How to use an AI contract generator without getting burned

An AI contract generator drafts and finds gaps. It does not decide enforceability.

The honest division of labour

Contract work splits into two jobs that feel like one. The first is deciding what the deal actually is: who owes what, by when, at what price, and what happens when it goes wrong. The second is expressing those decisions in language that survives a dispute in a particular legal system.

A model is genuinely useful on the first job and unreliable on the second. It has read enough agreements to notice that you never said when the invoice is due, who owns the design files, or what happens if the client goes quiet for six weeks. That gap-finding is the real value, and it is honest value — no legal judgement is required to observe that a document is silent on payment timing.

Fluency is not authority

The failure mode is specific. A generated clause reads like every other clause you have seen, so it inherits their credibility without having earned it. NIST's generative AI profile names this directly as confabulation: confidently stated content that is simply wrong.

“GAI systems generate and confidently present erroneous or false content in response to prompts.”

NIST AI 600-1, Generative AI Profile

In a report, a confabulation costs you a correction. In a contract, it costs you the thing the clause was supposed to protect — and you find out at the worst possible moment, when someone is already disputing it. So set the rule up front: the model drafts and flags, it never certifies. Ask it to say “this depends on your jurisdiction, ask a lawyer” rather than reassure you.

A contract is a list of decisions

Nine things that must be decided somewhere

Every blank you leave is a decision the draft makes for you, silently, usually in whichever direction the training data leaned. The table below is the checklist worth running any draft against — human or generated. The right-hand column is not a legal outcome; it is the argument you will be having.

Why IP ownership deserves its own paragraph

Take the row that catches people most often. Under United States copyright law, a commissioned work only counts as a work made for hire if it falls into one of nine listed categories — a contribution to a collective work, part of a motion picture or other audiovisual work, a translation, a supplementary work, a compilation, an instructional text, a test, answer material for a test, or an atlas — and the parties expressly agree in a signed written instrument that it is a work made for hire. The U.S. Copyright Office states plainly that if a work fails any of those requirements, it is not a work made for hire.

Software is not on that list of nine. Neither is a logo. So a US agreement that simply calls a contractor's code “work made for hire” and stops there has left the ownership question open in a way that is invisible until it matters. A model asked to draft an IP clause will often produce exactly that sentence, because it appears in thousands of real contracts. What it should do instead is flag ownership as a decision you have to make deliberately — assignment, licence, or something else — and send you to a lawyer to write it. See the Copyright Office's Circular 30 for the actual requirements, which apply to United States copyright only.

Brief the model before it drafts a word

Five inputs that change the whole draft

1. The deal in one sentence. Who does what, for whom, for how much. If you cannot write it, the contract cannot express it.
2. Both parties, as legal entities. Not “me and Dave” — the actual names, entity types, and countries. This changes which clauses are even relevant.
3. The jurisdiction, stated explicitly. Without it a model defaults to a generic US-flavoured template regardless of where you live. Name the country and state or region.
4. The worst realistic outcome. “They do not pay” and “their customers' data leaks because of my code” produce different liability sections. Say which one keeps you up.
5. Your red lines. The two or three things you will not sign away. Everything else is negotiable, and saying so up front stops the draft from quietly conceding them.

Get a file you can redline, not text in a chat window

Why the format is not a detail

Contract review happens in tracked changes. Two parties and their advisers pass a document back and forth, each version marked against the last, comments in the margin explaining why a cap moved. A draft that lives as scrollback in a chat interface cannot participate in that process at all. You end up copying it into Word by hand, which is both the moment formatting breaks and the moment clause numbering silently shifts.

So ask for the file explicitly. In OpenCraft, makeDoc takes Markdown and renders a real .docx alongside a doc-styled HTML preview, so you read it in the browser and download the editable original to redline. Because a re-generated file with the same name stacks as a new version of the same artifact rather than a fresh one, the draft history stays in one place while you iterate.

Bringing the other side's paper in

The commoner situation is that someone sends you a contract and you need to understand what you are agreeing to. You can attach a .docx or a PDF and have the text extracted for review — older formats such as .doc, .rtf, and .odt are not readable, so re-save those as .docx first. Then ask the useful question, which is not “is this a good contract” but “which of the nine decisions above does this document leave open, and which of them does it decide against me?”

Review the facts yourself, then send the law to someone else

The pass only you can do

Before anyone bills you an hour, check the things that need no legal training and are wrong surprisingly often. Are both legal names exact, including entity suffix? Do the dates form a possible sequence? Does every defined term get used, and does every used term get defined? Do the cross-references point at the clauses they claim to? Does the payment schedule add up to the total? Is the currency the one you agreed? Generated drafts inherit numbers from wherever the structure came from, and a stray “30 days” from a template is indistinguishable from one you chose.

The pass only a lawyer can do

Then there is the category the model cannot touch: whether a clause does what it says where you are. The United Kingdom offers a clean illustration. Under section 31 of the Consumer Rights Act 2015, a term of a contract to supply goods is not binding on the consumer to the extent that it would exclude or restrict the trader's liability under a list of the Act's own provisions. A drafting tool will happily write you a sweeping exclusion of liability. Nothing about the sentence looks wrong. Whether it has any effect is a function of who the counterparty is and which country's law applies — and that same clause carries a different answer in the US, in Germany, and in a business-to-business deal versus a consumer one.

This is the whole reason to keep the model on drafting duty. Ask it to list the clauses whose effect it cannot determine, hand that list to a lawyer with the draft, and you have converted an open-ended review into a specific set of questions. That is a cheaper hour and a better one.

When to stop generating and hire someone

Choose a lawyer for accountable judgment

An AI contract generator is a poor final authority, and for some agreements it should not be the starting point either. Go to a qualified lawyer in the relevant jurisdiction when the contract involves equity, options, or anything that divides ownership of a company; assignment of intellectual property you or your business depends on; employment, contractor classification, or anything touching an employment relationship; anything cross-border, where two legal systems and two tax regimes are in play; personal data at scale or regulated categories of it; real property, financing, or securities; or, the general test that subsumes all of these, anything you cannot afford to lose. A person with the right duty, context, and professional accountability has to sign off.

Choose a vetted template service for volume

For routine, repeated, low-value agreements — a standard NDA, a recurring services agreement on your own paper — a template maintained by lawyers for your jurisdiction beats a generated draft. The value there is not the words, it is that someone with liability keeps them current as the law changes and has thought about the edge cases you have not. A generator has no such duty and no update cycle. Use AI to fill the template in and to interrogate it, not to replace it.

Where the generator still earns its place

Between those two lies most of everyday commercial life: the small engagement, the one-off arrangement, the deal that is currently documented in a Slack message and nothing else. There, a structured first draft plus an explicit list of what it guessed at is a large improvement over the alternative, which is usually no written agreement at all. Draft it, check the facts yourself, budget an hour of review for anything with teeth in it, and keep the editable file.

For adjacent documents, see the AI proposal generator, which produces the scope language a contract usually points back at, or the wider OpenCraft use-case library.