ANCHOR Guidelines: A Constitutional Blueprint for Human-Centred Judicial AI in India
Keeping a human judge formally in charge does not by itself preserve judicial legitimacy; a governable judicial-AI regime must also make algorithmic influence, institutional operators, override incentives, and avenues for contestation independently auditable.
Review
Ray and Anand propose ANCHOR as a constitutional governance blueprint for judicial AI in India. The framework translates Articles 14 and 21, non-delegation, reason-giving, judicial review, and comparative AI-governance practice into five design commitments: Accountability, Non-delegation, Constitutional Conformity, Human Oversight, and Review. Its practical proposition is that AI may assist with research, summarisation, precedent discovery, pattern identification, and recommendations, but a human judge must retain final adjudicatory authority. Recommendations should be traceable to source material, departures should be reasoned, higher-risk uses should attract stricter liability, models should undergo bias testing, and litigants should retain routes for challenge and review.
The paper's contribution is institutional rather than technical. It treats judicial AI as an allocation of adjudicatory influence that must remain bounded by constitutional doctrine, rather than as a generic productivity tool. The most reusable move is the connection between the duty to give reasons and the requirement that AI-derived inputs be explainable and reviewable. The proposed pilot also improves the framework's testability by combining efficiency measures with judge-AI agreement, appeal and reversal rates, litigant trust, workload, consistency, and perceived fairness.
The central limitation is that ANCHOR secures formal human authority more clearly than it secures practical independence from algorithmic influence. A judge remains the legal decision-maker, yet the design asks judges to record reasons when departing from AI recommendations and proposes using those departures to recalibrate the system. That can create an institutional default around the model's recommendation, particularly if disagreement is administratively visible while agreement is frictionless. The paper does not specify how to prevent override logging from becoming an implicit conformity mechanism, who controls retraining decisions, what evidentiary threshold justifies recalibration, or how model changes are versioned and disclosed to litigants.
A related gap concerns oversight authority. ANCHOR proposes a dedicated monitoring team and an ethical review board, but does not fully define appointment, independence, removal, access powers, publication duties, procurement relationships, or responsibility when judges, vendors, court IT units, and oversight bodies disagree. Its appeal mechanism likewise needs a clearer procedural object: a litigant can contest an AI-related decision, but effective redress requires disclosure of where AI materially influenced the proceeding, preservation of model and data provenance, access to relevant logs, and a standard for determining whether an error was prejudicial.
The paper acknowledges these limits by presenting ANCHOR as a theoretical blueprint and calling for an 18-to-24-month pilot across three court settings. That appropriately bounds the claim. The framework therefore should be read as normative constitutional architecture awaiting institutional specification and empirical validation, not as a proven deployment model. Evidence that would materially strengthen it would include a pilot demonstrating that judges can override recommendations without conformity pressure, that affected parties can reconstruct and contest material AI influence, that oversight bodies can act independently of vendors and court administration, and that model updates do not silently change the evidentiary basis of adjudication.
ANCHOR matters because it moves judicial-AI governance away from abstract ethics and toward constitutional control points. Its next step should be to make those control points executable: define authority over deployment and retraining, preserve versioned audit evidence, specify disclosure and appeal rights, establish independence requirements for oversight, and test for automation bias and institutional deference rather than assuming that a human signature is sufficient proof of meaningful human judgment.
Key Insight
Keeping a human judge formally in charge does not by itself preserve judicial legitimacy; a governable judicial-AI regime must also make algorithmic influence, institutional operators, override incentives, and avenues for contestation independently auditable.