Ask a model to draft a patent background and it will hand you something good. Complete, well organized, the field described, the problem framed, the prior approaches laid out in order. It reads like diligence.
That is the problem.
A legal instrument is not prose. In prose, completeness is a virtue. In a patent, a brief, or a discovery response, completeness is a position. Every additional sentence is something you have now said on the record, and some of what a model says on your behalf, fluently and thoroughly, is something you would never have volunteered.
I have started calling this comprehensiveness as liability. The model is built to complete the thought, fill the section, and explain the point. Legal drafting often requires knowing what not to say. Every sentence can be defensible, and you still should not have said all of it. An experienced lawyer will usually see the problem. The risk is what happens when the draft is reviewed too quickly, or by someone who does not yet know which perfectly accurate sentence should never have been volunteered.
Silence is a tool
Every litigator already knows this. You do not over-answer an interrogatory. You do not concede a contested element in a statement of facts because the paragraph reads better complete. You do not rebut an argument the other side has not made yet. Knowing when to say nothing is half the craft.
A model does not know when to be silent. It answers the question you asked, fills the section you pointed at, and keeps going until the section looks finished. In an adversarial instrument, finished is not a neutral state. It is a set of commitments, made fluently, waiting for your signature.
Where it costs you in a patent
Patent drafting has two places where thoroughness is particularly dangerous, and models tend to say too much in both. Patent drafting is unusual because being articulate about your invention can cost you the invention.
The first is the background. When a specification characterizes something as prior art, with "conventionally," "it is known that," or "existing systems do X," that characterization is an admission. Under MPEP 2129, a statement identifying the work of another as prior art can be relied on against your claims for both anticipation and obviousness, whether or not the material would otherwise have qualified as prior art at all. That kind of framing comes easily to a model because it makes the background read smoothly. But there are two important limits. A background is not automatically an admission; it depends on whether you characterized the matter as prior art, and your own earlier work does not count against you unless it independently qualifies. The exposure is real, but it is not automatic.
It is also forum-dependent. The same admission that helps an examiner, or an opponent in district court, cannot be the basis of an inter partes review (Qualcomm v. Apple, Fed. Cir. 2022 and 2025), though it can still show what the skilled artisan knew. Which forum you end up in is not something the tool can know when it writes the sentence.
The second place is how the draft describes the invention itself. Characterize "the invention" as having a feature and a court can read that as a limit on every claim, including claims where you never wrote the feature in. In Honeywell v. ITT (Fed. Cir. 2006), the specification repeatedly described "the present invention" as a fuel filter, and the court confined the claim term "fuel injection system component" to fuel filters accordingly. Its reasoning was plain: the public is entitled to take the patentee at his word, and the word was that the invention is a fuel filter. The bar for reading a limitation in from the specification is high, and describing an embodiment or contrasting the prior art will not clear it. That matters with AI-assisted drafting because models are very comfortable making broad declarative statements about what "the invention" is. "The present invention provides..." can do considerably more legal work than the model, or the lawyer reviewing for readability, may intend.
None of this necessarily looks defective on the page. No grammatical problem, no missing antecedent basis, no malformed claim. It may pass examination cleanly. If that language survives review, a hostile reader may later use the extra language as a concession.
What to cut before you polish
The fix is not simply checking whether the draft is correct. It is reviewing for what an experienced lawyer would never have put there in the first place. Before you polish what is in the draft, decide what should not be in it at all.
What is this document conceding by being complete?
What would I have left unsaid?
In the background: am I calling anything prior art that I do not need to call prior art?
In the claims and summary: is any "the invention" language a limit I did not mean to set?
Part of the lawyer's job has always been knowing when to stop writing: which fact not to volunteer, which characterization to cut because it gives away more than it earns. A model will not stop on its own. It has no reason to know that the next fluent sentence is a door it is opening for the other side. It can make an inexperienced drafter sound more experienced. It cannot give them the experience required to know which perfectly good sentence should be deleted.
I write about using AI in legal practice without surrendering judgment, privilege, or the duty of competence at The Agentic Lawyer.
Educational only, not legal advice, and no attorney-client relationship is created. Views are my own. Attorney advertising in some jurisdictions.