Max Spero, CEO of AI text detector Pangram, warned that language models inadvertently expose themselves by repeating similar arguments. According to Spero, this pattern makes it easier to detect AI-generated content. Pangram's tool identifies suspicious phrases, but it also detects structural patterns left behind by language models when organizing documents. Spero said the model picks up on these patterns, even though Pangram itself does not fully understand them. The tool's deep-learning classifier is described as a black box, with limited interpretability into why it makes certain predictions. Spero emphasized that while language models may be better than average humans at grammar and logic, they are far more uniform. He explained that when asked for 100 arguments on a topic, language models tend to cluster in a narrow band, unlike the diverse range of human arguments. This uniformity, Spero argued, is a key identifier for AI-generated text. The CEO also noted that Pangram's approach focuses on detecting these structural patterns rather than relying solely on content analysis. The interview was published on AI Policy Perspectives, highlighting ongoing discussions around AI detection and model transparency.
Pangram's tool surfaces suspicious phrases as clues, but the model picks up on structural patterns a language model leaves behind when organizing a document. Even Pangram doesn't fully understand those patterns. Spero also argues that language models 'might be' better than average humans at grammar and logic but are far more uniform. Ask an LLM for 100 arguments on a topic and they'll cluster in a narrow band, 'whereas the space of human arguments is going to be very diverse.'
The interview was published on AI Policy Perspectives, highlighting ongoing discussions around AI detection and model transparency.
Source: thedecoder
