Supervised Intelligence Model (SIM)
A five-stage working method for professional work with generative AI in which three stages are reserved to the human alone (Legal Framing, Verification, Strategic Judgment) and two admit the machine (Pattern Expansion, Doctrinal Reconstruction), so that what leaves the professional's hands is human judgment with machine assistance inside it, never the reverse.
When this decides something for you
A junior has drafted a written submission with a language model and it reads well. A partner has to decide whether it may be filed. A judge has a summary of the pleadings prepared by a system and has to decide how far to rely on it. A compliance officer has a policy draft the model wrote and has to decide what a signature on it means. In each case the question is the same: what did the human do, and was it enough? The Supervised Intelligence Model is the answer stated as a method: five stages, three of them the human's alone.
The question
Where, in the production of professional work, must the human act alone, and where may the machine act under the human?
The premise
Responsibility for professional work cannot be delegated to what cannot be held to account. A machine may produce text, find patterns and propose authorities. It cannot be a respondent before a court, a disciplinary committee or a client. So the method must place the acts that create responsibility (deciding what the question is, confirming what is true, deciding what to do) with the person who will answer for them, and may place the rest with the machine.
How it works
The method has five stages, in order. The names are fixed.
1. Legal Framing (human only). The professional states the question, the thesis, the structure and the settled formulations before the machine is used. If no framing exists, the machine is not used to invent one; the framing is built first. A draft produced without this stage has no author.
2. Pattern Expansion (AI assisted). Within the framing, the machine may expand, connect and develop: find lines of authority to be checked, propose structures, draft passages, surface counter-arguments. Everything it produces is provisional.
3. Doctrinal Reconstruction (human dominant). The machine may draft full prose from the framing; the professional reworks it into an argument that is theirs. The test for where Stage 2 ends and Stage 3 begins is whether the passage expands a frame the human set or supplies a judgment the human has not yet made; the second is not permitted until the human has made it.
4. Verification (human only). Every citation, quotation, paragraph number, clause, date and figure is checked by the professional against the primary record: the court's own copy of the judgment, the gazette text of the statute, the archived original of the document. The machine never fills these from memory. Until checked, each is a visible placeholder, not a fact.
5. Strategic Judgment (human only). What to include, how strongly to put it, what to disclose, what risk to run, whether to file: these are decided by the professional and by no one else. The machine may recommend and must mark the decision as pending.
Three consequences follow. First, the method is a sequence, not a checklist: Stage 4 cannot be done before Stage 2 has produced something to verify, and Stage 5 cannot be done before Stage 4 has settled what is true. Second, machine assistance is legitimate only at Stages 2 and 3 and only inside a Stage 1 frame; a draft that arrives with its thesis supplied by the machine fails at Stage 1 whatever happens afterwards. Third, the human-only stages are where responsibility attaches, so a record of them (what was framed, what was verified against what, what was decided) is the record a tribunal will later ask for.
One case
Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd, 2026 INSC 668, Supreme Court of India, 2 July 2026. The tribunal decisions under appeal had relied on precedents that the respondent's own affidavit said were not cited by counsel but "obtained through its own research"; the Court set them aside, describing the result as "no decision at all" and a "subversion of the rule of law", and directed the Bar Council of India, "being the apex statutory body, to constitute a committee and deliberate... and prescribe a guiding principle to prevent such occurrences, along with the disciplinary action that will follow a violation of the norms". Read against the five stages, the failure was not that a machine was used. It was that Stage 4 did not happen: material entered a judicial decision without a human verifying it against the record. The Court's direction to the regulator is a direction to promulgate a method standard, which is what the five stages are.
The rule
The machine works inside the human's frame and under the human's verification; it never supplies the frame, the verification or the decision. A professional who can show the three human-only stages were performed has supervised; one who cannot has delegated.
Not to be confused with
Human oversight, as the term is used in the EU Artificial Intelligence Act and in most guidance, is a duty on the provider or deployer of a system to make oversight possible. SIM is a method for the professional who signs: it says what the human must do, in what order, and what record results. Human-in-the-loop names a design in which a person can intervene; SIM requires that the person framed, verified and decided, which intervention alone does not establish.
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: Activating the Missing Accountability Layer of the Solidarity Stack
- SSRN 7454698 (2026)
- section 2
Changelog
- 2026-09-25, version 1. First published.