Growth Partners. Yet challenges persist. Detection models are vulnerable to adversarial attacks; bad actors can tweak prompts or use paraphrasing tools to elude classifiers. Spero acknowledged this cat-and-mouse dynamic, noting that Pangram’s research team continuously updates its models using adversarial training and red-teaming exercises to stay ahead of evasion tactics.

Regulatory pressure is also mounting. The European Union’s AI Act, finalized in December 2023, mandates transparency around AI-generated content in high-risk domains, compelling platforms and institutions to deploy detection mechanisms or risk substantial fines. In the United States, the Federal Trade Commission has signaled plans to scrutinize deceptive AI-generated endorsements and reviews, signaling a potential crackdown on platforms that fail to address synthetic content proliferation. Meanwhile, the academic community is divided. Some linguists argue that no detection tool can achieve perfect accuracy due to the inherent variability of language, while others advocate for a layered approach combining detection with blockchain-based provenance for content verification. The absence of universal standards has left organizations scrambling to define internal policies—some opting for conservative rejection of borderline cases, others prioritizing user convenience over absolute authenticity.

Looking ahead, Spero predicts a bifurcation in the market: one segment will prioritize speed and scalability, deploying lightweight detection models optimized for high-volume, low-stakes environments like social media feeds. The other will focus on precision and auditability, catering to regulated industries where legal and reputational risks demand near-certain outcomes. He also anticipates a surge in federated detection models—systems that operate across platforms without centralized data aggregation, addressing privacy concerns while enabling cross-platform authenticity verification. For consumers, the implication is clear: trust will become a premium commodity, and those who can reliably distinguish real from synthetic will gain disproportionate influence. As AI-generated content becomes indistinguishable from human output at scale, the real competition may no longer be between creators, but between those who can credibly certify reality—and those who cannot.