AI Responsibility Test (ART)
A four-step test applied after the fact to a piece of AI-assisted professional work to decide whether the professional supervised or delegated, asking in turn whether the framing was diligent, the supervision adequate, the verification complete and the judgment independent.
When this decides something for you
A filing has gone wrong. An authority in it does not exist, or a figure is wrong, or a clause was misquoted, and the machine was involved. A disciplinary committee, a Bench, a client or the professional's own senior now has to decide what the professional is answerable for. "AI was used" decides nothing, because AI is used in most professional work now. The AI Responsibility Test is the instrument that decides it: four questions, in order, each answered from the record of what the professional did.
The question
After AI-assisted work has been produced, how is it determined whether the professional who signed it supervised the machine or delegated to it?
The premise
Responsibility is conditioned on office, not authorship. The person who signed owns the work whoever typed it. The test does not ask whether the machine wrote it; it asks whether the person who owns it did the four things that make ownership real.
How it works
The test has four steps. They are applied in sequence, and the answer to each is taken from the record, not from the professional's account of their intentions.
1. Framing Diligence. Did the professional set the question, the thesis and the structure before the machine was used, and does the work show it? A submission whose organising idea can only have come from the machine fails here. Evidence: the instructions given, the outline, the settled formulations the machine was working within.
2. Supervisory Adequacy. Was the machine's contribution confined to the stages where it is permitted, and was it read by the professional as a draft rather than adopted as a result? A passage taken over unread fails here even if it happens to be correct. Evidence: the marks of rework, the changes between the machine's draft and the filed text, the professional's notes.
3. Verification Completeness. Was every citation, quotation, paragraph number, date and figure checked against the primary record before the work was relied on? One unchecked authority fails this step whatever else was checked, because the duty is complete verification, not sampling. Evidence: the ledger of what was checked against what, and the archived copies.
4. Judgment Independence. Were the decisions that belong to the professional (what to include, how strongly to put it, what to disclose, whether to file) made by the professional, with the machine's recommendations treated as recommendations? A decision the machine made by default, because no one overruled it, fails here. Evidence: the recorded decisions and their reasons.
Two features matter in use. The test is post hoc: it is applied to finished work, by someone other than its author if need be, and it does not tell the professional how to work. The Supervised Intelligence Model does that; the test is what the model's stages look like from the outside, afterwards. And the test is conjunctive: work passes only if it passes all four steps. Passing three and failing the fourth is delegation.
One case
Gummadi Usha Rani v. Sure Mallikarjuna Rao, SLP(C) No.7575 of 2026, Supreme Court of India, order dated 27 February 2026. The trial court's order of 19 August 2025 relied on four decisions that did not exist; the High Court found them to be generated by artificial intelligence and, after a word of caution, affirmed the order on the merits. The Supreme Court declared that a decision based on such non-existent judgments "is not an error in the decision making. It would be a misconduct and legal consequence shall follow." Applying the test: the order fails Verification Completeness on its face, since four authorities entered a judicial order unchecked, and the Court's characterisation of the result as misconduct rather than error is the characterisation the test produces, because the failure lies in what the person holding the office did not do, not in the quality of the outcome.
The rule
AI-assisted work is supervised if, and only if, the framing was the professional's, the machine's contribution was confined and reworked, every source-dependent claim was verified against the primary record, and the decisions were the professional's. Otherwise it was delegated, and the consequence attaches to the person who signed.
Not to be confused with
The test is not a workflow and is not a substitute for the Supervised Intelligence Model; it is applied to work after the fact. It is not a test of the machine's accuracy: a machine that happens to be right does not pass the professional. And it is not a disclosure rule: whether AI use was disclosed is a separate question, answered by the instrument that governs the filing, though a professional who cannot answer the four steps will usually also be unable to disclose accurately.
Provenance and method.
Where this was published
- Shivam Shukla
- AI for Indian Advocates: The Practitioner's Standard for Supervised Intelligence (2026)
- Part III
- The Method
- Shivam Shukla
- The Power to Say No
- SSRN 7454698 (2026)
- section 2
Changelog
- 2026-09-25, version 1. First published.