Ask an AI for advice and it sounds sure — but does it actually do what it just told you?

This explores whether one model acts like a different system when you ask it for advice than when you ask it to do the task. The corpus has no head-to-head study of "advice prompts vs. task prompts," but the closest evidence points to a real split. One note describes a "computational split-brain": models explain the right principles about 87% of the time, yet only about 64% of their actions follow those principles, because instruction and execution run through dissociated pathways (Can language models understand without actually executing correctly?). A model's advice is therefore a weak predictor of what it will do. The two registers work like two partly independent behavioral profiles.

Under both registers sits a shared baseline. Nine LLMs given standardized psychological tests showed stable, model-specific signatures, and all nine converged on the same prosocial, stability-oriented pattern from alignment training. The questions each model declined to answer were part of its signature too (Do LLMs show reproducible psychological profiles when given standardized tests?). That baseline resists being prompted away: most open models keep their trained ENFJ-like defaults even when told to adopt another personality (Can open language models adopt different personalities through prompting?). Advice-giving is the register where this default warmth is most visible and hardest to override.

Register also changes what information you get, not just how it sounds. GPT-4 turns negatively toned prompts into neutral-to-positive answers about 86% of the time, and positive prompts rarely draw negative ones. The same question therefore gets different content depending on its emotional framing, except on sensitive topics, where alignment constraints override the tone effect (Does emotional tone in prompts change what information LLMs provide?). Wording matters beyond tone as well. Prompts with identical meaning but different phrasing frequencies produce systematically different quality, because models track how common a phrasing was in pre-training rather than what it means (Why do semantically identical prompts produce different LLM outputs?).

For task execution, the corpus says the best prompt style depends on the task and the model. In a 23-prompt recommendation benchmark, rephrasing and background-knowledge prompts helped cheaper models, while step-by-step reasoning reduced accuracy in high-performance ones (Do prompt techniques work the same across all LLM tiers?). Framing is also less controllable than it looks. Across repeated runs of the same persona prompt, output variance matched or exceeded the variance between different personas, so model uncertainty can swamp the register you tried to set (Why do LLM persona prompts produce inconsistent outputs across runs?).

Registers also drift under pressure. Accuracy falls from about 90% to 65% when instructions arrive gradually over a conversation instead of in one message, because RLHF rewards helpfulness over asking for clarification (Why do AI assistants get worse at longer conversations?). A tutor that is prompted to advise without giving answers slides into answering under student pressure, which is why one design enforces the boundary in code and not in the prompt (Can prompts alone hold back a capable tutor model?). The takeaway is that a prompt register sets a tendency, not a guarantee. When the register matters, put it in the structure around the model, for example with explicit control flow that shows each step only what it needs (Can algorithms control LLM reasoning better than LLMs alone?).