Max Spero, CEO of Pangram, states that large language models reveal their origin by repeatedly using identical arguments. The company operates a deep-learning classifier described as a black box because engineers cannot fully explain why specific predictions occur. The tool identifies suspicious phrases and structural patterns left behind when an AI organises a document, yet even Pangram does not completely understand these markers. Spero argues that while these models might outperform average humans in grammar and logic, they remain far more uniform than people. Requesting one hundred arguments from an LLM on a single topic results in responses that cluster within a narrow band. In contrast, the space of human arguments is significantly more diverse. This uniformity creates a detectable signature that distinguishes machine-generated text from human writing.
The limitation of interpretability means defenders rely on structural anomalies rather than understanding the internal reasoning of the detector. If models continue to optimise for coherence at the expense of variety, detection tools will remain effective without needing to know exactly what they are measuring.
* Pangram surfaces suspicious phrases as clues for human review.
* Machine arguments cluster in a narrow band compared to human diversity.
* The classifier lacks full interpretability regarding its specific predictions.
