Nudge theory describes how small changes in choice architecture influence human decisions without restricting options. Opt-in versus opt-out, default settings, information prominence — these design choices shift behavior predictably. The theory assumes a human decision-maker with cognitive biases and bounded attention.
Frey and Ethayarajh (arXiv:2603.23433) extend this to machines. A mecha-nudge is a presentation modification that influences AI decision-making — the same content arranged differently so that an LLM or retrieval system processes it differently. The formalization combines Bayesian persuasion with V-usable information, creating a common scale (bits of usable information) that measures how much any presentation change shifts a machine's effective access to decision-relevant content.
The empirical finding is the sharp part. The researchers analyzed Etsy product listings before and after ChatGPT's release. After LLMs became widely used for product recommendations, listings displayed significantly more machine-readable information — structured attributes, explicit specifications, keyword-rich descriptions. Sellers adapted their content to be more parseable by AI, not more persuasive to humans.
This is mecha-nudging in the wild: the choice architecture that matters increasingly is the one facing the machine, not the person. As AI intermediates more decisions (product recommendations, search results, content ranking), optimizing for machine readability becomes the dominant strategy. The nudge literature's assumption that the relevant decision-maker is human is quietly becoming wrong.
The through-claim: choice architecture is migrating from human-facing to machine-facing, and the migration is already measurable. The nudge that matters isn't how you present options to people — it's how you present options to the AI that presents options to people.