when-prior-ai-decisions-become-relevant
1 min read 6 views
When Prior AI Decisions Become Relevant
An AI system does not need to be the formal decision-maker to affect a legal controversy. It may select facts, summarise evidence, rank authorities or recommend a remedy. If those operations repeatedly react to irrelevant characteristics, the question is not merely whether the model is “biased”. The practical question is whether reliance on that particular deployment should continue.
I propose treating this as a functional impartiality challenge. This is not a claim that existing judicial-recusal rules automatically apply to language models. It is a structured way to ask whether an AI evaluator or decision-support system remains suitable for the task assigned to it.
Prior outputs can be evidence, but they are not precedents. A single failure may be serious without proving a stable pattern. Comparability matters: model version, system instructions, source bundle, memory state, language, date and procedural posture must be preserved or reconstructed.
The audit must also look beyond the final recommendation. Distortion may arise when the system selects facts, assigns credibility, retrieves authorities, frames the issue or changes the remedy. If its earlier summaries later enter memory or retrieval systems, an initial distortion may return as apparently independent support.
A proportionate response can include record preservation, controlled counterfactual testing, independent evaluation, exclusion of challenged outputs from memory, an alternate system or human review de novo. High-stakes reliance should stop when the system’s role is not inspectable or genuinely contestable.
This is a research framework, not a declaration that a universal doctrine of algorithmic recusal already exists. It is intended to support empirical validation and institutional collaboration.