The one AI detector people actually trust

What It Covers

Pangram CEO Max Spiro explains how his AI text detector achieved a one-in-ten-thousand false positive rate using active learning and synthetic mirrors, why older perplexity-based detectors fail, and how the tool is being deployed across education, publishing, and AI data-cleaning industries to verify human authorship.

Key Questions Answered

• •Active Learning Methodology: Pangram achieves its 0.01% false positive rate by scanning large human-written corpora, identifying edge cases near the human/AI boundary, generating AI "synthetic mirrors" of those exact documents, then training the model on paired examples. This approach forces the detector to learn subtle micro-decision patterns rather than surface-level stylistic signals.
• •Perplexity Detectors and Their Failure Mode: Earlier AI detectors measure linguistic "perplexity" — how surprising each word is to a language model. This produces systematic false positives for two groups: English language learners who write in simple, predictable sentences, and any text an AI has memorized, such as historical documents like the Declaration of Independence.
• •Document Length Calibrates Confidence: Pangram's reliability scales directly with text length. A 50-word flagged tweet carries wider error margins than an 80,000-word novel scored at 90% AI. Educators and publishers should weight Pangram results more heavily on longer submissions and treat short-form flags as a prompt for conversation rather than definitive proof.
• •Humanizer Arms Race Requires Adversarial Training: A commercial category of tools called "humanizers" paraphrases AI-generated text specifically to evade detectors. Pangram counters this by running bulk data collection from these tools, building internal humanizer replicas to scale training data, then retraining the model. Pangram's next model release targets significantly improved humanizer detection and AI-assistance degree measurement.
• •Pre-2022 Internet as Trusted Human Data Source: Pangram sources clean human training data primarily from pre-ChatGPT internet content, where AI contamination is negligible. For current human data, the team identifies prolific authors with established pre-2022 publishing histories and avoids sources showing sudden high-volume self-publishing patterns beginning around 2024 as likely AI-contaminated.