Skip to content
Computational Thinking
Enquire

Hands-on AI labs

The lab bench, not the lecture hall.

The work is done with your hands on the keyboard. Every session's live demonstration is first run on the class and then applied by you, on your own laptop, to a real chatbot — and checked, every time, by climbing the Ladder of Misinference on what it gave you. The labs are where the habit forms.

The standing rule

In every lab you must use AI — and you must disclose which tools you used and verify every citation and proposition before you rely on it. The model's confidence is not evidence.

The eight hands-on segments

One lab per session, on your own laptop

Every session ends in thirty guided minutes on your own laptop: a timed run of show, a fixed deliverable, and a verification checklist you sign off against before you submit. Half of every session is you doing the work.

Session 0130 minHands-on

Spot the Rung

In the lab

  1. 01On your own laptop, prompt a chatbot for a confident assertion on a point in your own area of interest, and capture it verbatim — formatting, citations and all.
  2. 02Assemble three statements: the chatbot's answer, an advocate's assertion on a legal point, and a proposition drawn from a reported case headnote.
  3. 03For each, apply the Ladder of Misinference (statement → fact → data → evidence → proof) and mark two things: the rung the claim presents itself as occupying, and the rung the evidence actually supports.
  4. 04Flag the AI-generated statement that is dressed as binding proof but is in fact an unverified statement — the Mata v. Avianca failure in miniature.
  5. 05For each statement, name the single verification step that would promote it one rung, and say honestly whether you could complete that step in ten minutes.

Prompts & tools

  • A generic AI chatbot (ChatGPT, Claude, or Gemini) to generate one of the three statements
  • The Ladder of Misinference (Edmans, May Contain Lies) as the analytic worksheet
  • A reported judgment or headnote from Indian Kanoon, SCC Online or Manupatra

Deliverable

A three-row Ladder worksheet: each statement placed on the rung it claims and the rung it earns, with a one-line justification and the specific verification step that would promote it.

Verification checklist

  • Each statement is assigned to a specific rung (statement / fact / data / evidence / proof).
  • The gap between how the statement presents itself and where it actually sits is identified.
  • The AI-generated statement is correctly flagged as an unverified statement masquerading as proof.
  • A concrete, performable verification step is named for each of the three statements.
Session 0230 minHands-on

Break It and Catch It

In the lab

  1. 01In pairs, train a small image classifier in the browser on two categories, and confirm it works on a fresh example.
  2. 02Retrain it with deliberately lopsided data, then record exactly what it now gets confidently wrong — the toy version of garbage in, garbage out.
  3. 03On your laptop, prompt a generic chatbot for authority on a deliberately niche point of Indian law, or on a judgment after its knowledge cutoff. Capture the output verbatim, including how certain it sounded.
  4. 04Attempt to verify every citation on a grounded source such as Indian Kanoon, and record the precise point at which verification failed.
  5. 05List the tells that marked the output as fabricated — plausible form, absent provenance, confident tone — and write one line explaining the mechanism: it predicted a plausible citation; it did not look one up.

Prompts & tools

  • Google Teachable Machine (or an equivalent browser classifier) for the train-and-break exercise
  • A generic, non-grounded AI chatbot to elicit the fabricated citation
  • Indian Kanoon (or another grounded legal-research tool) to attempt verification

Deliverable

A two-part record: the skew-and-fail note from your classifier, and one documented fabrication — the prompt, the fabricated citation verbatim, the verification attempt, the point of failure, and the tells.

Verification checklist

  • The classifier was deliberately skewed and its resulting confident failure is recorded specifically.
  • At least one fabricated citation or quote was elicited and recorded verbatim.
  • A genuine verification attempt against a grounded source was made and its failure documented.
  • The mechanism is explained in one line, connecting the toy classifier and the chatbot to the same cause.
Session 0330 minHands-on

Legal Prompt Portfolio I — decompose and pattern

In the lab

  1. 01Choose a realistic research question in your own area, and write out the one-line version of the ask you would previously have typed.
  2. 02Build and run a decomposed, staged prompt sequence — issues first, then the governing rule for each, then the authorities that would need checking — and compare it against the one-line version.
  3. 03Draft a clause twice: once from a description of what it should do, once from two exemplar clauses in a house style (few-shot). Compare the fit.
  4. 04Before each task, predict where the model will be strong and where it will be weak; afterwards, record whether the prediction held.
  5. 05In the margin of every output, mark each claim that would still need verifying before it could go near real work.

