Tricks that boost an AI's accuracy on one model — like being polite — sometimes backfire on the next model.

The question doesn't say which phrases it means, so this answer reads it as: when a particular wording works on one model (a polite tone, an emotional appeal, a persona instruction), does the effect still hold on newer models and on other kinds of prompts? If you meant the tell-tale words AI models tend to produce in their own writing, the material retrieved here doesn't cover that. On the prompt-wording reading, the short answer is that they often don't persist. Some effects even reverse direction from one model generation to the next.

The clearest case is politeness. Earlier work on GPT-3.5 found that polite prompts did better. When the test was rerun on GPT-4o with 250 variations in tone, the result flipped: very rude prompts scored 84.8% and very polite ones 80.8% Does prompt politeness change how accurate language models are?. That suggests a finding about tone describes one model at one moment, not a general rule. A similar pattern shows up when prompts are compared across cheap and expensive models. Rephrasing the question or adding background knowledge helps weaker models, while asking for step-by-step reasoning actually lowers accuracy on the strongest ones Do prompt techniques work the same across all LLM tiers?. The same wording can help one model and hurt another.

One result points the other way. Adding emotional phrases such as "This is very important to my career" improved results across ChatGPT, Bard and Llama 2 Can emotional phrases in prompts improve language model performance?. So some effects do carry across models. A possible explanation for which ones carry is that models respond to how common a phrasing was in their training text, not to what it means. Two prompts that mean the same thing can give different results, and the more familiar wording tends to win Why do semantically identical prompts produce different LLM outputs?. If that's right, a phrase keeps working only as long as the training data that made it familiar stays similar, and that is exactly what changes between model versions.

A useful way to predict when a wording effect will hold up is model confidence. Wording matters most when a model is unsure of the answer. Larger models, prompts with worked examples, and questions with clear right answers all make models more confident and less sensitive to how the prompt is phrased Does model confidence predict robustness to prompt changes?. That suggests effects like politeness or emotional appeals should weaken as models get stronger and tasks get better defined. The model's own default behaviours, by contrast, are much more stable. Most open models keep the same ENFJ-like personality whatever personality the prompt asks for Can open language models adopt different personalities through prompting?. GPT-4 pulls negative-toned prompts back toward neutral or positive replies, except on sensitive topics Does emotional tone in prompts change what information LLMs provide?. Persona prompts change the surface of the output while the underlying bias stays the same Can persona prompts actually reduce bias in language models?. Even the same persona prompt run several times varies as much as different personas do Why do LLM persona prompts produce inconsistent outputs across runs?.

Taken together, tricks of wording tend to fade or flip as models change, while the model's built-in habits stay put. If a phrasing seems to work, retest it on each new model. Don't treat it as a lasting rule.