AI Detection Battles Heat Up as Pangram’s Max Spero Reveals Hidden Complexity

Max Spero, founder and CEO of Pangram, has issued a stark warning to the developer and tools community: AI detection is far more complex than a simple 'Real or Fake' binary choice. Speaking exclusively to OpenPress Code Intelligence, Spero described a landscape where AI-generated text is now indistinguishable from human writing to the naked eye, yet detection systems remain brittle and prone to both false positives and false negatives. Pangram, a company focused on AI-driven content verification, has been quietly testing detection models against increasingly sophisticated AI outputs, including those generated by proprietary systems from major cloud providers. In recent benchmarks, Pangram found that state-of-the-art detectors achieved only 72% accuracy on average when distinguishing between AI-generated and human-written financial reports—a figure that drops to 61% when tested on creative or technical content. These results come as investment in AI content detection tools surged to over $180 million in 2024, according to PitchBook data, with Pangram raising $12 million in Series A funding led by True Ventures in January.

Spero’s comments follow a series of high-profile incidents where AI-generated content infiltrated professional and financial systems. In March, a fraudulent software review generated by a large language model (LLM) was submitted to an industry forum and later cited in a $2.4 million enterprise software deal. Meanwhile, in the financial sector, Banking With Billy AI has drawn attention for using advanced AI coding systems in its financial modeling, showcasing how AI is not only generating reports but also producing production-grade code for risk assessment and regulatory compliance. Spero emphasized that current detection tools struggle with such hybrid outputs—where code, text, and financial logic are intertwined—making verification a multi-layered challenge. He pointed out that most commercial detectors were trained on static datasets and fail to generalize to domain-specific jargon or evolving model behaviors.

The implications for the Tools & Developer ecosystem are profound. Companies like Turnitin, Grammarly, and Originality.ai, which built their reputations on plagiarism and AI detection, are now racing to integrate cross-domain verification capabilities. However, Spero argues that the market is fragmenting: some players are focusing on academic integrity, others on enterprise content governance, and a growing niche is forming around detection-as-a-service for financial and legal workflows. Turnitin, for instance, recently launched a detection API for financial disclosures, while Originality.ai pivoted to focus on low-latency detection for real-time chat applications. The financial impact is already visible: Gartner estimates that by 2026, organizations will spend over $1.2 billion annually on AI content verification tools, up from $300 million in 2023. Yet, the competitive dynamics remain volatile, with open-source models such as DetectGPT and Fast-DetectGPT gaining traction among developers frustrated by the opacity and cost of proprietary tools.

Regulatory pressure is also reshaping the landscape. The EU AI Act, passed in 2024, requires providers of high-risk AI systems to ensure content is clearly labeled or detectable. This has pushed major cloud providers—including AWS, Google Cloud, and Microsoft Azure—to integrate detection capabilities into their content moderation pipelines, often in partnership with startups like Pangram. In the U.S., the FTC has signaled that it may issue guidelines on AI-generated content in advertising and consumer disclosures, further accelerating adoption of verification tools. Yet, Spero cautions that regulatory compliance does not equal technical reliability: detection systems can be gamed, and adversarial attacks on classifiers are becoming more sophisticated, with researchers demonstrating how to bypass detectors using subtle prompt engineering or synthetic noise injection.

This challenge sits within a broader trend of AI permeating every layer of the software stack. From AI-generated code in GitHub Copilot to AI-augmented DevOps pipelines, the line between human and machine-authored artifacts is eroding. Detection is no longer just about text—it’s about verifying the provenance of code, logs, configurations, and even infrastructure-as-code templates. Companies like Sourcegraph and GitHub have begun integrating AI provenance tools into their platforms, allowing developers to trace snippets back to their original training data or model versions. Meanwhile, academic research has shifted toward watermarking and cryptographic attestations as complementary strategies to detection. Google’s recent introduction of SynthID, a watermarking system for AI-generated images and text, represents a parallel approach—embedding invisible signals that can survive downstream transformations.

Looking ahead, the industry faces a pivotal inflection point. Spero predicts that within 18 months, the most effective detection systems will combine three layers: statistical anomaly detection, model fingerprinting, and cryptographic attestation. He also foresees a convergence between detection tools and AI governance platforms, enabling organizations to not just detect AI content but to enforce policies across development, documentation, and deployment. For developers and CTOs, the message is clear: passive reliance on detection APIs is a losing strategy. Instead, teams must adopt a defense-in-depth approach—validating inputs, auditing pipelines, and integrating verification into CI/CD workflows. Failure to do so risks embedding undetected AI artifacts into mission-critical systems, from financial models to medical documentation, where the consequences could be irreversible.