Four scholars at the University of Minnesota: Nick Bednar, David Cleveland, Allan Erbsen and Daniel Schwarcz, decided to examine, under controlled conditions, issues pertaining to use of AI around the premise: whether leaning on AI early reduces cognitive engagement with the source material? And, does it erode the understanding needed for independent reasoning?

The authors, in their study, Artificial Intelligence and Human Legal Reasoning^[1], put about a hundred senior law students through four tasks in sequence, namely (i) synthesis of unfamiliar materials into a memo, (ii) closed-book comprehension test, (iii) application of law to a fresh set of facts, and (iv) finally, revision of the application memo. In the study, one group used AI for the opening synthesis and while the other did not. Neither group could use AI for the comprehension or application stages but both were permitted use for revision.

Initially, AI produced a gain far larger than earlier studies had recorded and the assisted synthesis memos were not merely quicker but substantially better. There was an improvement of roughly 50 to 70 per cent and the average student working without AI sat in the 25th percentile of quality while the average student with AI sat in the 71st percentile. The AI improved form and substance at once.

Contrary to fears, the gain did not come at the cost of understanding. When AI was taken away for the comprehension test, the two groups were indistinguishable: 3.86 correct answers against 3.88. More strikingly, the AI-exposed group also performed better on the later application task, even after AI had been taken away. But that finding comes with an important qualification. As Daniel Schwarcz has explained, once the researchers controlled for the quality of the students’ initial synthesis, the apparent advantage of prior AI use disappeared. In other words, AI did not seem to give students some lasting, independent boost in legal reasoning. Rather, it helped them produce a stronger synthesis at the outset and, with it, a better mental model of the material. That stronger foundation then carried forward when they were required to reason on their own.

At the revision stage everyone used AI and the effect was split by ability. Weak memos improved; whereas strong memos got worse. The lowest scorers gained nearly two points and the highest memos lost as many as eight points. For someone who had already produced something sophisticated, it offered a smoother, more confident, intellectually poorer alternative, and the alternative was accepted. AI compressed the range, thus, lifting the floor while lowering the ceiling. The real risk, then, is not using AI but ceasing to question it!

Fatigue and time were likely part of the explanation. By the revision stage, participants had already spent nearly three hours on demanding work and had only twenty minutes left. Those are exactly the conditions in which people may stop interrogating an answer and begin deferring to it.

The point, then, is not simply that AI helps or harms. It can do both. What matters is whether the person using it continues to think independently. The same tool that can scaffold a weaker lawyer’s reasoning may quietly weaken the judgment of a stronger one.

In this supervised and closed task, AI did not hallucinate more than the humans. The outcome is that verification works when someone performs it. But the important thing is that it only works when someone does it and not when someone does not.

On 2 July 2026, in Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd., 2026 SCC OnLine SC 1258, decided on 2 July 2026, the Supreme Court confronted an insolvency admission built on six authorities of which three did not exist. The Bank confirmed its counsel had never cited them and that they were the tribunal’s own research and had survived the first appeal. The Court set aside the order of the appellate tribunal and deprecated in the harshest of terms. Earlier, in Gummadi Usha Rani v. Sure Mallikarjuna Rao C.R.P. No. 2487 of 2025, decided on 21 January 2026, the Andhra Pradesh High Court noticed that four authorities relied upon by the Trial Court were AI-generated and non-existent, but declined to interfere because it considered the underlying legal reasoning sound. When the matter reached the Supreme Court in Gummadi Usha Rani v. Sure Mallikarjuna Rao, 2026 SCC OnLine SC 341, however, the Court treated the issue as one concerning the integrity of the adjudicatory process itself. It restrained the Trial Court from proceeding on the basis of the Advocate Commissioner’s Report and observed that a decision founded on non-existent and fake judgments was not merely an error in decision-making, but raised questions of misconduct and legal consequence.

Cognitive dependence leaves no such trace. A lawyer can receive an answer in which every citation is genuine and still miss that a holding has been stretched, a distinction quietly dropped, an inconvenient authority omitted, an argument framed in the wrong doctrinal terms. The work reads fluently from beginning to end. Nothing in it announces the defect. That is the failure the Minnesota study points to, and no cite-checker will catch it.

Law is not an information-retrieval profession but rather a judgment based profession. The hardest questions do not sit in a database. A lawyer has to distinguish a precedent, to name the assumption everyone else has missed, to see which technically available argument should not be run, to anticipate how a particular judge will receive it, to weigh a client’s commercial objective, and sometimes to recognise that strategically the strongest legal position is the wrong one. Lawyering is prejudging, reading, strategising in advance, how a matter will unfold and how the people in it will behave.

