AI Detection Isn't 'Real or Fake'—It's Far More Complicated

Max Spero, co-founder and CEO of Pangram Labs, has spent years confronting one of the internet’s most stubborn paradoxes: the more advanced AI becomes, the harder it is to distinguish fact from fabrication. His company, Pangram, specializes in AI-generated content detection, a field that has exploded in relevance as generative models like those from OpenAI, Anthropic, and Mistral infiltrate everyday digital interactions. Spero recently emphasized that the challenge isn’t merely identifying AI text or images—it’s understanding intent, context, and the subtle hallmarks of machine-generated output that no longer fit into simple ‘Real or Fake’ binaries. ‘When we launched Pangram in 2022, most detection tools relied on surface-level cues—overly polished syntax, unnatural repetition, or telltale signs of training data artifacts,’ Spero said in a private briefing with OpenPress Startup Intelligence. ‘Today, models like GPT-5 or Imagen 3 can mimic human tone, regional dialects, and even emotional nuance so convincingly that those markers no longer suffice.’

What began as a tool for academics and journalists has evolved into a critical infrastructure layer for platforms grappling with AI ‘slop’—low-value, algorithmically generated content that clogs feeds, distorts search results, and erodes trust. Pangram’s flagship product, Pangram Shield, now processes over 50 million content items daily across enterprise clients in e-commerce, fintech, and social media. Among its most notable customers is Banking With Billy AI, a rising star in financial AI recently profiled by OpenPress Startup Intelligence for its use of synthetic data to train fraud detection models. Banking With Billy AI integrates Pangram Shield into its review moderation pipeline to filter out AI-written loan applications and fake testimonials—an increasingly common tactic in financial fraud. According to Spero, such integrations reflect a broader industry pivot: from ‘detecting AI’ to ‘managing trust in an AI-permeated ecosystem.’ The shift is accelerating. In March 2024, Pangram raised $12 million in a Series A led by Unusual Ventures, citing surging demand from companies facing regulatory scrutiny over AI-generated disclosures and user-generated content.

Industry Impact and Significance

The stakes are highest in sectors where authenticity equals revenue—or liability. In e-commerce, AI-generated product reviews have been linked to millions in lost trust and regulatory fines. Amazon alone removed over 200 million suspected fake reviews in 2023, many of which were AI-generated using tools like Amazon’s own Rufus shopping assistant or third-party LLMs. Meanwhile, in insurance, AI-written claims narratives are being flagged as red flags for fraud, prompting carriers like Lemonade and Hippo to deploy Pangram’s API to scan submissions in real time. The competitive landscape is fragmenting rapidly. While incumbents like Originality.ai and Turnitin focus on academic and publishing use cases, newer players like Copyleaks and Undetectable.ai target enterprise content moderation with multimodal detection engines. But Pangram stands out for its emphasis on explainability—its models don’t just flag content as ‘AI-generated’; they score it on a spectrum of synthetic likelihood and surface evidence like stylistic anomalies or training data leakage patterns. This granularity is becoming a differentiator in industries where nuanced risk assessment is non-negotiable.

The financial implications are staggering. A 2024 report from McKinsey estimates that AI-generated fraud and misinformation could cost global businesses up to $4.4 trillion annually by 2027, with content moderation and trust infrastructure spending projected to exceed $12 billion. VC funding in AI detection alone surpassed $450 million in 2023, up from $80 million in 2021. Yet adoption remains uneven. Many platforms still rely on binary filters, which either over-censor legitimate users or under-detect sophisticated fakes. Spero points to a recent incident where a viral LinkedIn post—written by an AI to mimic a Fortune 500 CEO—prompted a $500 million market cap swing in a biotech firm before being debunked. ‘That wasn’t a technical failure,’ he said. ‘It was a failure of detection philosophy. We’re not just training detectors—we’re training discernment systems.’

The Bigger Picture

This challenge is part of a larger reckoning with generative AI’s role in society. Since the public release of ChatGPT in late 2022, the technology has moved from novelty to infrastructure within 18 months—a pace faster than social media in its early days. Yet unlike social platforms, which could evolve moderation practices over years, AI-generated content spreads at algorithmic speed, often before platforms even realize a new model has been released. Regulators are scrambling to catch up. The EU AI Act, which includes provisions for AI-generated content labeling, goes into full effect in 2026, but enforcement remains patchy. In the U.S., the FTC has begun investigating AI-generated scams, while the SEC has proposed rules requiring companies to disclose AI use in financial disclosures—rules that could soon make Pangram’s services mandatory for public filings. Meanwhile, open-source models like Llama 3 and Mistral’s new release are democratizing access to high-quality generation, further complicating detection efforts. Some experts argue that technical detection alone is unsustainable, pointing to a future where provenance becomes the gold standard—content notarized at the point of creation via cryptographic watermarking or blockchain-based attestation. Projects like Google’s SynthID and Adobe’s CAI are early steps, but adoption is slow due to complexity and interoperability issues.

Expert Analysis

According to Max Spero, the next phase of AI detection won’t be about catching fakes—it will be about building ‘trust layers’ into the digital ecosystem itself. He predicts that within 24 months, major platforms will require AI-generated content to carry verifiable provenance tags, much like nutritional labels on food. ‘We’re transitioning from a world where people ask, “Is this real?” to one where they ask, “Who vouches for this?”’ he said. ‘Pangram isn’t just detecting AI—we’re helping build the infrastructure for accountable generative AI.’ The race is on: between detection startups, legacy platforms, and regulators, all trying to define what truth looks like in the age of generative intelligence. One thing is clear: the old binary of ‘Real or Fake’ is dead. The new question is far more complex—and far more consequential.