2.1.2. Application of Preferences

Cette page resume la section correspondante du RFC 9969 et conserve les termes, dates et references RFC essentiels.

When data trains an LLM, the resulting model cannot selectively use only part of that data for a task, because inference uses the whole model and cannot identify specific input data for selective use.

As a result, preferences visible at crawl time are generally not available at inference time. Preferences such as "no military uses" or "non-commercial only" cannot be applied by a general-purpose "foundation" model.

AI vendors that want to comply face unattractive choices: omit such data from foundation models, or create separate models for each permutation of preferences. The issue is compounded because preferences change over time, while LLMs are created over long periods and cannot easily be updated to reflect those changes. Participants described this as making opt-out regimes "stickier".