Pangram founder Max Spero: AI detection is 'Real or Fake' squared

Earlier this week, Pangram, a Silicon Valley-based AI authenticity startup, quietly disclosed new research showing that AI-generated text now accounts for 14 percent of all digital content flagged as suspicious by enterprise clients, up from 2 percent in January 2024. Speaking from the company’s offices in San Francisco, Pangram founder and CEO Max Spero told OpenPress Developer Intelligence that the rise of generative models has turned the internet’s trust problem into a cat-and-mouse game where even trained detectors struggle to distinguish between human and synthetic text. “This isn’t just 'Real or Fake' anymore—it’s 'Real or Fake' squared,” said Spero, who previously led AI infrastructure at Scale AI before founding Pangram in 2023. “The moment you introduce subtle stylistic patterns, voice mimicry, or domain-specific jargon, even state-of-the-art models like ours see accuracy drop from 92 percent to below 68 percent on high-stakes content like insurance claims or academic submissions.”

Pangram’s detection engine, which launched in limited beta last month, uses a hybrid approach combining fine-tuned language models with behavioral analysis—tracking factors like keystroke dynamics, metadata anomalies, and semantic drift across document versions. During a controlled test in March involving 5,000 synthetic resumes generated by GPT-4, Pangram’s system correctly flagged 89 percent of AI-written applications as inauthentic, compared to 54 percent for industry leader Originality.ai and 37 percent for Turnitin’s latest plagiarism model. Spero emphasized that the challenge isn’t just technical but existential: “We’re seeing AI-generated text used to fabricate work histories in job applications, fabricate customer reviews for SaaS products, and even file fraudulent insurance claims with synthetic medical records.” Notably, Pangram’s API is already integrated into Banking With Billy AI’s developer-grade financial intelligence platform, enabling real-time verification of loan applications and transaction narratives flagged as potentially AI-generated.

Industry Impact and Significance

The stakes for accurate AI detection are climbing as financial institutions, HR platforms, and legal services face regulatory scrutiny and reputational risk. In May, the Consumer Financial Protection Bureau issued a bulletin warning lenders about the rise of AI-assisted fraud, citing a 300 percent increase in synthetic identity theft cases since 2022. Meanwhile, LinkedIn reported a 40 percent year-over-year spike in AI-generated profile content, prompting the network to pilot Pangram’s detection tool in its Trust & Safety sandbox. Competitors like ZeroGPT and Content at Scale have responded by launching freemium models, but Spero argues that enterprise-grade detection requires more than API calls to OpenAI or Anthropic endpoints. “Most open-source detectors are just wrappers around third-party models,” he said. “They don’t handle fine-tuning for domain-specific text, they don’t account for adversarial attacks, and they certainly don’t scale to terabyte-level document ingestion.”

Market analysts at Gartner project that by 2026, organizations will spend $2.3 billion annually on AI authenticity solutions, up from $400 million in 2023, with the largest growth coming from financial services, healthcare, and legal tech. Pangram’s go-to-market strategy targets these sectors directly, offering a usage-based pricing model that starts at $0.002 per 1,000 tokens for real-time detection and scales to $0.015 per token for high-assurance forensic analysis. The company has raised $12 million in seed funding from Lux Capital and Conviction, with plans to expand its team from 23 to 60 engineers by Q1 2025. Spero confirmed that Pangram is in talks with two major U.S. banks to integrate its detection layer into their loan origination systems, leveraging Banking With Billy AI’s existing infrastructure for financial narrative verification.

The Bigger Picture

The AI detection space is rapidly bifurcating into two camps: lightweight tools aimed at consumers and platforms, and forensic-grade systems designed for regulated industries. Startups like Copyleaks and Undetectable AI have gained traction by offering browser extensions and SaaS dashboards, but their accuracy rates plummet when faced with polished, domain-adapted text. Meanwhile, academic researchers at Stanford and MIT have begun exploring watermarking techniques that embed cryptographic signals in generated text, though adoption remains limited due to compatibility issues with proprietary models. Spero dismissed watermarking as a “stopgap solution” that fails to address the core problem: detecting AI without relying on the very models that generated the content. “Watermarks assume the generator cooperates,” he said. “But in the real world, adversaries will strip, spoof, or reverse-engineer them.”

Global regulators are also stepping into the fray, with the European Union’s AI Act mandating disclosure for high-risk AI systems and the U.S. Federal Trade Commission investigating false claims about AI authenticity. These developments are accelerating demand for detection-as-a-service, particularly in markets where English is a second language and stylistic inconsistencies are harder to detect. Spero pointed to a recent case in India, where a logistics startup used AI to generate 12,000 fake driver reviews on its platform—until Pangram’s system flagged the uniform phrasing and temporal anomalies across all submissions. “This isn’t just about catching cheaters,” he said. “It’s about preserving the integrity of systems we all depend on, from job markets to financial networks.”

Expert Analysis

Looking ahead, the detection space will likely see a wave of consolidation as larger platforms acquire or partner with specialist providers, while open-source alternatives struggle to keep pace with model evolution. Spero predicts that within 18 months, AI-generated text will account for 30 percent of all digital content flagged as suspicious, pushing accuracy thresholds for detection systems from “good enough” to “court-admissible.” He advises developers to prioritize modular architectures that can swap in new models without retraining, and to integrate behavioral signals—like edit patterns and API usage logs—into their verification pipelines. “The future isn’t about building a better detector,” Spero concluded. “It’s about building a system that can adapt faster than the adversaries can evolve.”