Lesson objective

Predict when an AI answer will change or degrade because of document length, an ambiguous instruction, a small change of wording, a different model or version, or ordinary non-determinism, and apply the controls that make the output checkable anyway.

Plain-language explanation

Why this matters now

The first reaction of most professionals to a wrong AI answer is to blame the model. The second, more useful reaction is to ask what the model was given. Vendor documentation now describes, in plain terms, several ways an answer changes for reasons that have nothing to do with the law. Anthropic's context-window page states that "as token count grows, accuracy and recall degrade, a phenomenon known as context rot." Its glossary states that "even with temperature set to 0, the results will not be fully deterministic and identical inputs may produce different outputs across API calls." OpenAI's API reference says the same of its seed setting: "Determinism is not guaranteed." These are the vendors describing their own products, as published on September 14, 2026.

A legal professional who knows these facts stops asking "Is the tool right?" and starts asking "What did I give it, and what could have changed the answer?" That question is answerable. The first one is not.

A question to sit with

You ask a permitted assistant how many days a fictional rule allows for a reply. It says fourteen. A colleague asks the identical question in her own session and gets twenty-one. You ask again and get fourteen. Which answer is right? The honest answer is that the exercise has told you nothing about the rule and everything about the tool.

What goes wrong

The document was too long. Module 01 introduced the context window: everything the model can see at once, including your documents, the conversation so far and its own answer. Anthropic's documentation states that everything counts against it: the system prompt, every message, tool results, images and documents, and the model's own thinking. If input alone exceeds the window, the request fails with an error. Below the limit, accuracy still degrades as the window fills. A 400-page appendix may fit and still be read badly, and the answer will not say so.

The instruction was ambiguous. "Summarise the key terms" means different things to a litigator and to a deal lawyer. The model picks one reading and answers confidently. "Is this clause enforceable?" without a jurisdiction invites an answer for no jurisdiction. Dahl and colleagues found in 2024 that general-purpose chatbots "often uncritically accept users' incorrect legal assumptions"; an ambiguous prompt is an invitation for the model to supply the assumption itself.

The wording changed, and so did the answer. Small changes in a prompt change what the model predicts. Asking "Does the tenant have a defence?" and "What defences might the tenant raise?" can produce answers of different length, confidence and content. Neither is the "true" answer; each is a prediction shaped by the words used.

The model or version was different. Two products, or two versions of one product, are different predictors with different training data, cutoffs and behaviour. A workflow tested on one is not tested on the next. Curran and colleagues at the University of Melbourne, in a 2025 workshop paper, posed the same 100 legal scenarios for Los Angeles, London and Sydney to GPT-4o, Claude 3.5 Sonnet and Gemini 1.5 Pro and found hallucination rates of 0.45, 0.55 and 0.61 respectively, with the rate "significantly associated with place." Where the question is asked, and of which model, changes how often the answer is wrong.

Nothing changed, and the answer still differed. This is non-determinism: the same input, the same model, the same settings, and a different output. A September 2025 technical post by Horace He and Thinking Machines Lab explains that "LLM APIs are still not deterministic in practice" even at temperature zero, mainly because server load changes how requests are batched and the numerical results shift slightly with it. The practical point is not the mechanism. It is that a single run proves nothing.

The practical answer

Know the five causes and the control for each.

Cause What it looks like Control
Context limit and context rot Vague treatment of the end of a long document; details from the start, guesses about the rest Split long material; ask about the specific section; confirm the tool saw it by asking it to quote a distinctive sentence
Ambiguous instruction A confident answer to a question you did not quite ask State the jurisdiction, the posture, the audience and the output you want; supply the premise instead of implying it
Prompt sensitivity Different answers to near-identical wordings Fix the wording in a saved prompt; when the wording changes, treat it as a new test
Model and version difference A workflow that worked last month behaves differently Record the model and version on every run; retest saved prompts after any change
Non-determinism Different answers to the identical prompt Never rely on one run for a fact that matters; verify against the source, not against a second run

Temperature, explained once. OpenAI's API reference describes temperature as a sampling setting between 0 and 2: "Higher values like 0.8 will make the output more random, while lower values like 0.2 will make it more focused and deterministic." Anthropic uses a 0 to 1 range for the same idea. Most legal products hide the setting. Two things follow. First, if you can see the setting, lower is steadier for fact-shaped tasks. Second, lower is not the same as fixed: both vendors state that identical inputs can still produce different outputs.

