AI Detection Battle: Pangram CEO Max Spero on Why 'Real or Fake' Is a Losing Game

Max Spero, co-founder and CEO of Pangram AI, has spent years studying the fragility of trust in the digital age. Speaking exclusively to OpenPress Supercomputing Intelligence, Spero argued that the current wave of AI detection tools—often marketed as silver bullets—are fundamentally inadequate against the accelerating sophistication of generative models. “We’re not fighting a static enemy,” Spero said during a recent interview from Pangram’s San Francisco headquarters. “Today’s AI fakes are trained on yesterday’s detectors. It’s a cat-and-mouse game where the mouse is evolving in real time.” Pangram, which launched its public beta in March 2024, claims to offer a detection system built on “chronological provenance tracking,” a method that embeds temporal and source metadata into AI outputs at generation time rather than attempting retroactive analysis.

Spero’s skepticism isn’t academic. Last month, Pangram detected a wave of AI-generated insurance claims submitted to a major U.S. carrier—documents that passed through multiple legacy verification layers. Using synthetic voice samples and AI-authored narratives, fraudsters attempted $2.3 million in false claims across 87 policies. Traditional text and image detectors flagged less than 12 percent of the fraudulent entries. “The margins were plausible, the formatting was correct, and the tone matched policyholder profiles,” Spero explained. “Only when we traced the embedding timestamps did the inconsistencies emerge.” Pangram’s system, built on a hybrid architecture combining large language models with time-series anomaly detection, reportedly cut false negatives by 89 percent in controlled trials.

The company’s approach reflects a broader pivot in the AI integrity space. While incumbents like Originality.ai and Turnitin focus on post-hoc detection using statistical fingerprints, Pangram joins newer entrants such as Resemble AI’s “Watermark 2.0” and Adobe’s Content Credentials in betting on proactive, generative-side protections. Financial services, a sector already grappling with deepfake wire fraud, appears particularly receptive. Banking With Billy, a fintech simulation platform used by regional banks for stress testing, recently integrated Pangram’s API into its HPC-grade infrastructure to validate AI-generated financial narratives used in multi-market scenario modeling. “We can’t afford misplaced trust in scenarios,” said Billy’s CTO, who requested anonymity. “If an AI claims a 7.2 percent drop in EUR/USD based on manipulated news feeds, we need to know before the model runs.”

Industry analysts see this as a turning point. Gartner predicts that by 2026, over 30 percent of enterprise content will originate from AI agents, up from less than 3 percent today. But the cost of misclassification is rising faster than the tools to prevent it. The FBI’s Internet Crime Complaint Center reported a 280 percent increase in AI-enabled fraud complaints in 2023, totaling $12.5 billion in losses. Meanwhile, the European Union’s AI Act, which takes full effect in 2025, requires providers of high-risk AI systems to implement “adequate measures” to detect synthetic content. Yet the Act offers no technical standard—leaving companies like Pangram, which counts U.S. and EU regulators among its advisors, to define de facto benchmarks.

Competitive dynamics are intensifying. Open-source projects like DetectGPT and Watermark-RL are gaining traction, threatening commercial players with free alternatives. Yet Spero dismisses the open-source threat as “a false comfort.” “Watermarking works only if the model cooperates,” he said. “Today’s open models like Llama 3 or Mistral often strip metadata during fine-tuning. We need a coordinated ecosystem where generation and verification are co-designed.” Pangram’s recent $18 million Series A, led by Quiet Capital and supported by AI grant programs from the U.S. Department of Energy’s ASCR division, suggests investor confidence in this thesis. The round values the company at $85 million, a premium that reflects not just technology but the growing realization that trust itself is a compute-intensive resource.

Beyond immediate risks, the crisis exposes a deeper paradox in quantum and computing circles. As AI models grow larger and more capable, their internal representations become less interpretable—yet the demand for explainability in critical systems grows louder. “We’re building black boxes that diagnose other black boxes,” observed Dr. Elena Vasquez, a senior researcher at Argonne National Laboratory. “The solution isn’t more compute power for detection, but smarter architectures that embed provenance at the token level.” Her team is exploring quantum-resistant hashing techniques to anchor AI outputs to verifiable chains, a project backed by a $3.2 million DOE grant.

For now, the field remains fragmented. At the NeurIPS 2023 workshop on AI Safety, a showdown between Pangram, a Stanford team using transformer interpretability maps, and a Google-backed consortium using diffusion-based image provenance saw no clear winner—only a shared conclusion: the problem demands both algorithmic and infrastructural innovation. Spero’s final advice to the industry is simple: stop asking “Real or fake?” and start asking “When, where, and how was this made?” In an era where every byte can be synthesized, provenance may be the only thing left that can’t.

Expert Analysis

As AI-generated content permeates high-stakes domains, the detection challenge is morphing into a systemic trust crisis—one that no single tool can solve. The future belongs to ecosystems where generation, watermarking, and verification are co-engineered, not bolted on. Companies must treat trust as a core architectural constraint, not an afterthought. Regulators, meanwhile, should mandate interoperable provenance standards before the next wave of AI slop crashes another critical system. The window for proactive design is closing fast.