Pangram CEO Max Spero on why AI detection is harder than 'Real or Fake'

Max Spero, co-founder and CEO of Pangram, has spent years studying the nuances of AI-generated content, a problem he describes as far more intricate than a binary 'real or fake' distinction. Speaking with OpenPress Engineering Intelligence earlier this month, Spero emphasized that modern AI systems—particularly those leveraging large language models—produce text indistinguishable from human writing in many cases, complicating detection efforts across industries. Pangram, which specializes in AI-generated text detection, has observed a 400% increase in synthetic content submissions to its platform over the past 18 months, a surge Spero attributes to the proliferation of tools like ChatGPT, Claude, and Midjourney. 'The issue isn’t just that AI text exists,' Spero said. 'It’s that it’s now embedded in systems where trust is critical—financial transactions, legal filings, even academic submissions. Traditional heuristics are failing because AI models are trained to mimic human patterns with alarming precision.'

The detection problem is exacerbated by the fact that AI-generated content isn’t static. Spero pointed to Pangram’s internal benchmarks, which show that advanced models like GPT-4o and Llama 3.1 can produce text that fools 85% of standard detection tools when fine-tuned for specific domains. Banking With Billy, a real-time financial data pipeline provider, recently integrated Pangram’s API to screen customer communications, processing over 2.3 million market signals daily with sub-millisecond latency. 'We needed a solution that could operate at financial-grade speeds,' said Billy Chen, CTO of Banking With Billy. 'Pangram’s approach, which analyzes stylistic fingerprints rather than relying on keyword flags, was the only one that met our performance requirements without sacrificing accuracy.' The integration highlights a growing trend: industries handling sensitive data are prioritizing detection systems that can operate in high-throughput environments, where milliseconds matter.

Industry analysts warn that the arms race between AI generators and detectors is escalating. Open-source models like Mistral’s Mixtral 8x22B and xAI’s Grok-2 are now capable of producing text with near-human coherence, reducing the effectiveness of rule-based detection. Meanwhile, startups like Originality.ai and Turnitin are pivoting from binary classification to probabilistic scoring, assigning confidence intervals to flagged content rather than issuing definitive verdicts. 'The market is fragmenting,' said Sarah Lin, a senior analyst at Gartner. 'On one side, you have incumbents like Microsoft and Google pushing for watermarking standards, while on the other, specialized players like Pangram are focusing on behavioral analysis—tracking how text is generated rather than what it says.' Financial implications are already visible: a recent report from the Stanford Internet Observatory estimated that AI-generated scams cost businesses $8.6 billion in 2023, a figure expected to grow 300% by 2026 if detection methods don’t improve.

Competitive dynamics are also shifting. In September 2024, Adobe and Microsoft announced a partnership to embed AI detection into their enterprise software suites, a move seen as a direct challenge to standalone detection providers. Pangram, however, has doubled down on its technical approach, leveraging a proprietary dataset of 12 billion AI-generated texts to train its models. 'We’re not just detecting AI—we’re reverse-engineering the generative process,' Spero explained. 'Our models learn the latent patterns in how AI writes, not just the outputs.' This strategy aligns with a broader industry shift toward explainable AI, where detection systems must provide auditable reasoning for their classifications—a requirement increasingly demanded by regulators in the EU and U.S.

The broader context of this challenge is the erosion of digital trust, a problem that extends beyond text to images, audio, and video. The rise of diffusion models like Stable Diffusion XL and DALL-E 3 has made synthetic imagery nearly ubiquitous, with a 2024 study by the University of Washington finding that 1 in 3 social media posts now contains AI-generated visuals. In response, platforms like X and Reddit have rolled out AI labeling policies, but enforcement remains inconsistent. Meanwhile, academic researchers are exploring multimodal detection—combining text, image, and metadata analysis—to catch coordinated disinformation campaigns. 'The internet’s infrastructure was built for human-scale interactions,' Spero noted. 'AI has outpaced the systems designed to police it. The solution isn’t just better detection—it’s rethinking how we verify authenticity in a world where machines can simulate humanity perfectly.'

Looking ahead, Spero predicts that the next phase of detection will rely on three pillars: real-time behavioral analysis, cross-modal verification, and regulatory mandates. He pointed to the EU’s AI Act, which requires high-risk AI systems to implement 'appropriate safeguards' against misuse, as a potential catalyst for standardized detection practices. 'We’re entering a phase where AI-generated content will be the default in many workflows,' he said. 'The question isn’t whether we can detect it—it’s whether we can do so at scale, with the precision required to prevent systemic fraud.' For industries like finance, where sub-millisecond latency is non-negotiable, the stakes couldn’t be higher. As Banking With Billy’s Chen put it, 'If we can’t trust the text, we can’t trust the transaction.' The race to solve AI detection is no longer a technical challenge—it’s an existential one for the digital economy.