Prompts & tools

  • A generalist AI model (Claude, ChatGPT, or Gemini)
  • Vague-vs-decomposed prompt pairs on a legal task (decomposition)
  • Few-shot prompts supplying exemplar clauses (pattern recognition)

Deliverable

Two portfolio entries — the decomposed issue-list sequence and the few-shot clause pair — each with prompts, outputs, a short critique, and the unverified claims marked. Feeds the Legal Prompt Portfolio in the recommended assessment scheme.

Verification checklist

  • Decomposition is shown: a staged sequence, not a single monolithic prompt.
  • Few-shot patterning is shown: exemplars supplied, and the output compared against the no-example version.
  • Each entry includes the prompt, the resulting output, and a critique.
  • Outputs containing unverified legal claims are marked as such rather than presented as verified fact.
Session 0430 minHands-on

Legal Prompt Portfolio II — specify and proceduralise

In the lab

  1. 01Write a full specification prompt — role, task, context, constraints, format — for a real legal task, and run it alongside a conversational version of the same ask.
  2. 02Take a tricky legal issue and prompt with IRAC as the procedure plus “reason step by step, showing your working”.
  3. 03Check each step of the reasoning against your own analysis, and mark the first step you disagree with — the conclusion is not the thing being checked.
  4. 04Explore three variants of one prompt, commit to a winner, and record in one line why it won (explore, then exploit).
  5. 05Record each prompt, output and critique, noting which steps of the chain-of-thought you verified and how.

Prompts & tools

  • A generalist AI model (Claude, ChatGPT, or Gemini)
  • Role + task + context + constraints + format prompt template (abstraction)
  • IRAC-as-prompt / “reason step by step” chain-of-thought prompts (algorithm design)

Deliverable

Two further portfolio entries plus an iteration log — prompts, outputs, critiques, and the verified steps of the reasoning. Feeds the Legal Prompt Portfolio in the recommended assessment scheme.

Verification checklist

  • A full specification prompt is shown and compared against a vaguer version.
  • The model was given a procedure (IRAC) and asked to show its working; the working was checked step by step, not just the conclusion.
  • Iteration is shown: three variants explored, a winner chosen, and the reason recorded.
  • All four pillars are now demonstrated across the portfolio (decompose, pattern/few-shot, abstract, algorithm/chain-of-thought).
Session 0530 minHands-on

The SIFT Drill and the Authentication Note

In the lab

  1. 01Take a viral legal-news claim — a forwarded “the Supreme Court has just held…” — and run all four SIFT moves: Stop, Investigate the source laterally, Find better coverage, Trace the claim to its origin.
  2. 02Continue until you reach the origin, or until you can state positively that there is no origin to reach.
  3. 03Ask a chatbot for three sources supporting a proposition in your field, and run SIFT on each: does the source exist, does it say that, is there better coverage? Record a verdict per source — verified, corrected, or fabricated.
  4. 04Write a five-line authentication note: if the first artefact were tendered as evidence, what would be needed to authenticate it, and what would your challenge be?
  5. 05Note where an emotional reaction — yours, or the room's — would have short-circuited the check.

Prompts & tools

  • Caulfield's SIFT method and lateral reading as the procedure
  • A generic AI chatbot to produce the three sources
  • Search, primary sources, and a grounded legal database to trace each claim
  • The electronic-records certificate regime (BSA 2023 s.63, successor to IEA s.65B) as the frame for the authentication note — verify to source before relying on it

Deliverable

A two-part verification log — one viral claim, three AI-supplied sources — with the SIFT steps and a verdict for each, plus a five-line authentication note framing what provenance the artefact would require.

