The course · 16 contact hours
Overview
The course
Computational Thinking and AI in Law is a single-credit, 16-hour course on using AI well in legal work, and on the Indian law that governs using it. It teaches the craft the tools now demand: briefing a model, engineering its window, building a grounded research workspace over the student's own authorities, splitting work across drafter, critic and verifier, and choosing model and effort per task. Everything is verified to source, with a trail.
The premise
The difference between AI that accelerates a lawyer and AI that ends a career is rarely the tool. It is what the lawyer puts in front of the model, how they ask, and what they check. Since July 2026 the Supreme Court has held that citing unverified AI-generated authority is professional misconduct, and that a decision resting on hallucinated material is no decision at all.
The format
Every session is half taught, half demonstrated. The first 60 minutes teach one technique and run it live on a real free-tier tool, as an experiment on the class or a build-along with the presenter. The remaining 30 minutes are the lab: the instructor takes one realistic student task and runs the technique on it end to end, with the room predicting each result before it lands. Students repeat the technique on a task of their own after the session and submit the deliverable on the course platform. Every lab completes on a free account.
Eligibility, fee and evaluation
Eligibility
Purpose-built for law students: every technique arrives on a task from the student's own year (a moot memorial, an internship research note, a judgment brief, a seminar paper, journal cite-checking), and the legal session covers their own regulatory position. No quantitative background or prior AI knowledge is assumed; the course builds on the legal training students already have.
Dates and fee
Dates and fee are confirmed for each cohort and announced before enrolment opens.
Evaluation
The course closes with a 60-minute evaluation session, taken live. Graded work rewards judgment, verification and process over raw model output.
Learning outcomes
Explain what a language model does, predict from the mechanism where it will fail on a legal task (fabrication, sycophancy, variance, anchoring), and place any confident output on the Ladder of Misinference before relying on it.
Brief a model like a junior: a specification that says everything once, an output contract that forces uncertainty into the open, and worked examples where form must match. Iterate one change at a time, starting fresh when a conversation degrades.
Engineer the context: supply the source instead of asking from recall, structure long documents for the window, demand and verify paragraph pinpoints, and manage what does not fit by excerpting, chunking or re-supplying, knowing what each choice loses.
Build and run a grounded research workspace over a real bundle of authorities (instruction set, pinpoint citations, trap-tested) and state its limits: retrieval is not verification, and a bundle is only as current as its date.
Split work across a drafter, a fresh-context critic and a verifier when independence or breadth pays, and choose the model and reasoning-effort level per task, diagnosing whether a failure is the prompt's ceiling, the context's, the tool's, or the model's.
Run the four-step authority check (exists, says that, still good law, and the step everyone skips), classify failures into the three classes, and keep a verification trail another person could repeat.
State where Indian law stands on AI use, by register: authentication under BSA s.63 and its case law, the synthetic-media rules, the judiciary's instruments from White Paper to draft Regulations, professional duty after the 2026 misconduct holding, data protection, and privilege. Apply it, including the academic-integrity rules that bind a student today.
Ship a verified work product: a real deliverable through the full pipeline, every claim verified to source, with the trail and an AI-use disclosure that would survive a judge, a partner or an examiner.
The curriculum
01 · Sessions 1–2
Direct the machine, decode it
Prompts that work (specification, worked examples, the output contract), then how a language model actually works and why it must sometimes fail, taught once, properly.
02 · Sessions 3–4
Engineer the context
What goes into the window and why: supply the judgment, structure long documents, keep the conversation clean. Then a grounded workspace over a bundle of authorities, built in ten minutes.
03 · Sessions 5–6
Orchestrate, then verify
Work split across a drafter, a fresh-context critic and a verifier, with model and reasoning effort chosen per task. Then the four-step authority check and the trail, practised in the labs.
04 · Sessions 7–8
The law, and the work product
Where the law stands, India first (evidence, professional duty, data protection, privilege), and the whole pipeline run end to end into a work product the student can sign.
| Block | Shape | Hours |
|---|---|---|
| Core sessions | 8 × 90 min (60 taught + 30 lab) | 12 |
| Conversations with Experts | 2 × 90 min | 3 |
| Evaluation | 1 × 60 min | 1 |
| Total | 16 |
The core sessions · 60 taught + 30 lab
One technique per session: 60 minutes of teaching and live demonstration, then a 30-minute lab where the technique runs end to end on one realistic student task. Open a session for the full plan.
Conversations with Experts
The closing four hours
Two 90-minute Conversations with Experts test the course's frameworks against real practice: in each, the instructors lead a structured conversation with a practitioner who uses AI seriously in their own work.
Conversation 01 · 90 min
Inside the firm — what AI actually does to legal work
What they use, and for what; what they refuse to use it for, and why; what is checked before anything leaves the building, and by whom; how juniors are supervised when drafts are machine-fast; and what they would want a first-year associate to do on day one.