The skill of a lawyer is built the slow way, preparing for situations which are essentially “out of syllabus”. A young lawyer becomes a good one by reading the difficult judgment, not its summary, learning where ratio ends and obiter begins, how stare decisis constrains, when res judicata bites.

What does it mean for a decision to be sub silentio or per incuriam, and when can either principle be invoked? The same is true of jurisdiction.

Knowing the law is only the first step. The harder skill is knowing what matters, what does not, and how a rule should be applied when the facts refuse to fit neatly within it. That instinct is built over time: through reading judgments closely, arguing both sides, making mistakes and learning why an argument failed. The Minnesota study is important precisely because its findings are more complicated than the familiar warning that AI simply makes us intellectually lazy. Students who used AI at the synthesis stage did not understand the material any less, and they later performed well even when the tool was taken away. But the revision exercise exposed a different danger. When AI was brought back, weaker work often improved, while some of the stronger work became worse. The problem, in other words, may not be that AI prevents us from thinking. It may be that, once an answer looks polished enough, we become less willing to keep thinking.

That is why the real question for lawyers is not whether they use AI, but whether they remain in control of the reasoning. AI can help organise authorities, test a proposition or expose gaps in an argument. It should not be asked to carry the entire intellectual burden. A useful rule is to use it only for work you remain capable of checking yourself. If you could not defend the argument before a sceptical partner or judge without returning to the tool, the reasoning is not yet yours.

Indian courts are beginning to approach the issue in much the same way. In Pooja Ramesh Singh, the Supreme Court confronted the consequences of fake citations and invented passages finding their way into adjudication and called for disciplinary safeguards against the use of hallucinated material. The Court’s Draft Regulations for Use of Artificial Intelligence in Courts, 2026^[2] go further by proposing disclosure where pleadings are AI-assisted and allowing judges to ask what system was used and how its output was verified. Regulation 43 has already attracted criticism, particularly because mandatory disclosure may be difficult to define and police. But beneath that controversy lies a fairly simple idea. The use of AI may be new; professional responsibility is not. The lawyer remains answerable for every proposition placed before the court, whether it came from memory, a junior, a search engine or a machine.

This disclosure mandate has drawn fire as the Supreme Court Advocates-on-Record Association (SCAORA) calls the routine disclosure unworkable and unjustified as advocates are already answerable for every filing under the Advocates Act, 1961 and the Supreme Court Rules, 2013 and warns that flagging AI use could prejudice a litigant or colour how a judge reads the document. SCAORA proposes a narrower certificate: not that AI was used, but that every citation, precedent and authority relied upon has been personally verified for accuracy.^[3]

A verification certificate may catch the fabricated case, but it does nothing about the genuine case cited for a proposition it does not support, the holding stretched a shade too far, and the authority left out. You can verify every citation and still lose control of the reasoning. The disclosure debate asks whether AI was used. The more difficult question raised by the Minnesota study is whether the lawyer still understands and owns the argument being made.

A lawyer must still be able to defend the argument and explain the reasoning behind it. Fluency can disguise a lack of understanding, especially when AI makes an answer appear complete. Pooja Ramesh Singh shows the cost of that failure: fabricated authorities entered the judicial process, survived appeal, and ultimately forced the matter back for reconsideration. AI can shorten the path to an answer. It cannot assume responsibility for whether that answer is right.

*Varun Pathak, Partner, Shardul Amarchand Mangaldas.

Assisted by Rudraditya Singh Panwar

[1] Nicholas Bednar, David R. Cleveland, Allan Erbsen & Daniel Schwarcz, Artificial Intelligence and Human Legal Reasoning, Minnesota Legal Studies Research Paper 2026-21 (2026), SSRN Abstract No. 6525800, HERE .

[2] Supreme Court of India, Draft Regulations for Use of Artificial Intelligence (AI) in Courts, 2026, Regulation 43; official draft: HERE .

[3] Supreme Court Advocates-on-Record Association (SCAORA), Comments and Recommendations on the Draft Regulations for the Use of Artificial Intelligence (AI) in Courts, 2026, pp. 43-44 (15 July 2026): HERE.

Disclaimer: The views and opinions expressed in this article are those of the authors in their personal capacity and do not necessarily reflect the views, positions, or policies of the firm, its clients, or The Bar Bulletin. This article is intended solely for informational and academic discussion and should not be construed as legal advice. Readers are advised to seek independent professional advice before acting on any information contained herein.