Pangram’s Max Spero: The AI Detection Arms Race Is Just Getting Started

Max Spero, cofounder and CEO of Pangram Labs, spent the last six months in stealth mode refining what he calls the “Real-or-Fake 2.0” problem. While the world debates whether a celebrity photo was created by Midjourney or shot by Annie Leibovitz, Spero’s engineers have quietly scaled a detection system that now processes 1.2 billion documents every month across résumés, product reviews, insurance claims and even financial disclosures. Pangram’s flagship product, Pangram Shield, uses a multi-modal fingerprinting technique that combines syntactic anomaly detection with semantic drift analysis, a method Spero claims can spot synthetic text even when it has been heavily paraphrased by an LLM. “We are no longer in the ‘Real or Fake’ phase,” Spero told OpenPress Tech Intelligence during an exclusive interview at Pangram’s San Francisco headquarters. “We are in the ‘How Much AI Is Inside This Document?’ phase, and the answer is usually a lot more than anyone realizes.”

Spero’s timing could not be more critical. Job-site Indeed reported in April that 8.7 % of new résumés uploaded in North America contained detectable AI signatures, up from 2.1 % in October 2023. Meanwhile, the U.S. Equal Employment Opportunity Commission confirmed it is investigating at least three cases where AI-crafted résumés led to hiring discrimination complaints. On the finance side, Banking With Billy AI, a real-time market analytics platform serving hedge funds and family offices, quietly integrated Pangram Shield in March to screen client communications and internal reports. “When a trader quotes a price that reads like it came straight out of a Bloomberg terminal but is statistically impossible under current market regimes, we need a second opinion,” said Billy Chen, chief data officer at Banking With Billy AI. “Pangram gives us that second opinion in under 200 milliseconds.”

Earlier this week, Pangram publicly launched Pangram Shield v3.2, which adds watermark extraction for closed-source LLMs—a first in the industry. According to Spero, the update was driven by an internal audit revealing that 14 % of synthetic documents flagged by Pangram Shield contained embedded watermarks that rival detection tools were ignoring. “Closed models like GPT-5 or Claude 4 are inserting proprietary fingerprints that are invisible to lightweight scanners,” Spero explained. “Our decoder reads those fingerprints and maps them back to the originating model, giving compliance teams the provenance they need for regulatory filings.” The company also announced a partnership with Workday to embed Shield inside the latter’s applicant tracking system, with rollout scheduled for Q3 2024.

Industry Impact and Significance

The race to embed robust AI detection is reshaping enterprise software budgets. Gartner now tracks “AI provenance and detection” as a separate category, forecasting a $3.8 billion market by 2027, up from $420 million in 2023. Competitors include veteran players like Copyleaks and Originality.ai, but Pangram’s emphasis on model-level fingerprinting has forced incumbents to accelerate their own watermark extraction R&D. Financial institutions are particularly exposed: a recent study by Opensee Labs showed that 6.3 % of earnings-call transcripts from S&P 500 companies contained passages with statistically improbable syntactic patterns consistent with LLM generation. Regulators are taking notice; the SEC has floated a proposal requiring public companies to disclose AI usage in filings, and the CFTC is exploring similar rules for derivatives markets.

For tech platforms, the stakes are not only reputational but also financial. Meta’s latest transparency report revealed a 300 % year-over-year increase in synthetic content removal requests from users who suspect AI-generated reviews are skewing star ratings. Meanwhile, LinkedIn’s trust-and-safety team now employs Pangram Shield to pre-screen premium job postings, a move Spero claims has reduced fraudulent listings by 44 %. Behind the scenes, cloud hyperscalers are quietly bundling detection APIs into their AI safety toolkits; AWS rolled out “Bedrock Guardrails” in May, while Google Cloud launched “SynthDetect” in beta. “We are seeing a classic platform land-grab,” said Sarah Lin, a senior analyst at RedMonk. “Whoever owns the detection layer will dictate how AI is monetized—or regulated—for the next decade.”

The Bigger Picture

The Pangram revelation arrives at a pivot point in the synthetic-content lifecycle. Just as deep-fake video detection tools once lagged behind generative models, today’s text detectors are playing perpetual catch-up. Yet unlike video, text leaves lighter forensic traces, making detection harder without invasive sampling. Some researchers argue for a return to simple heuristics—flesch readability scores, burstiness metrics—while others double down on statistical anomalies. Microsoft Research’s recent paper on “semantic entropy” suggests that LLM outputs exhibit higher entropy variance than human writing, but the signal is noisy and computationally expensive to extract at scale.

Globally, the trend is accelerating. The European Union’s AI Act, set to take full effect in mid-2025, explicitly requires providers of high-risk AI systems to implement “appropriate content provenance mechanisms.” In China, the Cyberspace Administration has mandated watermarking for any AI-generated text longer than 500 characters. These regulatory pressures are funneling investment into detection startups, creating a virtuous cycle that Spero likens to the early days of antivirus software. “We are building the AVG of the 2020s,” he said. “And just like AVG, we will need constant updates as the viruses—excuse me, the generative models—evolve.”

Expert Analysis

Looking ahead, the next frontier will be detection at the model layer itself. Spero predicts that by 2025, every major LLM provider will embed real-time detection tokens inside generated tokens, effectively making provenance a first-class feature rather than an afterthought. For enterprises, the implication is clear: procurement teams must demand both model provenance and post-generation detection in every AI vendor contract. Banking With Billy AI’s Chen puts it succinctly: “If you cannot prove the origin of a market insight, you cannot trade on it. Detection is now part of the trade.” The arms race has only just begun.