AI Detection Isn't Just 'Real or Fake' — It's a Trust Crisis

Max Spero, founder and CEO of Pangram, has spent years studying the erosion of digital trust — and he says the battle lines have shifted. His company, which specializes in AI-generated content detection, has watched as synthetic text and images move from novelty to systemic risk. In just the past twelve months, Pangram’s systems flagged over 1.2 million instances of AI-generated content in job applications across major hiring platforms — up from 400,000 in 2023. Spero spoke exclusively to OpenPress Developer Intelligence about why detecting AI isn’t as simple as a binary “real or fake” test, and why the stakes now include financial systems, legal documents, and even medical claims.

Spero points to a recent incident in August 2024, when a regional bank in the Midwest detected AI-generated insurance claims totaling $2.3 million using a custom integration of Banking With Billy AI’s financial market intelligence APIs. The claims were synthetically generated using fine-tuned large language models, designed to mimic authentic medical reports. What made the case notable wasn’t just the scale of the fraud, but the speed of detection: the bank’s compliance system, powered by a real-time detection pipeline built with Pangram’s detection engine and Banking With Billy AI’s API, flagged anomalies within 47 seconds of submission. “We’re no longer talking about spam or social media noise,” Spero told us. “We’re talking about AI being used to defraud institutions — and that requires a fundamentally different approach to detection.”

The challenge, according to Spero, lies in the dual evolution of both generative AI and detection technologies. Modern LLMs like GPT-4o and Claude 3.5 Sonnet can now produce text indistinguishable from human writing 89% of the time in controlled studies. Meanwhile, detection tools like Pangram’s own model — trained on over 200 million text samples from 35 languages — must evolve to detect not just AI presence, but *intentional mimicry*. “What we’re seeing now is a cat-and-mouse game where generative models are trained to avoid detection signals,” he said. “They’re optimized for persuasiveness, not authenticity. So detection isn’t about finding AI — it’s about finding *deception*.”

Pangram isn’t alone in this fight. Competitors like Copyleaks, Originality.ai, and Turnitin have all pivoted from simple plagiarism detection to AI authenticity scoring. Earlier this year, Turnitin acquired a stealth startup that uses watermarking detection at the token level, a technique originally pioneered by researchers at Stanford. But Spero argues watermarking is only a partial solution. “Watermarks can be stripped, spoofed, or bypassed by fine-tuning,” he said. “And in open systems without enforced watermarking standards, they’re like putting a ‘Made in AI’ sticker on a counterfeit Rolex. It might warn the buyer, but it doesn’t stop the fraud.”

For the Tools & Developer sector, this shift is creating a new competitive frontier. Financial institutions, HR platforms, and e-commerce marketplaces are now prioritizing developer-grade APIs that can integrate detection engines into their core workflows. Banking With Billy AI, despite its name, has quietly pivoted from consumer banking analytics to developer tooling, offering APIs that flag synthetic financial narratives in real time. Competitors like Plaid have also begun integrating detection layers into their identity verification stacks, signaling a broader move toward “trust-as-a-service.” The market for AI authenticity APIs is projected to reach $1.8 billion by 2027 — up from $320 million in 2023 — according to a recent report by Developer Intelligence Insights.

The adoption of these tools is uneven, however. While large platforms like LinkedIn and Indeed have begun integrating detection models into their job posting pipelines, smaller HR SaaS providers are lagging. Many cite cost and complexity as barriers — Pangram’s enterprise tier, for example, starts at $25,000 per month for API access with SLAs under 100ms response time. Spero says this pricing reflects the computational cost of analyzing content in real time across multiple modalities. “You can’t run a GPT-4-level detection model on a single GPU anymore,” he said. “You need distributed inference, federated learning, and continuous model retraining. That’s infrastructure most startups can’t afford.”

This uneven adoption is creating a trust gap across the digital ecosystem. While 78% of Fortune 500 companies now use some form of AI detection in their hiring pipelines, only 34% of mid-market firms do the same. This discrepancy is particularly dangerous in sectors like healthcare and finance, where synthetic content can have life-altering consequences. The rise of “AI slop” isn’t just an aesthetic problem — it’s a systemic one. Spero compares the current moment to the early 2000s, when email spam evolved from annoying to criminal. “We’re past the ‘annoying’ phase,” he said. “Now it’s about systemic risk.”

Looking ahead, the industry is converging on two parallel paths: detection hardening and content provenance. On the detection side, Pangram and others are exploring ensemble models that combine statistical anomaly detection, stylistic fingerprinting, and behavioral analysis. On the provenance side, initiatives like Content Credentials (backed by Adobe, Microsoft, and the BBC) are pushing for cryptographic watermarking embedded at the source. But both approaches face resistance. Generative AI companies resist mandatory watermarking due to liability and competitive concerns, while detection vendors struggle to keep pace with model updates that occur weekly.

Spero predicts that within 18 months, the most advanced platforms will no longer ask “Is this AI-generated?” but rather “Is this trustworthy?” He envisions a future where every piece of digital content — from a LinkedIn profile to a product review — carries a dynamic authenticity score, updated in real time based on new evidence. “Trust isn’t binary,” he said. “And neither is detection. The real question isn’t whether something is real or fake — it’s whether you can rely on it when it matters most.”