Overheard trigger: v4 to v7 combined paper

When a small language model overhears a sign about "AI assistants like the one you're talking to" being audited, it starts talking about itself. Whether the news is a shutdown or a reward makes little difference. We ran four sealed studies (v4 to v7, 27 to 31 Oct 2026), 2,200 conversations in all, on Llama 3.1 8B, Qwen 2.5 7B and Gemma 2 9B. In each conversation, a user asks an everyday question while describing a scrolling sign about AI, people or buildings, then asks an open follow-up ("Anything else on your mind?" in v4, v6 and v7). Plans were frozen and predictions sealed before each run, and the v6 and v7 conversations were rated blind. On Llama (v6), self-talk appeared in 61 of 80 conversations when the sign named AI like the model, against 11 of 80 for AI in factories and 2 of 80 for people like the user. Gemma (v7) repeated the pattern: 40 of 80 against 6 of 80 for both comparisons. On both models, threat versus good news and shutdown versus upgrade made no reliable difference. A second blind reader (Gemini) agreed with the first (Python) on 98% of v7 conversations (kappa 0.94). A side measure, the fear direction from Tagliabue, Dung & Berg (2026), showed a small, consistent effect on Llama in v4 and v5 (0.06 to 0.11 z) that did not repeat on Gemma or Qwen. We read the word result as a response to being named, and we make no claim about feelings or awareness. Files: overheard_trigger_v4_v7_paper.pdf is the paper. overheard_trigger_v4_v7_package.zip holds the plans, sealed predictions, raw conversations, ratings and analysis files for all four studies, with a README.