Judicial AI and the Irreparable Bias Problem
Bibliographic record
Abstract
Background: Courts are increasingly experimenting with large language models (LLMs) for tasks such as legal retrieval, drafting support, anonymisation, and triage. Yet the promise of efficiency collides with a structural problem: bias. Human adjudication already reflects cognitive and institutional biases; LLMs trained on past judgments and legal text inherit and sometimes amplify those biases. This article asks a focused question: If AI belongs in courts at all, what is the safe, lawful, and useful lane—especially with respect to bias? The inquiry is situated within fair-trial guarantees and emerging regulatory expectations. Methods: A staged analysis grounded in legal obligations and informed by relevant technical characteristics is employed. First, sources of human and judicial bias are mapped, along with points at which LLMs introduce or magnify bias. Second, hard- and soft-law guardrails relevant to bias control in the justice sector are synthesised. Third, two instructive case studies—COMPAS/Loomis (U.S.) and Ewert v. Canada—are examined to demonstrate how group-level disparities and model opacity can generate due-process risks and to identify remedies transferable to LLM-assisted workflows. Finally, an operational blueprint is derived and applied to identify low-risk, high-yield assistive uses for Ukraine. Results and conclusions: The analysis shows that fully impartial AI outputs are not attainable in adjudication; bias is ineliminable but can be bounded. For Ukraine, the rational path is to invest first in data curation, secure infrastructure, evaluation capacity, and procurement with audit rights, and to confine AI to retrieval, norm collation, drafting-hygiene checks, and “missed-norms” prompts. The contribution is a governance blueprint that ties specific LLM failure modes to enforceable legal duties and practical safeguards—offering courts a credible, bias-aware lane for AI that improves service while preserving rights.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.054 | 0.116 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.003 | 0.003 |
| Science and technology studies | 0.005 | 0.025 |
| Scholarly communication | 0.010 | 0.008 |
| Open science | 0.003 | 0.009 |
| Research integrity | 0.002 | 0.004 |
| Insufficient payload (model declined to judge) | 0.005 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".