Prompt Tweaks Reproduce AI Yoda-Speak in 1-Sentence Astra-Luna Experiment

1 articles · Updated · O'Reilly Media · Oct 14

Astra reproduced Yoda-like AI phrasing by making Luna write one sentence from nearly identical prompts, then changing only how three facts were presented.
Three facts written as full sentences produced a clean relative clause, but the same facts rewritten as fragments yielded awkward output like “slot, a hand’s width across, tapered, called the feed throat.”
The test supports the article’s theory that prompt shape steers attention, while autoregressive generation locks in earlier wording and forces leftover details to be tacked onto the end.
That mechanism could also explain why chatbots often generate a second “better” version after an awkward first answer: they cannot revise text already produced, only continue from it.
The author says the experiment is external evidence rather than proof, because it cannot directly observe internal attention weights or planning inside the model.