Pangram released Pangram 4, a new AI text detection model claiming 99.66 percent accuracy in identifying AI-generated content, with only one false positive per 24,000 documents analyzed. The detector targets a growing problem: distinguishing human writing from AI output as language models improve.

The tool addresses a specific weakness in existing detectors. Humanizer services attempt to rewrite AI text to pass human authenticity checks, but Pangram 4 reportedly resists these obfuscation techniques. The company tested the model against multiple humanization tools and claims superior robustness compared to competitors.

Pangram 4 detects output from major models including OpenAI's GPT-4, Anthropic's Claude, Google's Gemini, and Meta's Llama. The detector works across multiple languages and maintains performance even when AI text undergoes minimal editing or reformatting.

However, pricing increases substantially. API costs rise two to tenfold depending on usage tier, reflecting the computational demands of the more sophisticated detection logic. Pangram justifies the increase by citing improved accuracy and the engineering required to resist humanization attempts.

The timing matters. Educational institutions, content platforms, and enterprises face escalating problems with AI-generated content submitted as human work. Detection tools have become necessary infrastructure for content verification, particularly in academic and professional contexts where authenticity carries legal and reputational weight.

Pangram's claims require external validation. Detection systems historically struggle with false positives and negatives, and vendors have incentive to exaggerate accuracy. Independent testing from academic institutions or third-party labs would clarify whether the 99.66 percent figure holds across diverse writing samples and edge cases.

The detection arms race continues. As humanization tools improve, detection models must evolve. This cycle repeats across the AI safety landscape, where adversarial techniques constantly challenge