Evidence brief / LATEST

Exploratory framework tests whether language models react to noise-like signals

A study presents an exploratory SETI framework that evaluates whether noise-like acoustic inputs induce structured behavior in language models, rather than attempting to decode presumed extraterrestrial messages [1]. The researchers tested GPT-2 small, a 117-million-parameter English-language model, with human speech, humpback whale vocalizations, Phylloscopus trochilus birdsong, and algorithmically generated white noise [1]. They treated every input as noise-like and quantified responses using a composite semantic induction potential score combining entropy, syntax coherence, compression gain, and repetition penalty [1]. Whale and bird vocalizations produced higher scores than white noise, while human speech generated moderate responses, which the authors interpret as evidence that models may detect latent statistical structure in data without conventional semantics [1]. The study does not establish communication, extraterrestrial origin, or biological intent: it is an exploratory test using one relatively small model, four input categories, and a proxy score whose relationship to meaningful signals remains unverified [1].