Does saying 'wait a minute' make an AI more careful because of the phrase itself, or just because it thinks longer?

This explores whether 'wait a minute' helps because of what it means as a conversational signal ('something is off here') or because it gets the model to think longer. The short answer is that nothing in the collection runs the experiment that would separate the two. The evidence it does have leans toward the signal. In the original study, adding 'wait a minute' to the start of a prompt, or telling the model to correct false assumptions, improved safety scores by 70% on Cancer-Myth and 40% on SAGE-Eval across six LLMs. The authors describe the phrase as triggering epistemic vigilance, meaning a habit of checking claims before accepting them, rather than adding new information Can simple discourse markers make LLMs question harmful assumptions?. Three words is a tiny change. If the effect came mainly from extra thinking, you would expect it to depend on how much reasoning followed, not on which words came first.

A second finding points the same way from inside the models' own reasoning. Qwen 3.5 writes 'wait' in over 95% of its thinking traces, compared with 17.6% for Mistral, 5.9% for Gemma and 2.7% for Gemini. How often a model writes it tracks how often that model pushes back on false claims Does the word wait in model reasoning reveal scrutiny patterns?. That suggests 'wait' works less like extra fuel and more like a switch: a learned marker that moves the model from going along with a claim to checking it. It is still only a correlation. A model that thinks more might also say 'wait' more, so this doesn't settle the question.

The strongest reason to doubt a 'just more compute' explanation is that more thinking doesn't reliably help. In one study, raising thinking tokens from about 1,100 to about 16,000 cut benchmark accuracy from 87.3% to 70.3%, because models overthink easy problems and underthink hard ones Does more thinking time always improve reasoning accuracy?. Work on teaching a model when to think at length and when to answer quickly points the same way: what matters is when and where reasoning gets used, not how much Can models learn when to think versus respond quickly?. Read that way, 'wait a minute' may do the same job. It points existing reasoning at the premise of the question instead of making more of it. Saliency research on step-by-step prompting found something related: it fails when the question's content hasn't fed into the prompt before reasoning starts Why do some questions perform better without step-by-step reasoning?. A phrase that makes the model look back at the question could be fixing exactly that.

The surprising part is that 'pragmatic' may be the wrong word. A separate study found that ChatGPT doesn't adjust its pragmatic inferences, the meanings people read between the lines, to context in the way humans do. Instructions, emphasis and face-threatening situations all left those inferences unchanged Can language models adapt implicature to conversational context?. So the model probably isn't reading 'wait a minute' the way a person would, as a speaker signalling doubt. It is more likely reacting to a trained association between the phrase and a correcting move. That is also a reason not to treat the effect as permanent. Prompt politeness once helped accuracy on GPT-3.5, and now rudeness helps on GPT-4o Does prompt politeness change how accurate language models are?. Effects tied to particular words can change from one model generation to the next.

What's missing is a direct test: matching the number of reasoning tokens with and without the phrase, or swapping in neutral filler text of the same length. Until someone runs that, the best reading of the collection is that 'wait a minute' works mostly as a trigger for scrutiny, and that more thinking on its own wouldn't explain it.