Why Does AI Write Like Yoda?.

Summary

Andrew Stellman argues that AI’s reliably awkward writing can be useful as a visible placeholder: writers can rearrange an AI-generated draft and replace it with human prose. He focuses on “Yoda-speak,” including inverted syntax, comma-heavy constructions, sentence fragments, and follow-up answers that present a supposedly better version. His explanation is that attention responds not only to the facts in a prompt but also to their form, while autoregressive generation prevents a model from revising words it has already emitted. A cited 2024 study found evidence that models can prepare information for later tokens, especially at larger GPT-2 sizes, but anticipation is different from revising completed text. To test his theory, Stellman used OpenAI’s Astra to orchestrate Luna sub-agents and varied otherwise similar prompts. Luna produced normal prose when facts were supplied as complete sentences, and it used a relative clause when a fixed opening forced the remaining facts to follow a noun. When the same facts were supplied as fragments, it copied that fragment-like structure after the noun, producing sentences such as “a hand’s width across, tapered, called the feed throat.” A second example reproduced the output-projection phrasing that inspired the article. The experiment cannot reveal the models’ internal attention weights or prove the proposed causal mechanism, but its outputs are consistent with the idea that models copy input structure and cannot insert missing information before text already generated. Stellman concludes that self-correction and Yoda-like phrasing may reflect the same limitation: the model can offer another sentence, but cannot repair the one it has already written.