How AI Became the Backlog It Was Supposed to Fix

In February 2026, India's Chief Justice Surya Kant had a clear view of AI in the courtroom. After lawyers submitted petitions containing fictitious case law — including a non-existent case called "Mercy vs Mankind" — he condemned AI-assisted legal drafting in terms that left no room for ambiguity. ^[1]

We have been alarmingly told that some lawyers have started using AI for drafting. It is absolutely uncalled for. — Surya Kant

By March, he had launched a Supreme Court AI committee. By May, he was calling for an always-on automated judiciary and arguing that only technology could solve the courts' time problem. ^[2]

We should think of deepening technology and AI-based judicial architecture. — Surya Kant

In July, his own court set aside judgments from India's National Company Law Tribunal that had cited hallucinated legal precedents — decisions the Supreme Court described as legally void and an undermining of the rule of law. ^[3]

It is necessary for Courts to adopt a zero-tolerance mode for producing, citing or using AI-generated precedents without verification. — Supreme Court of India

Three stances, five months, one person. The contradiction is not Surya Kant's alone. The same arc is playing out across jurisdictions. Courts in at least five countries — China, India, the United Kingdom, the United States, and Canada — are making AI their answer to case backlogs that have grown unmanageable. India's Supreme Court alone carries 96,045 pending cases, with civil matters dating to 1986 and criminal matters from 1991; 75,402 new cases arrived in 2025, outpacing the rate at which the court disposes of them. ^[4] The UK's Crown Courts face a record backlog of more than 80,000 criminal cases across England and Wales. ^[5] In China, a pilot in Shenzhen boosted judge throughput from 495 to 744 cases per year — a 50 percent gain — and the Supreme People's Court has since expanded the system to 23 courts across 11 provinces. ^[6] The backlog-to-AI logic is shared, but the guardrails differ in timing and scope. China confined its system to research and drafting assistance from the start, with explicit guidelines that AI cannot issue final rulings or replace judges. ^[6] The UK's program, announced in June, scopes AI to paralegal functions — research, document summarization, grouping similar hearings — with no adjudication role. The Law Society issued its own caution about substituting technology for funding and staffing. ^[5]

While new technology should enhance access to justice, it cannot replace vital funding and additional court staff. — Ian Jeffery

In the United States, India, and Canada, the guardrails arrived after the line had already been crossed. In October 2025, two U.S. federal district judges admitted that staff had used ChatGPT and Perplexity to draft court orders containing significant factual errors that were erroneously docketed; one judge subsequently banned generative AI entirely from his chambers. ^[7]

My chamber’s policy prohibits the use of GenAI in the legal research for, or drafting of, opinions or orders. — Julien Neals

In India, the Supreme Court declared in March 2026 that AI-generated fake legal citations had become rampant in courts worldwide. ^[8] The court released draft AI regulations in June establishing human primacy — barring AI from independently adjudicating cases, passing sentences, or conducting risk-scoring for bail and recidivism. ^[9] By then, the encroachment was already documented: the National Company Law Tribunal had issued judgments citing hallucinated case law, and the Supreme Court was forced to set them aside. ^[3] In Canada, Chief Justice Richard Wagner warned in June that AI-generated hallucinated case citations had already appeared in Canadian legal proceedings. He drew a firm line: AI may summarize documents but must never be used to reach a legal conclusion. ^[10]

It should never be used in the analysis of the facts and the law with a view to obtain a judgment. — Richard Wagner

The pattern in these three jurisdictions is the same: adoption under backlog pressure, encroachment into adjudication, and a reactive guardrail written after the breach. That the UK and China built narrower limits from the outset does not resolve the larger problem — because even preemptive guardrails in two countries have not prevented encroachment in the others, and the errors now travel across borders through the same AI tools. When the Indian Supreme Court set aside the NCLT judgments in July, it used language that captured the nature of the threat. Fake AI precedents, the court said, were like a toxic gas leak in the legal system — invisible, insidious, and catastrophic. ^[3]

A decision of a Court or an adjudicating authority based on material which is fake and hallucinated is no decision at all, and it amounts to subversion of the rule of law. — Supreme Court of India

The comparison was not rhetorical excess. The court was describing a new category of judicial error — one that AI was supposed to eliminate, not produce. AI is not merely failing to save time. It is actively creating new judicial burdens. In Oregon, the Court of Appeals' Chief Judge Erin Lagesen ordered staff to begin tracking the hours spent on AI-fabricated filings, after sanctions in one federal case reached $110,000. ^[11]

In addition, to get a concrete sense of how much time the submission of fabricated authority likely produced by generative artificial intelligence is syphoning from the Court of Appeals’ core work of deciding cases, I have directed our staff and judges to track the time spent addressing fabricated authority so that Oregonians can have an accounting of the resources consumed in the event the situation does not resolve promptly. — Erin C. Lagesen

Separately, AI-generated deepfake evidence has begun appearing in court submissions. After a California court caught a deepfake video, security experts warned that newer AI models have eliminated the visual flaws that once allowed detection. One expert told judges the problem had almost certainly spread beyond the cases they had caught, and predicted a sharp rise in litigation volume. ^[12]

The system doesn’t have a way to test for authenticity; it’s almost a certainty that it has already happened elsewhere — Byron Long

These are not edge cases. They are a new category of work — fabricated filings to investigate, hallucinated precedents to reverse, deepfake evidence to authenticate — that consumes the very hours AI was deployed to reclaim. The throughput fix is expanding the throughput problem. Every human-primacy guarantee in this story — India's draft regulations, China's guidelines, Canada's warning, the U.S. judicial conference response — was written either after AI had already crossed the line it draws, or in anticipation of a crossing that had already occurred elsewhere. The operational strategy to collapse backlogs is, in practice, generating a new species of backlog, and the guardrails are being laid down on ground the technology has already passed over. On July 31, Surya Kant addressed a Commonwealth mediation conference and returned to the language of limits, drawing a distinction between what AI can mimic and what only a human judge can offer. ^[13]

AI may carry a human voice, but it cannot display human empathy or genuine apology. — Surya Kant

The line was still being drawn — even as the encroachment it guarded against had already happened, and was now producing the very backlog it was meant to cure.