Verification checklist

  • All four SIFT moves are shown for the viral claim, and the origin was actually reached or shown not to exist.
  • Each AI-supplied source was checked both for existence and for whether it says what the model claimed.
  • Verdicts are stated (verified / corrected / fabricated) with the evidence for each.
  • The authentication note names what provenance would be required, not merely that the artefact might be fake.
Session 0630 minHands-on

The Authority Check Drill

In the lab

  1. 01Ask a generic chatbot for five authorities supporting a proposition in your own area, and capture the output verbatim, citations and all.
  2. 02Run step one and two on each: does the authority exist on a grounded database (Indian Kanoon, SCC Online, Manupatra)?
  3. 03Run step three: open each authority that exists and find the paragraph — does it actually say what the model claimed?
  4. 04Run step four, the one people skip: is it still good law? Use a citator (SCC Online Note Up, Manupatra citation analysis).
  5. 05Re-run the same question by supplying a real judgment yourself and asking questions about that text, and compare the reliability and the availability of paragraph-level support.
  6. 06Classify each failure into one of the three classes: does not exist, says something else, no longer good law.

Prompts & tools

  • A generic AI chatbot to elicit the five authorities
  • Grounded databases: Indian Kanoon, SCC Online, Manupatra
  • Citators: SCC Online Note Up, Manupatra citation analysis
  • A real judgment, supplied to the model in full, for the grounded comparison

Deliverable

A five-row authority-check table — each authority with its four-step verdict, its failure class where applicable, and the trail — plus a short note comparing the recall-based answer against the answer given from a supplied document.

Verification checklist

  • All four steps were run on every authority, including the citator step.
  • Failures are classified into the three classes rather than lumped together as “wrong”.
  • The grounded comparison is documented: the same question answered from a supplied source.
  • The trail is complete enough that another person could repeat every check.
Session 0730 minHands-on

Judgment Decomposition

In the lab

  1. 01Choose a long, multi-issue judgment in your area of interest, from the supplied set or your own reading.
  2. 02Prompt the model in stages — the issues the court framed, the holding on each, what was obiter, and only then a summary — asking for paragraph numbers throughout.
  3. 03Open the operative paragraphs and read them against the summary.
  4. 04Mark every divergence and classify it: a flattened dissent, a lost per incuriam finding, a dropped distinguishing fact, an overstated holding.
  5. 05Compare with a neighbour who used a different judgment, and identify the flattening patterns you both found.

Prompts & tools

  • A generalist AI model for the staged summary
  • The judgment itself, from SCC Online, Manupatra, or Indian Kanoon
  • Staged prompts: issues → holdings → obiter → summary, with paragraph numbers demanded at every stage

Deliverable

The staged prompts and outputs, plus a marked-up comparison of the AI summary against the judgment with every divergence flagged and classified.

Verification checklist

  • The summary was built on a structure the student supplied (issues, holdings, obiter), not requested cold.
  • Paragraph numbers were demanded and actually used to check the summary against the judgment.
  • Divergences are flagged specifically — which dissent, which finding, which fact — and classified.
  • At least one flattening pattern is stated in general terms, transferable to the next judgment.
Session 0830 minHands-on

Hallucination Audit (signature drill)

In the lab

  1. 01Receive an AI-drafted legal memo seeded with all three classes of error: a fabricated authority, a real authority that says something materially different, and one that is no longer good law.
  2. 02Skim it first as you would a junior's draft, and note honestly what you would have accepted on a first read.
  3. 03Run the four-step authority check on every citation, and climb the Ladder on every proposition — in its legal reading, statement → fact → authority → binding authority → settled law.
  4. 04Mark each claim verified, corrected, or fabricated, with the trail and the corrected position.
  5. 05Compare your catch rate against the key, by error class, and name the step that would have caught the one you missed.
  6. 06Draft the AI-use disclosure you would attach to this memo if you were filing it.

Prompts & tools

  • Grounded legal-research tools and citators (Indian Kanoon, Manupatra citation analysis, SCC Online Note Up)
  • The four-step authority check and the Ladder of Misinference as the verification frameworks
  • The Calling Bullshit triage: who's telling me this? how do they know? what are they selling?
  • The AI-use disclosure template

Deliverable

An audit report covering every citation and proposition — the source consulted, the verdict, and the verification trail — plus a completed AI-use disclosure. The full version is submitted as the assessed Hallucination Audit.

Verification checklist

  • Every citation has been checked for existence, for content, and for current status.
  • Every legal proposition has been placed on the Ladder against binding authority.
  • Each seeded error is correctly identified and classified, with the verification method documented.
  • No new unverified authority has been introduced; the trail is complete and the AI-use disclosure is attached.

