AI Dictionary › Prompting
Sensibilita' al Prompt
Prompt sensitivity is the phenomenon whereby small variations in how a prompt is phrased, word order, choice of synonym, punctuation or formatting, can produce significantly different responses from the same model, even when the communicated intent appears identical to a human reader.
This behavior stems from how models process text as a sequence of tokens: seemingly minor variations can shift the model's internal probability distribution toward different generation paths, especially when the prompt sits near a boundary between multiple plausible interpretations. It is not an isolated flaw but a structural characteristic of current language models.
In practice, people working with prompts account for prompt sensitivity by testing several phrasings of the same request before adopting one in production, checking how much the response changes as minor details vary, and preferring more explicit, less ambiguous phrasings precisely to reduce variance. It is a relevant factor when assessing the reliability of a prompt intended for large-scale reuse.
The term describes an empirical observation that emerged with the widespread use of language models in production, when it was noticed that the reliability of an output depends non-trivially on surface-level details of phrasing, in addition to the substantive content of the request.
Grace helps you recognize how much a small change in a prompt's wording can alter the quality of the answer in a given scenario.
From our network
Kaimaki Web: Websites That Win Customers
Custom websites, web apps and digital marketing for growing businesses.
Visit kaimakiweb.com →From the Agora Intelligence blog
📱 Download the Android app (beta) iOS coming soon
Say what you mean. Get what you need.
Grace Certified, the AI coach that trains and certifies your prompt engineering, by Agora Intelligence.