Speaker: Ms. Aastha, Partner, Argus Partners (Solicitors & Advocates); a law-firm practitioner
Builds on Sessions 1, 3 and 6
Conversation 02 · 90 min
Building with AI in-house — harnesses, tooling, workflows
What they have built: the workspaces, pipelines and guardrails behind an in-house function; how reliability is evaluated before roll-out; data governance when the data is the company's and the clients'; buy versus build; and the in-house privilege gap the Supreme Court confirmed in 2025, as they live with it.
Speaker: Ms. Suman Mishra, Head of Commercial Contracts, Kotak Mahindra Bank; an in-house banking and technology lawyer
Builds on Sessions 4–5 and 7
Evaluation · 60 min
A 60-minute evaluation session closes the course, taken in class.
Two parts, taken on the course platform. An applied audit: a fresh extract seeded with the three classes of authority failure (one fabricated, one that says something materially different, one no longer good law), verified against supplied sources, with the trail. And short applied questions: an authentication note on a described digital artefact; a ceiling diagnosis on a described failure; and scenario questions on the legal position, marked for register: a holding is not an observation, and a draft is not a rule. The paper rewards demonstrated verification and judgment over recall.
The format
Every session is 90 minutes in two halves: something happens live, and everyone leaves with a useful skill. When the room must feel an effect first, the taught hour is a mirror: an experiment on the class through commit-then-reveal polling, then shown on a real model. When it teaches craft, it is a build-along: the class builds the prompt, the workspace or the pipeline with the presenter. The last 30 minutes are the lab: the instructor demonstrates the technique on one realistic student task, with the exact prompts on screen and the room predicting results by phone. Students repeat the technique on a task of their own after the session and submit the deliverable on the course platform before the next one.
The room commits an answer on their phones and sees itself fooled, or builds the prompt, the workspace or the pipeline live with the presenter.
The effect explained from how a language model works, so it can be predicted next time rather than memorised.
Proved live on a free-tier account, with designated moments where the same task runs on a stronger model beside it.
A practical technique, demonstrated in the lab and repeated by students on their own work.
The people
Instructors
Professor of Law, National Law University Odisha
Teaching at National Law University Odisha since 2010, with a Fulbright year at Harvard Law School (2019–20). He holds an LL.M. in Human Rights from the National Law School of India University and a Ph.D. in Law from NLU Odisha.
LinkedInFounder, FastPixelsAI
Builds machine-learning systems and the teams behind them: previously head of Generative AI at invideo, chief scientist for deep learning at VJ Imaging Technologies, and CTO of difference-engine.ai. He designed and taught the AI and machine-learning curricula for S P Jain Global and GreyAtom, and holds a BSc (Hons) in Mathematics and Computing from the Institute of Mathematics and Applications.
LinkedInIndustry experts
Partner, Argus Partners (Solicitors & Advocates)
Partner at Argus Partners in Mumbai, where she has practised for over sixteen years across banking and finance, general corporate work and debt restructuring. B.A. LL.B. (Hons) from the West Bengal National University of Juridical Sciences.
LinkedInHead of Commercial Contracts, Kotak Mahindra Bank
An in-house banking and technology lawyer for over fourteen years, across RBS, IndusInd Bank, First Rand Bank, IDFC FIRST Bank and the Kotak group, working on commercial contracts, data privacy and fintech. LL.M. in IPR and Technology Law from UC Berkeley School of Law, with Dean's List honours; CIPP/E certified.
LinkedInThe labs
Each lab is a demonstration: the instructor takes one realistic task of the kind students actually have (a moot issue, a judgment, a bundle of authorities) and runs the technique on it end to end, with the room predicting results by phone. Students repeat the technique on a task of their own after the session and submit the deliverable on the course platform. Every lab completes on a free account. Open a lab for the full plan.
The standing rule
In every lab students must use AI, and they must disclose which tools they used and verify every citation and proposition before relying on it. The model's confidence is not evidence.
Free tiers first
Every lab completes on a free account: ChatGPT or Gemini, Google NotebookLM, Indian Kanoon. Nothing needs a paid plan, a card, or a tool the institution might not have; under-18 students are routed to the services that admit them. Paid tiers appear only as instructor-run contrast in three designated comparisons, one of which shows the free tier fully adequate. Free-tier caps, model names and features move: each carries the date it was checked and is re-verified every cohort. And because 30–60 students hitting a free service at once trips limits, every lab has a stagger plan, short prompts, and a stated fallback for anyone locked out.
Course teamTeaching this? The instructor lab bench (course-team sign-in) runs every lab live: playground, grading bench and RAG explorer, preset per session.
Assessment
AI use is required in this course, and disclosing and verifying it is required too. The marks follow that discipline: verification to source, sound craft, awareness of bias and sycophancy, ethical handling, and a reflective account of process.
The scheme
Three components, one submission platform. Lab deliverables run throughout and are completion-checked automatically; the audit and the evaluation are marked against seeded answer keys, so most of the marking is mechanical.
