The Pitfalls of Using AI in Legal Drafting

AI drafting tools fail in predictable, recurring ways. Knowing the pattern is most of the defense — here's what actually goes wrong, and why it keeps happening even at careful firms.

1. Fabricated or misquoted case citations

This is the most publicized failure mode, and for good reason: generative AI models can produce case names, citations, and quoted holdings that sound entirely plausible and simply don't exist, or that misstate what a real case actually held. This happens because the model is generating text that statistically resembles legal writing, not retrieving verified facts from a legal database.

2. Confidentiality exposure through casual tool use

An associate pasting a client's deposition transcript or settlement facts into a consumer AI chat tool to get a quick summary may be creating a confidentiality problem without realizing it. Depending on the tool's terms of service, that data may be logged, retained, or in some cases used to improve the underlying model. This is a quiet risk because nothing about it looks like a mistake in the moment — it just looks like someone being efficient.

3. Tone and register drift

AI-drafted language often defaults to a generic register that doesn't match how your specific court or opposing counsel expects a filing to read — sometimes overly formal, sometimes surprisingly casual, and often inconsistent within the same document if multiple sections were drafted separately. Left uncorrected, this can make a filing read as obviously AI-generated, which invites exactly the scrutiny you don't want.

4. Stale law

AI models have a training cutoff, and legal rules change after that cutoff constantly: statutes get amended, cases get reversed on appeal, procedural rules get updated. A model can confidently cite a rule that was accurate as of its training data and is no longer good law today.

5. Over-reliance without supervision

The subtlest pitfall isn't a single bad citation — it's a gradual shift where an experienced attorney stops reviewing AI output as carefully as they'd review a junior associate's work, simply because the output usually looks polished. Model Rule 5.3 treats AI-assisted work the same as work product from a non-lawyer assistant: it requires supervision, not just spot-checking when something looks off.

6. Formatting and jurisdiction mismatches

A model trained broadly across jurisdictions can produce a pleading that's substantively sound but formatted for the wrong court — wrong caption format, wrong page limits, missing a certificate required by local rule. These are easy to miss because they don't affect the argument's persuasiveness, only its procedural acceptability.

How firms actually avoid these pitfalls

• Treat every AI-suggested citation as unverified until independently confirmed in a primary source
• Set clear rules on what client information can and can't be entered into which tools
• Keep the same supervision chain for AI-assisted drafts as for any associate's work product
• Build jurisdiction-specific templates rather than relying on the model's general knowledge of court rules
• Periodically audit actual usage across the firm, not just what's written in the policy

For a full breakdown of how these risks map to the Model Rules specifically, see our guide on AI ethics rules for attorneys. If you'd rather have someone map these risks against your firm's actual current practices, that's exactly what our AI ethics and compliance audit does.