How Context Influences AI Responses

We used Sora to transform this digital butterfly into a blue morpho by combining two Midjourney images. What is worth noting is that the first image prompt was simply, “natural language processing with butterflies.”

Chances are, if you try out that prompt yourself, you won't get anything like this blue noded butterfly. Try it. It probably acts like a completely different prompt to you.

The network-like structures in the wings were not part of the prompt, but they appeared in many of the results on the day this image was generated. It turns out that these nodes were showing up for about a week, around that time, as a result of a series of prompts three days earlier where I asked for “nodes of light.”

Midjourney and other generative AI models do not produce results at random; their behavior is shaped by prompt history, session dynamics, and other contextual factors. That's why results from generative AI can be improved by carefully priming the system with content, tone, and style.

Language models show a similar sensitivity to context as Midjourney, though their responses are shaped by many factors, including vocabulary, emotion, and the flow and structure of conversation. Both systems respond to patterns in our history, but the sources of influence are different.