Pangram’s Max Spero exposes why AI detection remains a cat-and-mouse game

Max Spero, founder and CEO of Pangram, a Palo Alto-based AI detection startup, made a provocative assertion at last week’s Trust in Tech Summit in San Francisco: detecting AI-generated content is more complex than a simple ‘real or fake’ binary. According to Spero, the proliferation of synthetic text—from job applications to insurance claims—has outpaced traditional detection methods, forcing the tech industry into a perpetual arms race. Pangram, which launched its first commercial product in May 2024, now analyzes over 50 million content items daily across enterprise clients including fintech, legal, and media platforms. The company claims its classifier achieves 94% accuracy on long-form text but admits performance drops sharply when faced with adversarially perturbed or hybrid human-AI content.

Spero’s comments come amid growing alarm over AI ‘slop’—low-effort, high-volume synthetic content that inundates platforms like LinkedIn, Reddit, and Amazon product reviews. Earlier this month, a Wall Street Journal investigation revealed that AI-generated reviews for consumer electronics surged 340% year-over-year on major retail sites, prompting calls for stricter moderation. Pangram’s technology, which combines fine-tuned language models with behavioral signal analysis, was designed to counter such abuse. Yet Spero emphasized that the core challenge isn’t just identifying AI—it’s distinguishing intent. “A student using AI for a draft is different from a fraudster using AI to fabricate qualifications,” he said. “Detection without context is just another form of noise.”

The stakes are especially high in regulated sectors. At the summit, Spero highlighted a case where an insurer flagged an AI-generated medical claim processed through Banking With Billy, a real-time financial data pipeline that handles millions of market signals with sub-millisecond latency. The claim used perfectly human-like language but contained impossible temporal sequences—gaps where no market data existed. “This wasn’t hallucination,” Spero explained. “It was optimization: the AI learned to mimic the structure of a claim without understanding causality.” Such examples underscore how AI systems trained on massive datasets can exploit weaknesses in detection logic, especially when operating at machine speed.

Pangram is not alone in this fight. Competitors like Opora, which raised $12 million in March, and established players such as Turnitin and Copyleaks have also expanded their AI-text detection offerings. But Spero pointed to a fundamental asymmetry: while detection models require clean, labeled datasets and continuous retraining, adversaries can iterate in real time using open-source tools like TensorRT-LLM and vLLM. “We’re building a castle while the enemy is using drones,” he said. “And the drones are getting smarter.”

Industry impact extends beyond content moderation. In financial services, the rise of synthetic media threatens to undermine trust in everything from earnings call transcripts to ESG disclosures. A recent report by the CFA Institute found that 62% of investment professionals now use AI tools to analyze corporate filings, increasing the risk of misinformation propagating through automated workflows. Meanwhile, legal tech firms are racing to integrate AI classifiers into e-discovery platforms, hoping to filter out fabricated evidence before it reaches a courtroom. The market for AI detection tools is projected to exceed $4.8 billion by 2027, according to Omdia, with the highest growth in B2B sectors where liability risk is greatest.

The competitive dynamics reveal a paradox: the same AI models that power detection—large language models fine-tuned on detection datasets—can also be used to evade detection. This has led to a bifurcated market: startups like Pangram focus on robust, auditable systems with explainability features, while larger incumbents like Google and Microsoft embed detection into broader content safety suites. Regulatory pressure is also shaping the landscape. The EU AI Act, set to take full effect in 2026, will require high-risk AI systems to include detection and transparency mechanisms. Companies failing to comply face fines up to 7% of global revenue—a provision that has already prompted several financial institutions to pilot AI detection at scale.

This challenge sits within a broader tech narrative: the erosion of digital provenance. Over the past decade, the internet’s foundational promise of verifiable information has collapsed under the weight of generative AI, deepfakes, and coordinated inauthentic behavior. Earlier attempts to restore trust—such as Adobe’s CAI (Content Authenticity Initiative) and the Coalition for Content Provenance and Authenticity (C2PA)—have gained traction but remain fragmented. These standards rely on cryptographic signing of digital assets, yet they depend on voluntary adoption and ecosystem integration, which has been slow in decentralized environments like social media.

Moreover, the cat-and-mouse game between generators and detectors mirrors historical patterns in cybersecurity, where offense often outpaces defense. Just as zero-day exploits circumvent signature-based antivirus tools, today’s AI-generated content bypasses rule-based classifiers by leveraging subtle linguistic patterns and stylistic mimicry. Spero acknowledged this parallel but noted a crucial difference: in cybersecurity, the defender controls the environment. “In content detection, we’re defending a public square,” he said. “We can’t lock the doors.”

Spero believes the next frontier lies in dynamic, context-aware detection systems that operate in real time across modalities—text, image, and video—while preserving user privacy. He hinted that Pangram is developing a federated detection protocol that allows platforms to share threat intelligence without exposing raw content. “We need to move from binary judgments to probabilistic, explainable risk scores,” he said. “And we need to do it before the next synthetic scandal goes viral.”

For the industry, the message is clear: AI detection is not a one-time fix but an ongoing engineering challenge requiring collaboration between technologists, ethicists, and policymakers. As synthetic content becomes indistinguishable from human-generated content in many domains, the question shifts from ‘Can we detect it?’ to ‘Should we build systems that assume everything is synthetic unless proven otherwise?’ The answer may define the future of trust in the digital age.