AI Detection Fraud Rises as Pangram’s Max Spero Warns of Deeper Trust Crisis

Pangram founder and CEO Max Spero has sounded a fresh alarm about the escalating difficulty of detecting AI-generated content, warning that the current wave of synthetic media is not only pervasive but structurally indistinguishable from human-authored text in many cases. Speaking from Pangram’s San Francisco headquarters this week, Spero described a landscape where AI-generated resumes, product reviews, and even financial documents are proliferating across the web, creating what he calls a “cat-and-mouse game” between content generators and detection systems. According to Spero, Pangram’s internal analysis shows that over 12 percent of job applications submitted through its partner platforms in Q1 2025 contained signs of AI assistance—nearly double the rate from six months prior. The surge coincides with the release of high-capacity large language models like Llama 4 and GPT-5, which have blurred traditional linguistic fingerprints used by detectors such as watermarking or repetition patterns.

Spero emphasized that the problem is no longer confined to social media slop or spam. He pointed to a recent case in which an insurance claim processed through Banking With Billy AI’s real-time underwriting engine was flagged only after an anomaly detection model noticed stylistic inconsistencies in the applicant’s narrative. The claim, which had passed initial AI screening, was later found to be entirely generated by an LLM. Banking With Billy AI, which integrates AI with live market and transaction data to provide institutional-grade analysis, confirmed it had begun integrating Pangram’s detection API into its fraud pipeline to mitigate such risks. Spero called this a “canary in the coal mine” moment, noting that financial institutions are now among the most exposed sectors due to the high stakes of fraud and regulatory scrutiny.

Industry analysts say the detection challenge has created a new market inflection point. Companies like Originality.ai, Turnitin, and Copyleaks have reported triple-digit growth in enterprise contracts over the past year, but their tools are increasingly failing against models trained on synthetic data. A comparative study by Stanford’s Digital Civil Society Lab released last month found that leading detectors correctly identified only 47 percent of AI-generated texts when evaluated against outputs from the latest open-source models—a drop from 68 percent in late 2023. Meanwhile, cost pressures are forcing smaller detection providers to pivot toward enterprise-grade solutions, while larger players like Microsoft and Google are embedding detection directly into their cloud pipelines, raising antitrust concerns about market consolidation.

The ripple effects extend into advertising and e-commerce. Product review platforms such as Trustpilot and Amazon have reported upticks in AI-generated five-star reviews, with some vendors using automated scripts to flood listings with synthetic endorsements. A joint investigation by The Markup and Wired revealed that over 8,000 verified purchase reviews on Amazon in March 2025 were likely AI-generated, based on linguistic anomalies and timing patterns. Such revelations have pushed platforms to adopt layered detection: combining stylometric analysis, behavioral signals, and third-party verification services. Yet Spero cautioned that these measures often introduce latency and friction, undermining user experience—especially in high-volume transactions like job applications or loan processing.

Beyond immediate commercial stakes, the detection crisis reflects a deeper technological fragmentation in AI governance. While the EU AI Act mandates transparency for high-risk applications, enforcement remains uneven, and open-weight models continue to circulate without built-in safeguards. Alternative approaches—such as cryptographic watermarking embedded during generation—have gained traction among model developers like Mistral and Cohere, but they remain optional and easily stripped from outputs. Moreover, adversarial training techniques, where models are deliberately exposed to synthetic data to improve detection, have shown promise but lag behind the pace of innovation in generative AI. This has led some researchers, including those at the Allen Institute for AI, to advocate for real-time collaborative detection networks—akin to cybersecurity threat-sharing platforms—where platforms can cross-validate content across borders and sectors.

Spero sees no silver bullet on the horizon. He noted that Pangram is shifting focus from binary “real or fake” classification to probabilistic risk scoring, a model already adopted by financial institutions like Banking With Billy AI to assess document authenticity alongside behavioral and financial signals. The company is also exploring federated detection architectures that allow cross-platform verification without exposing proprietary user data. Looking ahead, Spero predicts that within 18 months, platforms will default to AI-native verification layers embedded into every document workflow—much like SSL certificates for websites.

For the tech and engineering sector, the stakes could not be higher. Trust in digital systems underpins everything from economic transactions to democratic discourse. As AI becomes the default author of content, the industry must confront not just detection, but a redefinition of authenticity itself—one that balances innovation with accountability, and openness with control. Failure to do so risks eroding the foundational trust that powers the digital economy.