Between Assistance and Adjudication: Defining the Role of AI in Judicial Decision-Making

[Keshav and Rajbeer are students at Gujarat National Law University.]

On 2 July 2026, the Supreme Court of India overruled the decision of the National Company Law Tribunal (NCLT) and the National Company Law Appellate Tribunal (NCLAT) in relation to a Section 7 matter under the Insolvency and Bankruptcy Code 2016 (IBC), on the basis of the consideration of judicial precedents that did not exist and were created by artificial intelligence (AI). This piece uses the standard court articulated as a benchmark against which to test the Supreme Court’s own draft Regulations for the Use of Artificial Intelligence in Courts 2026 and to set out the amendments the draft still needs.

The Standard the Judgment Sets

The judgment lays down three clear rules. First, AI can help with judicial research, but it cannot replace the judge’s own reasoning. Second, checking AI- generated material is not an option, there should be a zero tolerance rule for it. Third, if any fabricated material becomes part of a judicial order that order becomes invalid, even if it did not change the final outcome. The draft regulations published recently deal with the same issues, but not always as strictly.

Gaps in the Draft Regulations and the Amendments Needed

The court makes it clear that AI should help judges in their decisions, but it should never be a replacement for them. Although the court is willing to adopt the use of AI in the judiciary, the court has made it very clear that the “adjudication shall be totally and absolutely under the control of the human judge and there shall be human in the loop”. Regulation 4 of the draft regulations reinforces this stance and makes it clear that AI should not override human judgement.

However, this assertion is harder to sustain in practice. Even though the NCLT’s order was made by a judicial officer, it rested on 6 fictitious or distorted citations introduced through the Tribunal’s own research, and the all-human NCLAT reiterated the same citations without any independent verification. This shows that there is a risk in “human in the loop” that does not verify and not in AI displacing the judge. Regulation 4 should therefore be amended to pair its declaration of human control with a mandatory, auditable verification step, so that “human in the loop” is a demonstrable practice rather than an assumption.

Verification as a legal duty

This judgment treats verification not only as a good practice but also as a legal duty binding on both bar and bench. Advocates must not cite unverified authorities, and judges must independently verify the material they rely on; failure on either side attracts what the court calls a policy of “zero tolerance”.

The draft regulations recognise the same concern, but qualify it. Regulation 8 of the draft requires judicial officers to exercise reasonable care before relying on AI-generated material and holds them accountable for resulting errors, yet it permits verification to be skipped where reasons are recorded in writing. In some cases, it also presumes that approved AI tools have already satisfied the requirement. Regulation 19 of the draft likewise allows the use of AI-based legal research and citation checking, provided always that they have first obtained the relevant approvals and supervision.

There is a clear discrepancy between the two frameworks. While the court permits no exception to the duty of verification, the regulations create one through a written-reasons carve-out and an unexplained presumption that “approved” tools have already undertaken the requisite verification. Neither provision specifies what constitutes an adequate recorded reason or the standards a tool must satisfy to qualify as “approved”. This creates a wide, self-certifying gap in a duty that the judgment treats as absolute. The regulations should therefore either remove this exception altogether for adjudicatory use or replace it with an objective approval regime incorporating defined accuracy benchmarks, periodic re-certification, and a public register of approved tools. The reliance on AI tools should be based on demonstrable evidence of reliability rather than merely written note in the record.

AI hallucinations and the validity of judicial decisions

The draft regulations define an AI “hallucination” as output that is fabricated, misleading, or unsupported by reliable sources; it includes non-existent or misrepresented judicial precedents which is the same problem the court identifies. Their solution, however, is narrower. Regulation 43(6) is directed primarily at situations where a party or advocate places fabricated AI-generated material before the court; it does not clearly extend to a case where the fabrication originates within a court’s or tribunal’s own research and which is then carried forward unverified.

Here, the court takes a wider stance and chooses to do so deliberately; it draws no differentiation between fabricated authority presented by the lawyer, the litigant, or the court. As soon as the fabricated authority makes its way into the decision-making process, the decision becomes unlawful. The regulations, in turn, mostly deal with the use of AI technology by litigants and courts’ users, thereby failing to tackle the use of AI in court processes properly.

Regulation 43 (6) should be amended to state expressly that fabricated material vitiates a decision regardless of its source, with accountability extended to judicial and tribunal staff on the same footing as advocates and litigants.

No fast-track mechanism for time-sensitive regimes

The impact of such contamination does not stop there as far as doctrine is concerned. The consequences of such contamination are not only doctrinal. Where a hallucination is discovered late in the process, the only route back is a full remand, precisely the delay a time-bound regime like the IBC, built around quick resolution and preserving asset value, can least afford. Such a delay becomes all the more significant since IBC itself relies on quick resolution and maximisation of the value of assets. Such delays can lead to the exact thing that the code sought to avoid, i.e., a loss in asset value. Hence, it needs to be understood that the problem of AI hallucinations in the courts does not involve only technical difficulties but can delay insolvency proceedings. The regulations should be amended to include a summary recall or correction mechanism for orders shown to rest on fabricated AI-generated material, so fixing the error does not itself become as costly as the error.

Conclusion

The lasting significance of the present case lies not merely in setting aside the two flawed insolvency orders, but in establishing important principles for the responsible use of AI in adjudication. When the judgment is read alongside Supreme Court’s draft regulations, it exposes both  the strengths and gaps within the proposed Regulations. While the court insists on strict human oversight and verification, the regulations leave open a written-reasons exception to that duty, a narrow reading of what counts as a disqualifying hallucination, and no dedicated mechanism for correcting AI-tainted orders quickly. Closing these gaps through the amendments to Regulations 4,8,19 and 43 (6) discussed above, is what would make the standard the court has set enforceable in practice, rather than aspirational in principle. As AI becomes increasingly integrated into the justice system, the regulatory framework must evolve to safeguard the accuracy, accountability, and public confidence in judicial decision-making.