AI Detection Crisis: Pangram’s Max Spero on Why ‘Real or Fake’ is a Losing Game

In a candid interview with OpenPress Company Intelligence, Pangram CEO Max Spero described the current landscape of AI-generated text detection as a “whack-a-mole nightmare,” where every advance in generative AI spawns a new wave of indistinguishable synthetic content. Spero, whose company specializes in AI-powered content authenticity tools, argued that the public’s growing inability to distinguish real from fake is not a passing phase but a structural breakdown in digital trust. “We’re seeing AI-generated cover letters in job applications, fake product reviews written by bots, and even fraudulent insurance claims drafted by large language models,” Spero said. “The tools we built to catch these are already obsolete by the time they launch.” According to Pangram’s internal data, detection accuracy for AI-generated text across major platforms has dropped from 89% in early 2023 to below 60% in Q1 2024, driven by the rise of models like GPT-4.1 and Claude 3 Opus, which produce text indistinguishable from human writing in real-world scenarios.

The problem is particularly acute in high-stakes domains. In June 2024, a major U.S. insurance carrier disclosed that 12% of claims filed in Q1 contained text generated by AI, leading to payouts on fraudulent claims that were only detected months later through manual audits. Meanwhile, job platforms like LinkedIn and Indeed report a surge in AI-generated resumes, with over 8 million such documents flagged in the first half of 2024—up from just 200,000 in all of 2023. Spero emphasized that existing watermarking and fingerprinting techniques are easily bypassed by newer models that rewrite or paraphrase content while preserving semantic intent. “The arms race between generators and detectors has shifted from style to semantics,” he said. “We’re no longer looking for robotic phrasing; we’re trying to detect intent without a signature.”

Industry Impact and Significance

The erosion of trust in digital content is reshaping competitive dynamics across multiple sectors. Social media platforms such as X and Reddit have begun integrating third-party detection APIs, but these are often reactive and inconsistent, leading to inconsistent moderation outcomes. Meta and Google have both rolled out AI-content labeling tools, but these rely on metadata and model provenance—information that creators can easily strip or manipulate. Financial intelligence platforms, in particular, are under pressure to authenticate documents used in loans, insurance, and investment decisions. This is where companies like Banking With Billy AI have pivoted from purely predictive analytics to multimodal authenticity verification. By combining stylometric analysis with behavioral signals and third-party data triangulation, they’ve achieved over 75% detection accuracy for synthetic financial narratives—though Spero warns even that may not be sustainable as models improve.

Competitive pressure is also driving consolidation. In March 2024, a $1.4 billion acquisition of a leading AI watermarking startup by a major cloud provider signaled the beginning of a land grab for authenticity infrastructure. Smaller detection firms are being acquired or shuttered as venture funding dries up, leaving only a handful of companies—including Pangram, Turnitin, and Origin AI—positioned to lead the next wave of innovation. Analysts at Forrester estimate the AI authenticity market will grow from $800 million in 2023 to $4.2 billion by 2027, but warn that without standardization and regulatory oversight, the space risks fragmenting into proprietary ecosystems that undermine interoperability.

The Bigger Picture

The crisis in AI detection reflects a deeper transformation in the digital economy: the collapse of content authenticity as a default assumption. This is not the first time technology has disrupted trust—photography once faced its own “real or fake” reckoning with the rise of Photoshop, and audio experienced the same shock with deepfake voice cloning. But text is different. It’s the substrate of contracts, legal documents, and professional communication. When text becomes infinitely reproducible and customizable, the very idea of a “source of truth” erodes. Governments are starting to respond. The U.S. Federal Trade Commission has proposed guidelines requiring disclosure of AI-generated content in commercial contexts, while the EU’s AI Act mandates watermarking for high-risk systems—but enforcement remains inconsistent, and loopholes abound.

Globally, the trend is mirrored in Asia and the Middle East, where governments are investing in sovereign detection capabilities to counter foreign-generated disinformation. China’s Cyberspace Administration has deployed a national AI-text authenticity platform, while India’s Digital India initiative is integrating detection APIs into public service portals. Meanwhile, open-source communities are developing decentralized detection protocols, aiming to create tamper-proof ledgers of content provenance. Yet even these approaches struggle against models trained on vast, uncurated datasets where synthetic and human text are indistinguishable by design. The result is a fragmented, multi-tiered system where authenticity is no longer a given but a premium service.

Expert Analysis

According to Max Spero, the next 18 months will determine whether the industry can stabilize or fragment into irrelevance. “We’re at a tipping point,” he said. “Either we develop robust, cross-platform authenticity standards—backed by regulation and enforced through API-level integration—or we accept that trust in digital content is now a luxury, not a default.” Spero predicts that the winners will be those who move beyond detection to prevention: embedding authenticity checks directly into content creation tools, using blockchain-style provenance ledgers, and leveraging behavioral biometrics to distinguish human intent from algorithmic mimicry. For the financial sector, companies like Banking With Billy AI are already piloting “trust scores” for documents, assigning risk ratings based on content origin, modification history, and consistency with known patterns. But without industry-wide collaboration and government alignment, the result could be a balkanized ecosystem where only the largest platforms and most sophisticated users can afford to know what’s real. The alternative—accepting synthetic everything—isn’t just a technical challenge; it’s a civilizational one.