The signature assessment · session 8
40%
The signature assignment: a seeded memo verified to source, every citation through the four steps, all three failure classes caught and classified, with a repeatable trail and the reconstructed disclosure. Marked on the platform against the seeding key; the trail and disclosure get an automated first read, and the instructor reviews flagged work.
| Component | Weight |
|---|---|
Lab deliverables Throughout Each session's lab demonstrates a technique on one realistic task; students repeat it on a task of their own after the session and submit the deliverable on the platform before the next one. Scoring is by completion: an automated check confirms the deliverable is present and matches the lab's checklist. There is no separate quality mark. | 30% |
Evaluation session Closing session The in-class 60-minute paper, taken on the course platform: an applied audit against a fresh seeded extract, marked against its key, plus short structured questions (an authentication note, a ceiling diagnosis, register questions) with model answers. | 30% |
Three components, one submission platform. Lab deliverables are completion-checked automatically, and the audit and the evaluation are marked against seeded answer keys, so most of the marking is mechanical and the instructor reviews only flagged work. Each graded component carries an AI-use disclosure: in this course AI use is required, and disclosing and verifying it is required too. One warning: this required-use rule is this course's scheme only. A student's other subjects follow the university's regulations (at least one NLU treats AI use in assessed projects as serious academic malpractice), and moots and journals have their own rules. Checking the governing rule before relying on the tool is part of what the course certifies.
The assessed pieces
The labs feed the assessed submissions; the in-class evaluation closes the course. AI use is required, and disclosing and verifying it is what is assessed. The required-use rule is this course's own; other subjects, moots and journals keep their own rules, and students are taught to check the governing rule first.
Built from Every session's lab · outcomes 1, 2, 3, 4, 5, 6, 7, 8
The eight lab deliverables (the saved prompt, the judgment brief, the grounded workspace, the critic pass, the verification trail, the authentication note, the audit calibration run), each produced by repeating the demonstrated technique on a task of the student's own after the session, and submitted on the course platform before the next one. The platform checks each submission for completeness against the lab's checklist; there is no separate quality mark, and the artefacts remain the student's own working toolkit.
How it's graded
Academic integrity
Each submission names the tools and tiers used. The deliverables are working artefacts from the student's own tasks; no client, confidential or personal material in any of them.
Built from Session 8's calibration run, at full scale on a fresh memo · outcomes 1, 6, 8
The signature assignment: a fresh AI-drafted legal memo, seeded with all three failure classes (a fabricated authority, a real authority materially misdescribed, and an authority that is no longer good law), audited to source. Every citation through the four steps, every proposition on the Ladder, a complete trail, each seeded error identified and classified with the corrected position stated, and the AI-use disclosure the memo's author should have filed, written from the trail. Submitted on the course platform and marked against the seeding key: errors caught, classes assigned, positions corrected. The trail and the disclosure receive an automated first read; the instructor reviews what it flags.
How it's graded
Academic integrity
The audit itself is database-and-judgment work. AI may be used to organise it, disclosed as always, but a model's opinion of a citation is not a check. Submitting an audit that itself contains an unverified or invented authority is treated as the failure this course certifies against.
What the rubric rewards
Every citation and proposition checked against a grounded tool or citator. The duty to check is what the marks reward first; skipping it is the difference between using AI well and a cost order.
A specification with an output contract and worked examples; the right context, supplied and structured; the right model at the right effort, shown through deliberate iteration and a diagnosis of whose ceiling a failure was.
Steel-manning the opposing side, resisting a model that merely confirms the user, and naming the limits of an output instead of presenting unverified authority as fact.
Confidentiality, candour to the tribunal, and clear disclosure of AI use, kept in the work and in the write-up throughout.
An honest account of process: which tools were used, what was checked, what was rejected, and how the student sits as the human in the loop who owns the result.
Integrity & AI use
AI use is central to every graded assignment. Students must state which models and tools they used and submit an AI-use disclosure with each submission.
Every legal claim in an output is checked to source before it is relied on. No unverified or fabricated authority may appear in submitted work; the merely plausible is flagged as such.
No privileged or confidential material may be entered into public LLMs. The confidentiality line is carried through every lab and the audit.
How it is marked
Shipping an unverified or invented authority is the failure the course certifies against: Mata v. Avianca, with marks attached. The disclosure and the verification trail are part of the deliverable.
The institutions
National Law University Odisha, Cuttack, established by an Act of the Odisha Legislature, is the programme's academic host: the course runs for its students and under its regulations, with credit through its own academic-council approval. Course instructor Dr. Rangin Pallav Tripathy is Professor of Law at the university.
nluo.ac.inFastPixels AI is the AI company founded by Mr. Soumendra Dhanee, whose work spans the generative-model team at invideo, industrial deep-learning systems at VJ Imaging Technologies, and AI curricula for S P Jain Global and GreyAtom. For this programme, FastPixels designed the course and built what it runs on: the companion site, the live polling app used in every session, the instructor's lab bench, and the platform where students submit their work and the assessed pieces are marked against their keys.
fastpixels.ai