Few topics attract more confident wrong opinions than language models in legal work. Some people insist the technology is a liability waiting to happen and refuse to touch it. Others treat it as a near-lawyer that can be trusted to draft binding language. Both positions are myths, and both lead to bad decisions.
This article works through the most durable misconceptions about prompting for legal and compliance writing and replaces each with the accurate picture. The accurate picture is less dramatic than either extreme: a well-controlled drafting process that is genuinely useful for a defined set of tasks and genuinely dangerous when used outside them.
Getting the myths right matters because they shape behavior. A firm that believes the "it will fabricate everything" myth bans a tool that could safely accelerate routine drafting. A firm that believes the "it basically knows the law" myth ships errors. The goal is calibration, not enthusiasm or fear.
Myth: The Model Knows the Law
This is the most consequential misconception because it invites overtrust.
The reality of trained patterns versus authority
A model has absorbed enormous amounts of text, including legal text, but it does not hold an authoritative, current, jurisdiction-specific knowledge of the law. It produces statistically likely legal-sounding language. Sometimes that language is correct; the model cannot tell you which times. Treating its output as legal knowledge rather than as a draft is the root of most serious errors.
Why grounding changes everything
The accurate move is to supply the actual statute, contract, or policy and have the model work from that text. This shifts the task from "recall the law" (which it cannot do reliably) to "restate and structure this supplied material" (which it does well). The myth dissolves once you stop asking the model to be the source of authority.
Myth: It Will Inevitably Fabricate, So It Cannot Be Trusted
The opposite error, equally common among skeptics.
Fabrication is a controllable failure mode
Yes, ungrounded models invent citations. But fabrication drops dramatically when the model drafts from supplied sources and is instructed to flag anything it cannot ground. The risk is real but it is a controllable failure mode, not an immovable property of the tool. Dismissing the technology because of uncontrolled fabrication is like refusing to drive because cars can crash without seatbelts.
Verification closes the residual gap
Even with grounding, a verification step that checks every citation, figure, and date catches the residual errors. The combination of grounding plus verification makes the practice defensible. The detailed risk treatment in The Quiet Liabilities Buried in Prompting for Legal Text shows how far controls can shrink the danger.
Myth: Better Prompts Eliminate the Need for Lawyers
A favorite of vendors and a misreading of what prompting does.
Prompting changes the task, not the responsibility
A strong prompt produces a stronger draft. It does not move professional responsibility off the human who signs the document. The lawyer's job shifts from drafting from scratch toward reviewing, correcting, and standing behind the result, which is different work but not absent work.
The judgment that does not transfer
Deciding what to argue, how aggressive a position to take, what a client actually needs, and whether a clause creates unacceptable exposure are judgment calls grounded in responsibility and context the model does not have. Prompting accelerates production; it does not supply judgment.
Myth: One Good Prompt Works for Everyone
The belief that a clever prompt, once written, is a permanent asset for the whole team.
Context and jurisdiction break universality
A prompt tuned for one jurisdiction, document type, or risk posture often fails in another. Force, terminology, and required disclosures vary. The accurate picture is a library of context-specific prompts with owners, not a single magic prompt. The companion piece on Standardizing AI Drafting Across a Legal and Compliance Function lays out how that library should be structured.
Prompts drift with models
A prompt that behaved well on one model version can behave differently after an update. Treating any prompt as permanently solved ignores how the underlying technology shifts beneath it.
Myth: Plain-Language Conversion Is Low-Risk
Because it feels like simplification rather than legal work, this use case is widely underestimated.
Simplification can drop substance
Converting dense legal text to plain language can quietly remove a qualifier or condition that changes meaning. The output reads cleaner and is materially less accurate. Always check the plain version against the original for lost substance, not just for readability.
Readability is not fidelity
A simpler sentence is not automatically a faithful one. The myth conflates the two. The accurate practice treats plain-language conversion as a substantive task requiring the same verification as any other legal drafting.
Myth: Using AI Means Disclosing It Everywhere or Nowhere
Two opposite misconceptions cluster around disclosure, and both are wrong.
The all-or-nothing framing is false
Some assume that any AI involvement must be disclosed in every document; others assume that because the human reviewed and signed it, disclosure never matters. Neither holds. Disclosure obligations depend on context: the forum, the rules that apply, and what was actually represented. The accurate stance is to know your specific obligations rather than adopting a blanket policy in either direction.
Provenance is what makes disclosure tractable
The practical reality is that you cannot disclose what you did not record. Teams that capture provenance, the source material, the prompt, and the human review, can answer disclosure questions accurately if they arise. Teams that treat the question as settled in advance get caught flat-footed when a specific rule says otherwise. Recording how a document was produced is the move that survives whichever way the obligation falls.
Myth: The Skill Is Writing Clever Prompts
A final misconception locates the value in the wrong place.
The value is in the controls, not the cleverness
People imagine the expert is the one with the most ingenious prompt. In legal and compliance work, the expert is the one with the most disciplined process: grounding, verification, and human approval. A plain prompt run through strong controls beats a clever prompt run through none. The cleverness myth leads teams to chase prompt tricks instead of building the controls that actually protect them.
Judgment about when not to use it
Part of the real skill is recognizing tasks where the model should not be used at all, because the stakes, novelty, or judgment involved exceed what controlled drafting can safely support. Knowing the boundary of safe use is more valuable than any prompt phrasing, and it is precisely the judgment the cleverness myth ignores.
Frequently Asked Questions
Does the model actually understand legal concepts?
It produces statistically likely legal-sounding language based on training, which is not the same as authoritative, current knowledge of the law. It can restate supplied material well; it cannot be trusted as the source of legal authority. Grounding it in real source text is what makes it useful.
If fabrication is controllable, why does it still happen so publicly?
Because the people in those cases used ungrounded models and skipped verification. Both controls were available and unused. The public failures are failures of process, not proof that the technology cannot be used safely.
Will prompting eventually replace junior legal drafting work?
It changes the shape of that work toward reviewing and correcting strong drafts rather than producing them from scratch. The professional responsibility for the result does not transfer to the tool, so the human role persists even as the task changes.
Is there a universal best prompt for legal writing?
No. Prompts that work depend on jurisdiction, document type, and risk posture, and they drift as models update. A maintained library of context-specific, owned prompts is the realistic alternative to a single universal prompt.
Why treat plain-language conversion as risky?
Because simplification can silently drop a material qualifier, producing text that is more readable and less accurate. Readability and fidelity are different properties, and the conversion needs the same verification as any other legal draft.
How should a skeptical firm calibrate?
Reject both extremes. The tool is neither a near-lawyer nor an inevitable liability. With grounding, verification, and human approval, it safely accelerates a defined set of drafting tasks, which is the accurate and useful middle position.
Key Takeaways
- The model does not know the law; it produces legal-sounding language and must be grounded in real source text.
- Fabrication is a controllable failure mode, not a reason to dismiss the technology outright.
- Prompting changes the task but never transfers professional responsibility off the human who signs.
- No universal prompt exists; context and model drift demand a maintained, owned library.
- Plain-language conversion is substantive work, not low-risk simplification, and needs full verification.
- Calibration beats both fear and enthusiasm: useful within defined tasks, dangerous outside them.