AI and the Law I — Judgments, Doctrine and Comparative Analysis
Apply the four pillars to the core problem of legal research: extracting the reasoning from long, multi-issue, multi-judge judgments without inheriting what the machine flattened.
The hook
“A nine-judge bench, six separate opinions, several hundred pages. A model will summarise it in ten seconds — and the summary will be fluent, confident, and quietly missing the dissent that decides your case.”
What you'll be able to do
- Decompose a lengthy judgment into issues, holdings and obiter before prompting for a summary — so the summary has a structure you can check.
- Extract a ratio with paragraph numbers, and use the model as an index into a long document rather than as a substitute for reading it.
- Trace how a principle has evolved across a line of cases, and verify every link in the chain.
- Say precisely where AI summarisation flattens what matters — dissents, per incuriam findings, distinguishing facts — and where comparative analysis is genuinely strong.
On the syllabus
- Decomposing a lengthy judgment into issues, holdings and obiter before prompting for a summary
- Extracting the ratio with paragraph numbers; using the model as an index, not a substitute
- Tracing a doctrine across a line of cases, and verifying every link
- Where AI summarisation flattens: dissents, concurrences, per incuriam findings, distinguishing facts
- Comparative analysis — judgments, statutes and clauses side by side, and what it cannot see
In short
Session 7 is where the computational-thinking skills of Sessions 3 and 4 meet a real problem of legal research: reading long, multi-issue, multi-judge judgments efficiently and accurately. Decomposition first — issues, holdings, obiter — then a staged prompt; abstraction to extract the ratio with paragraph numbers; a procedure for tracing a doctrine across cases with every link checked; and throughout, the discipline of reading the model's summary against the judgment rather than instead of it. The hour maps exactly where summarisation is reliable and where it flattens the details a matter turns on, and the hands-on half hour has students catch that flattening themselves.
Why it matters for using AI well
Let the model index, structure and draft the first pass; never let it be your only reading. The summary is a hypothesis and the judgment is the evidence — decompose, extract with paragraph numbers, and compare, and you get the speed without inheriting the flattening.
What they leave with
The skill
Structure the judgment yourself, demand paragraph numbers, and read the summary back against the operative paragraphs.
The insight
A model compresses toward the most common phrasing, so the first thing it loses is the exception — and the exception is usually your case.
The moment they remember
The room votes on whether a single passage is ratio, obiter, or from a dissent — and splits badly, which quietly establishes that the distinction is hard for humans before anyone blames the machine. Then the model's summary goes up beside the operative paragraphs and the students find what was flattened, themselves, out loud. The delight is not watching AI fail; it is the specific professional pleasure of being the person in the room who spotted the missing dissent — which is exactly the identity the course is trying to build.
In this session
- 01
Decompose before you summarise: ask the model first for the issues the court framed, then for the holding on each, then for what was obiter — and only then for a summary. A summary built on a structure you supplied can be checked against that structure; a summary requested cold cannot be checked against anything.
- 02
Abstraction as extracting the ratio: the prompt that asks for “the principle that binds, in one sentence, with the paragraph number” is the machine version of the skill you learned in your first year — and the paragraph number is precisely what makes it checkable.
- 03
Navigating rather than reading: where models genuinely help is locating the passage on a point in a four-hundred-page judgment, listing which opinions addressed which issue, and building a table of issues by judge. Use the model as an index, and read the passages yourself.
- 04
Tracing a doctrine: staged prompts that ask a model to trace how a principle has evolved across a line of cases produce a plausible, coherent narrative — and coherence is exactly what the model optimises for, whether or not each link is real. Every link goes through the Session 6 check: does the case exist, did it say that, and was it the majority?
- 05
Where summaries flatten: dissents and concurrences collapsed into “the court held”; a per incuriam finding lost; the distinguishing fact that made a ruling narrow dropped because it did not fit the pattern. These are precisely the details on which a matter turns, and they are the first casualties of compression toward the most common phrasing.
- 06
Comparative analysis: setting two judgments, two statutory regimes or two contractual clauses side by side is a task models do well, because it is reshaping text they have been given. What they miss is what is not on the page — the defined term three pages earlier, the amendment since the training cutoff, the proviso in a different instrument.
- 07
The Indian frame for AI as navigation rather than authority: in Jaswinder Singh v. State of Punjab (P&H HC, 2023) the court consulted ChatGPT for a broader picture of bail jurisprudence and expressly did not decide the bail on it; the Supreme Court's White Paper (Nov 2025) puts the principle as “Judges must remain the ultimate decision-makers, AI may assist, but it cannot substitute human judgement.” Read it precisely: the paper's express prohibition on *delegating* decision-making to AI is described as the Delaware Supreme Court's rule, not adopted as an Indian one.
- 08
The comparison discipline: the AI summary is a hypothesis about the judgment. Read the operative paragraphs against it, mark every divergence, and keep the log — which, incidentally, is also the habit that makes you a faster reader of judgments.
