The Supreme Court’s draft Regulations for Use of Artificial Intelligence in Courts, 2026 draws a necessary line. AI may assist legal research, drafting, translation and other court-facing tasks, but it must not decide cases, predict outcomes or replace judicial decision-making.

The next question is narrower and more operational: when AI-assisted material is filed in court, what record should accompany it?

Disclosure is a start; it is not proof. A lawyer may disclose that an AI tool was used to prepare a pleading. A party may disclose that a transcript, translation, image, video, audio clip or summary was generated, enhanced or analysed using AI. A registry may know that an AI system helped generate a notice, cause-list entry or communication. But disclosure alone does not answer the harder questions that arise if the output is challenged.

What was the source material? What was changed? Which tool was used? Was the output checked by a human? Were citations verified? Was the original file preserved? Does metadata exist? Can the final document be traced back to the material on which it is based?

That is why court AI governance needs a filing provenance record.

The problem is not hypothetical. Indian courts and legal professionals are already confronting AI-generated fake citations and non-existent authorities. The Bar Council of India’s position is that advocates cannot avoid responsibility for fake citations by blaming technology. The Supreme Court has also expressed concern over AI-generated non-existent judgments being relied upon in judicial work.

Those cases show the first category of risk: the AI output is simply false. But the next category is more subtle. AI may alter evidentiary character without obvious fabrication. A translation may change meaning. A summary may omit context. An enhanced image may make uncertain detail appear clear. A cleaned audio clip may remove noise in a way that affects interpretation. A transcript may silently convert ambiguity into certainty. A pleading may include AI-assisted legal propositions that appear well-structured but are unsupported by the record.

In each case, the Court’s concern should not stop at whether AI was used. The concern should be whether the use can be reconstructed.

A workable provenance record need not be complicated. For most filings, it could be compact. It should identify the nature of AI assistance, the category of tool used where known, the source document or material, the human reviewer, the verification step and whether the original file or source material has been preserved. Where electronic evidence is involved, courts may also need metadata, hash values, chain-of-custody information, tool logs or other integrity markers, depending on the dispute.

This does not mean every AI-assisted filing should become a mini-forensic proceeding. That would be disproportionate and would slow court process. The point is different: when authenticity, source, alteration history, evidentiary weight or authorship becomes material, the Court should not be left with only a checkbox disclosure.

Judicial discretion must remain central.

An AI-assisted record should not be automatically inadmissible merely because AI was used. That would be overbroad. A lawyer may use AI to improve formatting, prepare a first translation draft, organise a chronology, compare documents or test a legal argument. Some of these uses may be harmless or low-risk. Others may affect the substance of what is placed before the court.

The correct response is, therefore, risk-sensitive. Routine formatting help should not attract the same scrutiny as AI-enhanced audio, AI-generated images, AI-assisted translation of witness material, or AI-supported summaries of voluminous records. A court may need only a basic declaration in one case and a fuller integrity statement or expert examination in another.

This is where the draft regulations can be strengthened.

A binary distinction between permitted and prohibited AI uses is useful, but it does not fully capture the middle ground. There is a difference between an AI scheduling assistant and a tool that shapes the text of a judicial summary, registry communication or evidence-related filing. The latter should attract stronger records, logs and verification controls.

The same logic applies to AI-assisted judicial or registry content. If AI materially assists in generating a court notice, cause-list entry, translation, summary or registry communication, the institution should be able to verify later whether AI was used, what system was involved, what human review occurred and whether the final document is traceable. Without that record, auditability becomes theoretical.

A filing provenance record would also help advocates. It gives lawyers practical discipline before relying on AI-assisted output. Instead of asking only whether the tool produced something useful, the lawyer must ask: Can the source be produced? Can the citation be verified? Can the original document be shown? Can the court understand what AI changed and what the human reviewer accepted?

This is not anti-technology; it is basic litigation hygiene.

The filing record should be even stronger for AI-assisted or synthetic evidence. Images, audio, video, transcripts, translations, summaries and data extracts can now be generated or materially altered in ways that are not obvious to the eye. Where such material is relied upon to prove identity, conduct, intent, communication, location, authorship or chronology, source preservation matters.

Courts should be able to direct production of the original file, metadata, hash values, system logs, provenance records or supplementary AI-use disclosures where the issue is material. Where required, courts may also seek forensic examination or expert review.

The key is to avoid two weak extremes. One extreme is blind acceptance: AI use is disclosed, so the court moves on. That is insufficient. A disclosure that cannot be tested does little when authenticity is disputed.

The other extreme is blanket suspicion: AI was used, so the material is inherently unreliable. That is also wrong. Many AI-assisted uses may be routine, transparent and properly verified.

The better position is controlled reliance. AI-assisted filings should be allowed where appropriate, but their source, review and integrity should be capable of being shown when challenged.

The Supreme Court’s draft regulations already move the conversation beyond vague AI ethics. They recognise that AI used in courts must remain subordinate to judicial authority. The next step is to make that principle operational in filings, pleadings, registry communications and electronic evidence.

For Indian courts, the minimum practical question should be simple: If AI helped create or alter this material, can the person relying on it show what happened? If the answer is yes, the court has a basis to assess relevance, authenticity and evidentiary weight. If the answer is no, disclosure has become a label without a record.

AI may assist the legal system. It should not make the legal record harder to verify.

Abhishek G Sharma is Founder and CTO of EU AI Compass and Move78 International, a Hong Kong-based cybersecurity, AI governance, and enterprise risk advisory firm.