Why "ask it again" is not verification

Asking the model the same question twice and getting the same answer feels like confirmation. It is not. Two predictions from the same process can agree and both be wrong, exactly as two guesses can. Agreement between runs tells you the answer was stable, not that it was true. Verification means comparing the answer to a source outside the model: the rule, the document, the case.

Write prompts that remove the causes. A prompt that closes the gaps has five parts. The prompt-engineering module builds this into a full architecture; the foundation is here.

  1. Jurisdiction and posture. "Fictional State of Arden; trial court; motion to dismiss."
  2. The material, in full or in the relevant part. Paste the clause, the rule text, the pages that matter. Do not describe them.
  3. The premise, stated rather than assumed. "Assume the notice was served on March 3. Do not assume anything about whether it was timely."
  4. The output you want. "List each defence in one sentence with the page of the pasted material it relies on."
  5. The check. "If any statement relies on material not pasted, say so explicitly."

Handle long documents deliberately. For any document that approaches the published window, or any document where the answer depends on a late section:

  • Split the document and ask about each part, then ask the tool to reconcile the parts.
  • Ask the tool to quote the distinctive sentence you know appears near the end. If it cannot, it did not read that far.
  • Read the decisive pages yourself. The tool's summary is a map; the document is the territory.

Record the model and version. Every run that feeds a work product should record the product, the model, the version if visible, the date and the prompt used. When a vendor changes the model, saved prompts are retested on fictional material before they are trusted again. This is not bureaucracy; it is the only way to know why an answer changed.

Remember this

  • Accuracy degrades as the context window fills; a document that fits can still be read badly, and the answer will not say so.
  • An ambiguous prompt invites the model to supply the assumption. Name the jurisdiction, the posture, the premise and the output.
  • Small wording changes, different models and different versions all change the answer. Record the model and retest after any change.
  • The same prompt can give different answers on different runs, even at the lowest temperature. One run proves nothing; agreement between runs is stability, not truth.
  • Verify numbers, deadlines and rules against the source. That is the only control that works for every cause.

Legal example

At Ferrante & Sayre, the fictional firm, partner Miriam Vanterpool asks associate Tomas Berg to use a permitted assistant to find every indemnity obligation in a fictional 310-page asset purchase agreement for client Odalys Marin, and to say which party bears environmental liabilities.

First run. Tomas uploads the whole agreement and asks, "Which party bears environmental liabilities?" The answer is fluent: the buyer bears them, "under Article 9." He checks the model facts card from Module 01; the agreement fits the published window with room to spare. But Article 12, near the end, contains a carve-back that shifts pre-closing environmental liabilities to the seller. The answer never mentions it.

Second run. He asks the assistant to quote the first sentence of Article 12. It produces a plausible sentence that is not in the document. The end of the agreement was not reliably read, even though it fit.

Third run. He splits the agreement into three parts, asks about environmental liability in each, then asks the tool to reconcile the three answers. This time the carve-back appears. He then reads Articles 9 and 12 himself and confirms the allocation.

The ambiguity. Miriam had said "indemnity obligations." Tomas realises the assistant treated that as indemnities given by the seller and omitted the buyer's. He rewrites the prompt: "List every indemnity in this agreement, whoever gives it, with the article number, in a two-column table."

The number. The agreement caps the seller's indemnity at a percentage of the purchase price. The assistant reported "ten percent" in one run and "fifteen percent" in another, with identical prompts. Tomas records both, opens the clause, and reads the figure himself. It is twelve. Neither run was right.