The assessed work

Three pieces of assessed work, plus participation

The labs feed three assessed submissions, alongside the in-class evaluation session. Across all of them, the grade rewards judgment, verification, and process — not raw model output. Using AI is required; disclosing and verifying it is the assessed skill.

Legal Prompt Portfolio

25%

Beginning in Sessions 3 and 4 and continuing through the course, students build and document a portfolio of prompts, outputs and critique across a set of legal tasks — an issue list, a clause, a judgment summary, a counter-argument — demonstrating the four computational-thinking pillars as prompt-craft: decomposition, pattern/few-shot, abstraction (role + task + context + constraints + format), and algorithm/chain-of-thought (IRAC as a procedure). Iteration discipline — explore variants, then exploit the winner — must be shown.

How it's graded

  • Soundness of prompt strategy: clear use of decomposition, few-shot, abstraction, and chain-of-thought.
  • Quality and relevance of outputs for each legal task.
  • Depth of critique: why each prompt worked or failed and how it was refined.
  • Demonstrated iteration (explore-then-exploit) rather than single-shot prompting.
  • Critical awareness of model limits, sycophancy, and the need to verify outputs.

Academic integrity

Use of AI is required and central to this assignment; students must disclose which models and tools were used and verify any legal claim in an output before relying on it. Outputs containing unverified or potentially hallucinated authority must be flagged as such, not presented as fact. Confidential or privileged material must never be entered into public LLMs. The AI-use disclosure form accompanies the submission.

Hallucination Audit

25%

Students receive an AI-drafted legal memo deliberately seeded with all three classes of authority failure — one fabricated, one real but materially misdescribed, and one no longer good law — and must verify every citation and proposition to source using grounded databases, citators, the four-step authority check and the Ladder of Misinference. They submit an audit report flagging each error, documenting the verification method, and stating the corrected authority. This is the signature Session 8 assignment, operationalising the duty to check that separates using AI well from a cost order.

How it's graded

  • Correctness and completeness of verification: every citation and proposition checked to source.
  • Sound use of citators and grounded tools, and of the four-step check and the Ladder.
  • Accurate identification and classification of each seeded error, with a correct corrected authority.
  • Awareness of professional duties (competence, candour to the tribunal, confidentiality) in the write-up.
  • No new unverified authority introduced; the verification trail is clear and reproducible.

Academic integrity

Students are required to use AI for this course but must disclose and verify. For the audit, students must document every verification step and source, must not paste privileged or client data into public LLMs, and must complete the AI-use disclosure form stating which tools were used. Grading rewards judgment, verification and process, not raw model output. Submitting any unverified or invented authority is treated as the very failure the assignment teaches against.

Capstone

20%

Submitted after Session 8: a verified AI-assisted work product in the student's own discipline — for law students a research memo or a contract; for others a research brief, policy note or analysis — accompanied by a verification trail and a short reflective account of process and ethics. The capstone synthesises the whole course around one rule — the model's confidence is not evidence, and an unverified AI answer is never shipped — and situates the student as the human in the loop who owns the work product.

How it's graded

  • Correctness and verification of the work product: every citation and proposition checked to source.
  • Quality and professionalism of the AI-assisted work product itself, judged by the standards of the student's discipline.
  • Completeness of the verification trail (tools used, what was checked, what was rejected).
  • Depth and honesty of the reflective account of process and ethics (confidentiality, candour, disclosure).
  • Reflective insight into the student's role as human in the loop and the responsible use of AI.

Academic integrity

Students must use AI and must disclose it: the capstone requires a documented verification trail and an AI-use disclosure form identifying every tool used. No unverified or fabricated authority may appear in the work product, and no privileged or client data may be entered into public LLMs. Grading explicitly rewards judgment, verification and reflective process over raw model output; shipping an unverified AI answer is treated as a failure of the core competency the course certifies.

Where this goes next

See how the labs add up to a grade — and where they sit in the syllabus.

The assignments above carry the marks; the full weighting, integrity policy, and disclosure requirements live on the assessment page, and the session-by-session plan lives on the syllabus.

See the assessment & integrity policyView the full syllabus

Every deliverable here is graded work in waiting.

The portfolio entries, the audit and the capstone are assembled from what you produce in these thirty-minute segments — so the work is done by the time it is assessed.

How the work is gradedSee the sessions