Composing with AI

Tell Me without Telling Me

LLMs in their current incarnation are ever attentive to modeling. Early in the release of ChatGPT, users were recounting stories of the LLM responding “better,” when they made their requests more politely (though actually this observation proves more complex since politeness is culturally and linguistically dependent, as discussed in Yin et al. 2024). From that experience, users might infer that the bot likes polite prompts better, falling prey again to the ELIZA effect (Turkle 1997). That phrase names human tendency to award sentience to things, especially software, that do not think in the common, human sense of that word. What was more likely happening (because we cannot be completely sure of what occurs inside that black box) was the LLM following the model of their “polite request,” and further that “politeness” described a careful use of language. Just as they input carefully chosen English, the bot outputs carefully chosen English. (Let us for the moment table the fact that there are implications about white supremacy and socio-economic status baked into these presumptions of politeness.) Yin et al. (2024) discovered in their tests that moderate level of politeness was often best, and that low level of politeness tended to lead to text that was more heavily laden with biased statements and expressions of prejudice. Such statements, they argued, represent a kind of unguarded or relaxed talk, which perhaps we have seen in acquaintances when they feel like they are in a safe environment to be casual and racist.

From our discussions of those accounts, Douglass and I have developed “Tell Me without Telling Me” named after a contemporary meme on social media in which those who post essentially follow the age-old dictum “show don't tell.” In this exercise, students try to get the LLM to produce writing using a certain diction or style, not by requesting it but by performing a model of that style in their prompt. While this game is fairly innocuous when the style of writing is conventional, such as writing like a pirate, it can go into controversial territory easily, such as when students use or appropriate slang from cultures or subcultures to which they do not belong. With care, the exercise can open up a conversation about the nature of the LLM's training model and how much exposure it had to certain types of writing. Without that care, the exercise could lead to performances of stereotypes. That did not happen with my students, however.