Legal AI Fundamentals
Legal AI must work within constraints: jurisdiction, sources, accountability. It's not generic AI with a law filter.
- Doba čtení: 16 min
- Končí kvízem
Co tato kapitola pokrývá
- What Makes AI 'Legal'
- Jurisdiction Constraints
- Source Hierarchy
- Accountability Requirements
- Hallucination Risk
- Confidentiality
- Training Data Issues
Kapitoly kurzu jsou napsány anglicky. Zbytek Academy je přeložen.
TL;DR (Core Understanding)
Legal AI fails when it pretends law is text. Law is context + jurisdiction + sources + accountability. Generic chatbots generate language. Legal AI must generate defensible reasoning with traceable sources, scoped assumptions, and jurisdictional control. Anything else is a liability generator.
1) Why "AI for law" is not just NLP
Most AI tools treat law as:
- •Documents to summarize
- •Clauses to remix
- •Answers to sound confident about
That works until:
- •Jurisdiction changes
- •A rule conflicts with another
- •A regulator asks "based on what?"
Legal AI is not about fluency. It's about constraint handling.
2) The four constraints every legal AI must respect
If one is missing, the output is dangerous.
Context
Facts, roles, timeline, intent, constraints. Without context, law is meaningless abstraction.
Jurisdiction
Which country, which state, which regime, which date. Law changes by place and time. Always - HAQQ's Jurisdiction Finder narrows it down.
Sources
Statutes, regulations, cases, guidance. "Common practice" is not a source.
Accountability
Who owns the output? If the answer can't be defended, it shouldn't exist.
Time as a Fifth Constraint
A legal AI output must always answer:
"Valid as of when?"
Without a date anchor:
- • Statutes drift
- • Case law reverses
- • Compliance advice decays
Legal AI without temporal awareness is a liability generator.
3) Why generic chatbots fail in legal work
They:
- •Blur jurisdictions
- •Invent sources
- •Generalize edge cases
- •Ignore hierarchy of norms
They optimize for plausibility, not defensibility.
A confident hallucination is worse than no answer.
Legal coherence ≠ logical consistency
Legal systems tolerate:
- • Contradictions
- • Exceptions
- • Overlapping regimes
Precedent can conflict. Regulation can lag reality.
AI trained to optimize logical consistency will:
- • Oversimplify
- • "Resolve" contradictions that legally exist
- • Invent clarity where none exists
Legal AI must preserve acceptable inconsistency, not eliminate it.
4) Hallucinations aren't bugs. They're design tradeoffs.
This is the section to feel rather than read. The engine under every assistant predicts the next piece of text from the patterns it absorbed; it does not look an answer up. Complete the governing-law clause below one token at a time, then drag the temperature and watch the distribution flatten.
One step, many branches
A model does not pick the next words. It holds a distribution over them.
This Agreement shall be governed by and construed in accordance with the laws of …
The second fan hangs off the first branch, and every other seat fans the same way.
- Thicker branch, likelier continuation
- The MENA seats, thin on purpose
- A collapsed tail, not one candidate
Follow the two fattest branches: England and Wales at 33%, then without regard to at 46%. Multiply them and the likeliest two step reading of this clause is 15%, about fifteen chances in a hundred. That is the version a model states most confidently, and there are more steps in the sentence after it.
The probabilities are chosen to teach, not measured from a model. What is not invented is the shape: a corpus of English language US and UK contracts really does leave the DIFC and the ADGM at the bottom of this list, however well they fit the matter in front of you.
Nothing here knows where your client trades. The ranking records what was drafted most often, and it compounds with every word that follows.
Token 1 of 3
This Agreement shall be governed by and construed in accordance with the laws of …
The governing law
- England and Wales42.4%
- the State of New York30.2%
- the State of Delaware11.2%
- Singapore4.6%
- the Dubai International Financial CentreMENA3.7%
- Switzerland2.9%
- the Abu Dhabi Global MarketMENA2.1%
- the Kingdom of Saudi ArabiaMENA2.1%
- France0.8%
0 · Always the top candidate1.5 · More surprises
Read the bottom of the first list. A model trained mostly on English-language US and UK contracts gives the DIFC and the ADGM a few percent even when they are the only right answer for the matter in front of you. The ranking reflects what was drafted most often, not what should govern your deal.