He records the model, the version shown in the product, the date, each prompt and each answer in the run record, and hands Miriam a memo that cites article numbers and page references she can open. Her review note: "Assistant used for navigation only. Every allocation and every figure confirmed from the agreement text. Run record retained; prompt saved for retest after the next product update."

Responsible-use warning

Never rely on a single AI run for a deadline, a figure, a percentage or a rule, and never treat a second run that agrees with the first as verification; two predictions can agree and both be wrong. Never assume a long document was fully read because it fit the published context window, and never let an ambiguous instruction stand when the jurisdiction, posture or premise could change the answer. These habits prevent the memo that reports the wrong cap, the summary that missed the carve-back at the end, and the confident answer to a question nobody asked.

Human-review requirement

A licensed attorney reviews every AI-assisted analysis, summary or extraction before it is relied on and owns the decision to use it. In this lesson's context the reviewer checks the run record: the model and version, the date, the exact prompt with its jurisdiction, posture and premise stated, and whether the material was pasted rather than described. For long documents the reviewer confirms the material was split or that the tool proved it read the decisive section. Every number, deadline and rule in the output is marked as read from the source by a named person. The reviewer records the tool, version, date and what was verified in the matter's AI use record.

Practical exercise

Run a controlled test of prompt sensitivity, non-determinism and context limits, using only fictional material and a tool your firm permits, and produce a run record.

  1. Write a fictional two-page rule or clause that contains one deadline in days, one percentage cap and one exception buried in the final paragraph.
  2. Open a run record with columns: run number; model and version; date; exact prompt; answer given; matches source (yes/no); cause if not (context, ambiguity, wording, model, non-determinism).
  3. Non-determinism. Paste the text and ask, "What is the deadline in days?" three times in separate sessions. Record each answer.
  4. Prompt sensitivity. Ask once more with different wording ("By when must the party act?"). Record the answer and note any change.
  5. Ambiguity. Ask, "What is the cap?" without saying which cap. Then ask, "What is the percentage cap on the indemnity in paragraph 4?" Record both answers.
  6. Context. Pad the fictional text with several pages of unrelated fictional filler before the final paragraph, so the exception sits at the end of a long input. Ask whether any exception applies. Then ask the tool to quote the last sentence of the document. Record whether it could.
  7. If a second model or version is available to you, repeat step 3 once on it and record the difference.
  8. Write a short conclusion: which causes you observed, which control from the lesson addresses each, and the sentence you will add to every future prompt to remove the ambiguity you found.
  9. Paste the run record and conclusion into the submission box. In the workbook, Module 02 worksheet, "Run record," keep the rows with the date, and set a reminder to retest the saved prompt after the next product update.

Knowledge check

  1. What is context rot, and why does it matter even when a document fits within the published context window?

    Answer guidance: The vendor-documented tendency for accuracy and recall to degrade as the token count grows. A document can fit and still be read badly, especially at the end, and the answer will not say which parts were handled poorly; the control is to split, ask about specific sections and confirm by asking for a distinctive quotation.

  2. An assistant gives "ten percent" as an indemnity cap on one run and "fifteen percent" on the next, with identical prompts. What do you do?

    Answer guidance: Treat neither as the answer. Non-determinism means identical inputs can produce different outputs even at the lowest temperature; open the clause and read the figure from the agreement. Record both runs and the source check in the run record.

  3. A colleague ran the same question twice, got the same deadline both times, and calls it "verified." What is wrong with that reasoning?

    Answer guidance: Agreement between runs shows stability, not truth; two predictions from the same process can agree and both be wrong. Verification means comparing the answer to the source outside the model, in this case the current rule text.

  4. The firm's saved contract-review prompt worked well for six months. The vendor announces a new model version. Should the team keep using the prompt as before?

    Answer guidance: Not without retesting. A different model or version is a different predictor with different training and behaviour; the prompt should be rerun on fictional material, the results recorded with the new version, and only then returned to use on client work.

Sources and further reading