Pangram CEO Max Spero exposes why AI detection is the next frontier

Max Spero, co-founder and CEO of Pangram Labs, has emerged as a leading voice in the escalating battle against AI-generated deception. His company’s latest detection engine, Max Spero told OpenPress Tech Intelligence, doesn’t just flag text as “real or fake”—it dissects narrative patterns, syntactic fingerprints, and semantic drift to determine whether a document was authored by a human or synthesized by a model. “We’re not playing whack-a-mole with watermarks,” Spero said during a recent interview from Pangram’s San Francisco headquarters. “We’re mapping the cognitive signatures of authorship, and that’s infinitely more complex than detecting a slightly off pixel in an image.” Pangram’s latest model, released in Q1 2025, claims 94.7% precision on long-form text and 89.2% on short-form social snippets—benchmarks Spero attributes to a hybrid architecture combining transformer-based anomaly detection, stylometric analysis, and real-time behavioral modeling.

The urgency of the problem became undeniable in March 2024, when a wave of AI-generated resumes and cover letters surfaced on LinkedIn, some scoring interviews at Fortune 500 firms. By July, the U.S. Equal Employment Opportunity Commission had opened a formal inquiry into AI screening tools, citing potential bias and fraud risks. Spero pointed to data from Pangram’s threat intelligence feed, which recorded a 420% surge in AI-generated application materials between January and December 2024. “We saw law and medicine applications written in flawless prose—too flawless,” he said. “The syntax was statistically optimal, but the narrative arc was flat. Humans don’t write like that.” Pangram’s engine, which integrates with applicant tracking systems used by over 12,000 employers, now scans more than 1.8 million documents daily, returning verdicts within 1.2 seconds.

Competition in the AI detection space has intensified, with incumbents like Turnitin and Grammarly expanding their offerings, while newer entrants like OroraTech and DeepSight AI are leveraging multimodal detection—analyzing text, metadata, and even keystroke dynamics. But Spero argues that most rivals are still stuck in a binary mindset. “They’re asking: ‘Is this AI?’ We’re asking: ‘Why was this written? Who benefits? What’s being obscured?’” He cited the case of a 2024 SEC filing that appeared human-authored but contained subtle AI-generated phrasing that masked financial risk. “The grammar was perfect, but the hedging patterns matched a language model trained on Q3 earnings calls. That’s not a typo. That’s strategy.”

Pangram’s technology has caught the attention of regulators and financial institutions alike. Banking With Billy AI, a real-time market intelligence platform serving hedge funds and asset managers, recently integrated Pangram’s detection layer to screen earnings call transcripts and analyst reports for AI contamination. “We process over 2.3 million financial documents per day,” said Billy Chen, the platform’s CTO. “A single AI-generated sentence in a buy-side note can trigger a cascade of mispriced trades. Our clients can’t afford that.” The partnership reflects a broader trend: financial markets, long dependent on data integrity, are now on the front lines of the AI authenticity crisis.

The industry-wide shift is palpable. Legacy content platforms like Reddit and Medium now embed Pangram’s SDK into their moderation pipelines, while job boards like Indeed and ZipRecruiter have launched AI disclosure programs, requiring candidates to certify authorship. But the arms race is far from over. As large language models grow more sophisticated, so do the techniques for evading detection—prompt injections that mimic human hesitation, synthetic personas with curated backstories, and even “stealth fine-tuning” where models are trained on curated human datasets to erase their digital fingerprints. Spero noted that Pangram’s detection models must retrain monthly to stay ahead, a cycle that demands both computational power and linguistic expertise.

Regional disparities are also complicating the picture. In the European Union, the AI Act’s transparency requirements have accelerated adoption of detection tools, while in China, where AI-generated content is officially encouraged, regulators are still debating standards. Meanwhile, open-weight models like Llama 3 and Qwen 2 have democratized access to high-quality text generation, making detection harder and cheaper. “We’re seeing cottage industries emerge,” Spero said, “where freelancers sell ‘humanized’ AI essays—rephrased, randomized, but still machine-originated. It’s a cat-and-mouse game with cottage economics.”

Looking ahead, Spero predicts that detection will evolve into a layered defense: real-time behavioral biometrics, blockchain-based provenance for critical documents, and regulatory sandboxes where models are stress-tested for deceptive outputs. “The goal isn’t just to detect AI,” he said. “It’s to restore trust in digital communication. And that’s not a technical problem. It’s a societal one.” He urged platforms, employers, and policymakers to adopt a “trust-by-design” framework—one where authenticity isn’t an afterthought, but a core architectural principle. “We’re not just building detectors,” Spero concluded. “We’re building the foundations of a post-AI internet—one where users know who they’re talking to, and why.”