The four-step mirror
Run it on the class. Then on the machine.
An experiment on the room, the same effect explained in the model, a live demonstration on a real chatbot, and a named takeaway skill.
Ratio or obiter? — spot what the summary flattened
On the class
A passage from a multi-judge judgment goes up. The room votes: is this proposition the ratio, obiter, or from a dissent? The split shows how hard the distinction is even for humans.
In the model
A model summarising the same judgment compresses toward the most statistically common phrasing — “the court held” — and the dissent, the concurrence and the per incuriam finding are the first casualties.
Live chatbot
The presenter asks a chatbot to summarise the judgment, then puts its summary next to the operative paragraphs: the room finds what was flattened.
The skill
Decompose into issues, holdings and obiter before you ask for a summary — and read the summary against the judgment, not instead of it.
Trace the doctrine — and check every link
On the class
The room names a doctrine everyone knows and votes on how many cases they think it took to get from its origin to its present form.
In the model
Asked to trace a doctrine across cases, a model produces a coherent narrative — and coherence is exactly what it optimises for, whether or not each link is real.
Live chatbot
The presenter asks for the line of cases with holdings, then checks each one on Indian Kanoon in front of the room: which exist, which said that, which were the majority.
The skill
A traced doctrine is a chain of claims. Verify each link — exists, says that, was the majority — before you rely on the chain.
Side by side — what comparison catches and misses
On the class
Two versions of a clause go up; the room lists the differences it can spot in sixty seconds.
In the model
Reshaping text it is given is the model's strength: it will produce a clean comparison table in seconds. What it cannot see is the defined term elsewhere in the contract, or the amendment to the statute since its training.
Live chatbot
The presenter runs the comparison, and the room finds the difference the table missed.
The skill
Use AI for the side-by-side; supply everything it needs to see; and read for what is not on the page.
Hands-on · on your own laptop
Judgment Decomposition
Take a long, multi-issue judgment and build a summary you can actually check. Prompt in stages — issues, holdings, obiter, then summary — demanding paragraph numbers throughout; then read the operative paragraphs against the summary and mark every divergence.
Run of show · 30 minutes
- 0–5 min — Choose a long, multi-issue judgment in your area, from the supplied set or your own reading.
- 5–18 min — Run the staged sequence: the issues the court framed, the holding on each, what was obiter, then the summary — asking for paragraph numbers at every stage.
- 18–26 min — Open the operative paragraphs and read them against the summary. Mark every divergence: a flattened dissent, a lost per incuriam finding, a dropped distinguishing fact.
- 26–30 min — Compare with a neighbour who used a different judgment and identify the flattening patterns you both found.
Deliverable
The staged prompts and outputs, plus a marked-up comparison of the AI summary against the judgment with every divergence flagged and classified.
Key sources & cases
Indian research stack: SCC Online, Manupatra, Indian Kanoon, CaseMine
The grounded tools used to open the judgment itself and to check every case in a traced line of authority.
Kevin D. Ashley, Artificial Intelligence and Legal Analytics (2017)
What machines can and cannot do with legal texts — the ground under the session's map of reliable versus flattened summarisation.
Jaswinder Singh v. State of Punjab (P&H HC, 2023)
The Punjab & Haryana High Court consulted ChatGPT for a “broader picture” of bail jurisprudence — and did not decide the bail on it. A real Indian example of AI as navigation, not authority.
Supreme Court of India, White Paper on Artificial Intelligence and Judiciary (Centre for Research and Planning, Nov 2025)
Verified 2026-08-26 against the official PDF. Note the exact title — “Artificial Intelligence and Judiciary”, with no “the”. States that “Judges must remain the ultimate decision-makers, AI may assist, but it cannot substitute human judgement” (p.10) and that “AI may assist the judges, but cannot replace them” (p.65). Guideline 14 requires all AI-derived information to be independently verified before reliance; Guideline 15 forbids using one generative tool to verify another; Guideline 12 directs that private, confidential or legally privileged information not be input into ANY AI tool. On deployment it *suggests* courts prioritise secure in-house tools over “open-source or publicly accessible” ones — it does NOT restrict cloud AI, and never uses that framing.
Richard Susskind, Tomorrow's Lawyers
How legal reading and legal work change when the machine does the indexing.
Readings
- Kevin D. Ashley, Artificial Intelligence and Legal Analytics (2017)
- Supreme Court of India, White Paper on Artificial Intelligence and Judiciary (Centre for Research and Planning, Nov 2025)
- Jaswinder Singh v. State of Punjab (P&H HC, 2023)
- Richard Susskind, Tomorrow's Lawyers — how the reading changes when the machine indexes
Next session
Session 08 / 08
AI and the Law II — Duty, Confidentiality and the Verified Work Product
Own the work productSixteen hours, one professional discipline.
Using AI well is not a knack — it is competence, candour and verification, practised until they are habits you could defend in court.