AI-generated legal drafts pose hidden risks
Clients who use AI to generate draft legal documents are posing new challenges — and potential risks — for their attorneys, warns Indiana University Law Prof. Josh Kubicki, a legal tech expert.
“Most clients aren’t disclosing when they’ve used AI,” Kubicki writes in the latest edition of his Brainyacts newsletter.
“And even when they do, they often forget the more important part: what they told the AI to get the draft.”
From ordinary citizens to “sophisticated in-house counsel,” clients are increasingly using AI to create first drafts of contracts, memos, emails, issue spotters, and litigation narratives, Kubicki writes.
That could pose a serious risk, Kubicki warns.
“If a client pasted confidential strategy, facts, or matter details into ChatGPT or another public LLM, there’s a real chance they’ve waived privilege or confidentiality without realizing it,” Kubicki writes. “Many consumer-grade AI tools store prompts. Some even learn from them.”
Months later, as the case progresses, the prompt will be long forgotten, but the attorney “may still be held responsible for how it shaped the draft,” Kubicki writes.
Kubicki recommends that attorneys consider some hard and fast rules.
They include asking — every time — if AI was used, requesting the prompts, flagging potential waiver risks, and building a “process for documenting this in your intake or engagement letter.”
Clients who use AI to generate draft legal documents pose other significant challenges, Kubicki warns.
Clients may expect to save money by taking the initiative — without realizing how much additional work they are creating for the attorney or outside counsel.
The draft must be reviewed line-by-line, the prompts requested, read, and evaluated for privilege, and more time must be spent communicating with the client about “what AI got wrong,” and “why that matters,” Kubicki notes.
“That’s not free,” Kubicki writes. “That’s not even neutral. That’s additive.”
He recommends lawyers develop a clear process for identifying AI-originated work, a protocol for reviewing prompts and evaluating risk, and language to explain, “without condescension,” why review “still matters.”
Kubicki concludes with a recommended checklist.
- Was AI in any way used to draft this?
- Which tool was used?
- Do you have a copy of the prompts?
- Was confidential or privileged information entered into the AI tool?
- Are you comfortable with us reviewing, revising, or redrafting this from a risk standpoint?
- Do you understand that review of AI work may involve additional time and cost?
