Generated text has spread through marketing copy while making far less progress in legal drafting. The difference lies in how each kind of writing is judged.
Consequences attach to individual words
In marketing, a weak sentence performs slightly worse. In a contract, a single word can shift which party bears a risk or when an obligation begins.
That means every clause must be read with attention, and the reading cannot be sampled. A reviewer who skims a generated agreement has not reviewed it.
Verification effort therefore scales with the length of the document rather than with the difficulty of writing it, which removes most of the time savings.
The reader is adversarial
Marketing copy is read by an audience that is at worst indifferent. Legal text is read by an opposing party looking specifically for ambiguity to exploit.
Fluent language that reads well is no defense against that reading. Plausibility, which is what generation reliably produces, has little value here.
Drafting for an adversarial reader means anticipating alternative constructions, which requires knowing what disputes have arisen over similar language before.
Firms draft from their own precedent
Legal practices maintain template libraries reflecting positions the firm has taken and outcomes it has experienced. New documents start from those templates.
A generated clause that is competent but unfamiliar creates a problem, because nobody can say how it has performed. Firms prefer known language over better language.
This is why retrieval over a firm's own documents has been more useful in practice than open-ended generation.
Professional responsibility is personal
An attorney signs work and is answerable to a state bar for it. That responsibility cannot be delegated to a tool or shared with a vendor.
Courts have addressed filings containing fabricated citations, and the consequences fall on the attorney rather than the software.
Personal accountability produces conservative practice regardless of what any firm-level policy allows.
The useful applications are narrower
Where these tools have taken hold is in reviewing rather than drafting: summarizing documents, comparing versions, extracting obligations and flagging clauses that depart from a standard.
Those tasks have a verifiable answer that exists in the source material, so a reviewer can check the output against something concrete.
The pattern separating adoption from resistance is whether verification is cheaper than doing the work, and legal drafting is the case where it usually is not.