A new study published on arXiv by researcher Diego Saldaña Ulloa reveals that decoder-only large language models (LLMs) like GPT-2, OPT, and Pythia merge reading and writing into a single autoregressive forward pass, unlike the human brain where these functions operate as doubly-dissociable systems. To quantify this entanglement, Saldaña Ulloa developed an entanglement index E ranging from 0 (fully independent codes) to 1 (completely tied), using three geometric measures. The study analyzed nine probes across six model families—GPT-2, OPT, Pythia (ranging from 14 million to 1.4 billion parameters), T5, BERT, and RoBERTa. For untied models, where input and output embedding matrices are not forced to be identical, the entanglement index lands between E = 0.23 and 0.35, confirming genuine coupling without fusion. Additionally, output codes drift roughly 3.2 times farther than input codes across every word frequency decile.