The candidates and their probabilities are illustrative, chosen to teach: not measured from a real model. The temperature maths is the genuine one: each probability is raised to the power 1/T and renormalised, exactly as a sampler does it.
Two consequences follow from that mechanism and neither is a defect anyone can patch out. A plausible sentence and a true sentence are produced the same way, so a fabricated citation arrives in exactly the register of a real one. And the patterns stop at a date.
Knowledge cutoff
Three marks on one timeline, and you are always standing to the right of all of them
A timeline in three steps, in order: the training data ends, the model is released months later, and later still a leading case is overturned. The first step is inside the model's world. The second and third are after the cutoff and do not exist for it.
Training data ends
The last thing the model ever read
Everything up to here is in the weights. Nothing after it is, and the model was given no way to notice the edge.
Released, months later
You start using it
Training, evaluation and safety work take months. On the day the model ships, its knowledge is already that many months old.
After the cutoff
A leading case is overturned
The authority the model will cite is no longer good law. Nothing told the model, so nothing in its answer will hesitate.
Everything after the first mark does not exist for the model. Ask it about the overturned case and you will not get a refusal, you will get the superseded position, stated with complete confidence and in the right register. To work from recent law, the tool has to go and fetch that law and put it in front of the model.
Language models are built to:
- •Continue text
- •Sound coherent
- •Fill gaps
Law punishes gaps.
In legal work:
- •Fabricated case law = malpractice risk
- •Wrong statute = invalid advice
- •Outdated rule = compliance failure
This is why "be careful" is not a strategy.
In law, a beautifully wrong answer is worse than no answer at all. Fluency can hide fake law. Judge a legal AI the way you'd judge a chess engine - look for the Stockfish of law, the tool that values correctness over confidence. The next chapter makes this a concrete test: The "I Don't Know" Test.
5) Traceability is non-negotiable
Legal outputs must answer:
- ?Which source supports this?
- ?Which version/date?
- ?What assumptions were made?
If you can't trace it, you can't trust it.
Traceability is what turns AI from toy to tool.
6) Auditability beats creativity
Legal AI must produce:
- ✓Repeatable results
- ✓Inspectable reasoning
- ✓Reviewable steps
Creativity is useful in strategy.
It's toxic in compliance.
7) Reasoning ≠ explanation
Many tools "explain" nicely.
Legal AI must reason:
- •Apply rules to facts
- •Weigh conflicts
- •Surface uncertainty
Pretty explanations without reasoning are cosmetic.
Interpretive Latitude (why answers are ranges, not points)
Judges and regulators operate within bounded freedom.
Law often allows:
- • Multiple valid interpretations
- • Competing readings
- • Discretionary outcomes
Legal AI must surface:
- • Interpretation ranges
- • Risk bands
- • Decision sensitivities
Single "correct" answers are usually wrong.
8) Liability doesn't disappear with automation
If AI drafts:
- 1.You review
- 2.You approve
- 3.You are responsible
Courts don't subpoena models. They subpoena people.
This is why legal AI must be designed for human-in-the-loop, not human-out-of-the-way.
9) Regulations are already catching up
Legal AI is now intersecting with:
- •Professional responsibility rules
- •Data protection regimes (e.g. the European Union's)
- •AI governance frameworks
Dates, versions, and scope matter.
If your AI can't anchor itself in time, it's unusable.
10) The hidden requirement: legal systems thinking
Good legal AI models:
- •Hierarchy of norms
- •Procedural constraints
- •Adversarial logic
- •Risk framing
They don't just answer. They structure uncertainty.
Chapter 11 Takeaway
Legal AI isn't about sounding smart. It's about surviving scrutiny.
Next chapter makes correctness a concrete test, then gets practical and sharp: prompting for legal work properly, without turning AI into a liability cannon.
Visual Summary
Visual: Four Constraints Box
If one is missing, output is dangerous
Visual: Generic Chatbot vs Legal AI
- • Optimizes for fluency
- • Blurs jurisdictions
- • Invents sources
- • No accountability
- • Optimizes for defensibility
- • Jurisdiction-bound
- • Source-controlled
- • Traceable outputs
Visual: Traceability Chain
Visual: Human-in-the-Loop
Chapter 11 Quiz
Question 1 of 6What are the four constraints every legal AI must respect?