Simplified Supervision: Chat GPT Lessons for Legal Writing

/

If you’ve spent time on Chat GPT, you’ve figured out that it’s important to provide AI with the right kinds of instruction prompts. Otherwise, you just end up with a bunch of nonsense that looks good, but is worthless in practice.

Why is it that we haven’t learned the same lessons when working with flesh AI (a.k.a. associates)? I am frequently amazed by the supervisor instructions given to associates that simply don’t work. As a result, the supervisors end up doing the writing tasks of the associates, and later bitch about it to us.

In this article, you can learn how to get better writing out of your associates by simply applying your new insights gained from AI. In particular, you can learn how to avoid the following types of prompt mistakes:

  • ambiguous or vague instructions

  • overly long instructions

  • instructions that assume model knowledge

Ambiguous and Vague Instructions

Imagine that you told Chat GPT to fix up your writing by “avoiding long sentences”. Poor Chat GPT would be left with all sorts of unanswered questions like the following:

  • What is the maximum number of words per sentence? (Is there a minimum?)

  • Should we be focusing on an average number of words per sentence? (In other words, does a short sentence make up for a long sentence?)

  • Are there any exceptions to this avoidance rule? (For example, what if I have to quote an annoyingly long statute?)

  • If I find a long sentence, where exactly should I break it up?

  • If you would prefer that I make sentences shorter without breaking them up, what words can I delete? (And, what words should I keep?)

If you give the same instruction to your associates, don’t be surprised if they react with the same levels of uncertainty. And, when you create this uncertainty for them, you might just be muddying the waters (i.e. despite your efforts to fix things, you’ve just made them worse).

But, the funniest thing is this - I don’t think you even have a problem with physically long sentences. (If you did, you probably should have picked a different profession.) If you are like most lawyers, your real problem is that you don’t like sentences that feel unnecessarily long. To solve this latter problem, you would be better off shifting the attention of your associates to structural issues.

For example, you can fix a lot of legal sentences that feel long by merely making sure that the important actions aren’t hidden in nouns. To solve this problem, you just need to give the associates the following instruction: “make sure that you place your important actions in verbs”. See below for a demonstration.

Original Version (Main Actions in Nouns)

The denied access to this specific dataset gathered by the defendant is not only an abuse of a dominant position that infringes competition law but above all a threat to public health.

Revised Version (Main Actions in Verbs)

When the defendant denied access to this specific dataset that it had gathered, it not only abused a dominant position but, above all, threatened public health.

Lengthy Instructions

Although Chat GPT is a super-brain, it can still get distracted and make mistakes if you provide it with overly-long prompts. Not surprisingly, your associates can also get distracted and make mistakes if you ask them to focus on too many instructions.

For example, if you mark up a 10-page advice letter with detailed explanations about 30 separate mistakes and expect the associate to read your explanations and learn from them, then you are going to be disappointed. Although your associates are definitely bright and hard-working, it’s next to impossible for anyone to process and absorb so many lessons at one time.

You are much more likely to get serious improvement in their writing if you just highlight the three most important mistakes and encourage them to focus on these for the time being. Once you see improvement in these areas, you can switch their attention to three new sets of problems. Although this minimal, focused approach requires some patience on your part, you will find that it gets faster results in the long run.

Instructions that Assume Model Knowledge

You understand the risks of asking Chat Gpt to conduct a legal analysis of your client’s situation and provide some legal advice. Currently, Chat Gpt lacks the model knowledge to properly think like a lawyer.

You can see a similar problem when we ask young associates to provide clients with direct answers to their legal questions. In most cases, we receive legal advice that does everything it can to avoid providing a direct answer. (In some cases, the associates ignore the question completely.)

Strangely, I oftentimes hear supervisors (and clients) complain about this problem. And yet, nobody is asking a very obvious question. Do these associates possess the model knowledge to answer questions directly? In most cases, the answer is “NO”. They were never taught to provide such client-friendly answers in law school – our professors were more interested in us showing off our legal knowledge. And, most firms spend little time telling their young associates about how to provide a direct answer. We just kind of assume that this whole process is obvious.

But, it’s not obvious. Young lawyers struggle with providing direct answers because we’ve been told over and over that the law is complex and dangerous. They mistakenly think that their job is to talk about the law and let the client figure things out on their own. I am sure that there are plenty of more reasons, but you get my point.

If you want to get your associates to provide client-friendly answers, you need to provide instructions that address the following:

  • how to identify the client’s real question to stop the law student urge to answer general questions,

  • how to utilize the terminology of the client’s question in your answer to avoid giving an answer that sounds like a law journal article,

  • how to structure an answer that is both direct for the client AND legally safe for the lawyer, and

  • how to propose constructive alternatives when the original answer doesn’t sufficiently solve the client’s underlying problems.

Additional Tip: If you want to use the above approach, you might also want to reinforce the importance of the above steps with a vivid explanation about how pissed off clients can get when you ask them to pay for advice that isn’t useful.

Conclusion

In summary, you can get much more out of your associates if you learn from the lessons of Chat Gpt and focus on giving your associates better legal writing instructions, instructions that are (i) sufficiently specific, (ii) suitably short, and (iii) fill in any model knowledge gaps.

You can also save yourself substantial time, energy and frustration if you let us provide your associates with the right writing instructions. If you liked the above approach regarding main actions in verbs, your associates can learn similar instructions in our Clear Legal Writing course. If you would like us to give your associates the right instructions for providing clients with direct answers, sign them up for our High-Value Writing course.