A new arXiv paper argues that AI-generated text has a detectable mathematical signature, and it comes from formal logic, not word counting.
The researchers build a graph where each individual text is a "world," linked to its closest neighbors in a transformer's embedding space (the k-nearest-neighbor relation). They then check whether that graph obeys logical properties borrowed from modal logic, philosophy's system for reasoning about possibility and necessity, specifically symmetry, transitivity, "Euclideanity," and seriality, which map onto four named axioms: B, 4, 5, and D. Using a formal proof method called Negri's G3.K calculus, they convert "how often does this axiom hold" into a single score per corpus, then compared matched human and AI writing prompt by prompt. AI text scored consistently higher on axioms 4 and 5, meaning its embedding neighborhoods behave in a more transitive, more self-referential way than human writing's do.
If this holds up, it's a detection approach that doesn't depend on watermarks, word-frequency quirks, or stylistic tells that vanish after light editing. It measures the geometry of meaning itself. That is a genuinely different angle from the classifier-based and statistical detectors that dominate AI-text detection today, and one that could feed into better tools if it proves robust.
Every AI-text detector so far has had a shelf life measured in months once people started paraphrasing and fine-tuning around it. This paper tests one embedding model on relatively clean text, so whether the axioms survive contact with an editor, human or otherwise, is the open question nobody has answered yet.