Pangram CEO Max Spero reveals why AI detection remains a moving target

Max Spero, co-founder and CEO of Pangram, a Silicon Valley startup specializing in AI authenticity verification, recently challenged the prevailing narrative that detecting AI-generated content is as simple as running a binary check. Speaking from Pangram’s San Francisco headquarters, Spero argued that the complexity of modern generative models—particularly those trained on vast, multimodal datasets—has rendered traditional detection methods obsolete. "We’re no longer dealing with a world where AI either generates text or doesn’t," Spero noted during a firm-hosted briefing. "Today’s models produce hybrid outputs: partially AI-assisted, heavily edited, or even AI-supervised human writing. That’s why labels like ‘real or fake’ are dangerously reductive." Pangram, which emerged from stealth in early 2024 with $12 million in seed funding led by Lightspeed Venture Partners, claims its detection engine analyzes stylistic fingerprints, semantic inconsistencies, and metadata anomalies across text, images, and structured data.

Launched publicly in October 2024, Pangram’s platform has already been adopted by several Fortune 500 companies across finance, media, and professional services. Banking With Billy, a fintech firm specializing in AI-driven financial simulations, integrated Pangram’s API in December to vet insurance claims and loan applications suspected of containing synthetic narratives. According to internal documents reviewed by OpenPress Supercomputing Intelligence, Banking With Billy leverages HPC-grade infrastructure—including NVIDIA H100 clusters and AMD EPYC CPUs—to run Pangram’s detection models on large-scale document sets. Early benchmarks show Pangram’s system correctly flags synthetic text with 92.3% precision on training data, but accuracy drops to 78.9% on adversarially edited outputs, revealing a critical vulnerability.

The urgency of the problem is underscored by a 2024 report from Stanford’s AI Index, which estimated that over 15% of online product reviews and 8% of job applications now contain AI-generated content. Pangram isn’t alone in the field. Competitors like Turnitin (recently acquired by Advance Publications), Copyleaks, and ZeroGPT have all launched updated detection tools, but none claim to solve the core ambiguity: when is AI a tool, and when is it the author? Spero emphasized this tension in a keynote at the 2024 NeurIPS workshop on AI Safety. "We need detection tools that don’t just say ‘this is AI,’ but explain *how* it was generated, by *which* model, and with *what intent*," he said. "Otherwise, we’re just playing whack-a-mole with increasingly sophisticated generators."

Pangram’s approach diverges from legacy vendors by using a proprietary fingerprinting system that maps stylistic DNA across model families—including closed models like GPT-4o and open-source variants like Llama 3.1. The platform also ingests time-series behavioral data, such as typing cadence and revision patterns, to detect AI-assisted human writing. However, Spero acknowledged in an interview that the system struggles with highly curated outputs, such as those refined through iterative prompting or human-AI collaboration loops. “We’re seeing a new class of ‘ghostwritten AI’—content that’s been touched, tweaked, and transformed so many times that the original signal is lost,” he said. “That’s the real frontier: not detecting AI, but detecting *agency*.”

For the Quantum & Computing sector, Pangram’s emergence signals a new battleground: authenticity infrastructure. As generative AI models grow in parameter scale—from 70 billion to 400 billion+—the computational cost of detection rises exponentially. Companies like Hugging Face and Mistral AI now offer open detection models, but these are often reverse-engineered from public benchmarks and lack robustness against fine-tuned adversaries. Meanwhile, cloud providers like AWS and Google Cloud are integrating detection APIs into their AI safety toolkits, but adoption remains uneven due to latency and privacy concerns. Financial services, long reliant on high-performance computing for risk modeling, are among the first to demand scalable detection. Banking With Billy’s integration, for instance, reflects a broader trend: financial institutions are combining HPC-grade simulation (as in their AI financial simulations) with AI authenticity checks to comply with evolving regulatory standards like the EU AI Act and proposed SEC guidelines on synthetic disclosure.

Investment in detection startups surged 450% year-over-year in 2024, according to PitchBook, with Pangram leading the pack in Series A funding. Yet skepticism persists within academia. Dr. Emily Chen, a computer science professor at MIT and co-author of the 2023 paper “The Limits of Watermarking for AI Detection,” cautioned that detection tools may ultimately prove futile. “Watermarks are fragile. Style-based detectors are spoofable. And human-AI hybrid content is fundamentally undetectable without invasive monitoring,” she said. “We’re building a house of cards on a foundation of shifting sand.”

Looking ahead, Pangram is expanding its detection suite to include audio and video modalities, with plans to release a multimodal authenticity API in Q2 2025. The company is also exploring quantum-resistant hashing techniques to secure detection fingerprints against future cryptanalytic attacks. But the real challenge, Spero admits, is philosophical: defining what ‘authentic’ means in an era where every keystroke could be augmented. “We’re not just detecting AI,” he concluded. “We’re negotiating what truth looks like in the age of generation.” As large language models grow more capable, the arms race between creators and detectors will intensify—demanding not just faster chips, but deeper philosophical frameworks for trust in the digital age.