Wharton professor Ethan Mollick wrote: “The best use of AI is in your narrow domain of expertise, where you can understand its limits, and its value. That helps you become a ‘centaur,’ integrating your best human abilities with the capabilities of AI work, so that your output becomes a blend of both, like the half-person half-horse of myth.”* That logic certainly applies to contracting tools. If you understand the issues raised by your deal, you’re in a good position to craft effective prompts and playbooks and to assess outputs, re-prompt, reassess, and create effective contracts. But that’s not quite enough. You should also understand the AI itself: what it does well and what it doesn’t.

Issue-Spotting and Overcorrecting

AI also does a good job with issue-spotting. We’ve got to notice an issue to address it, but we can’t catch everything. Your AI can’t either, but it can supplement your efforts, leading to good centaur work.

AI issue-spotting, however, has a downside, at least if you rely on it too much. Most models are overcautious – particularly legal models: systems fine-tuned on legal sources. Have you ever reviewed an AI-generated e-mail or memo about a legal concern? Often the AI modifies every statement with a long list of complex, barely relevant caveats. That same AI might also give you vastly more protective contract-terms than you need, with exceptions and special clauses that bog down negotiations. The centaur process needs your judgment about what matters and whether an issue outweighs the cost of fixing it.

This AI overcaution strikes me more as a feature than a bug. You might reject nine out of ten issues your AI raises, and you might not like how it handled the tenth. But if number ten matters and you wouldn’t have caught it, AI issue-spotting has done you a big favor.

The Clause’s First Draft

AI also does a particularly good generating the first draft of a clause, from scratch or by editing existing language (from your playbook, the other side, etc.). But a sophisticated-looking clause might hide the fact that the AI didn’t really follow the prompt, and it might hide bad assumptions or failure to address context. A clause that works in lots of settings – praised over and over in the AI’s training data – might not work for your deal, and the AI might not understand.

So your part of the centaur checks for fit-to-prompt and bad assumptions. You should also ask for and review the AI’s sources of information and, where possible, check them.

Your Prompts and Playbooks

Of course, all this effective centaur work starts with prompts and playbooks that push the AI in the right direction. The AI’s long list of issues, if it produces one, could still miss what matters. So you’ve got to give it the issues you know. For example, it’s not enough to prompt an AI to draft or review an indemnity – not even after describing the parties’ relationship. Your prompt, playbook, or other instructions should ask specific questions, like, “to trigger this indemnity, would the indemnified party have to accuse the indemnitor of wrongdoing?” – or, “would this IP indemnity cover a claim about the software’s interface with a third party product?” (For these issues, see Chapter II.L of my book, The Tech Contracts Handbook.)

Fortunately, you don’t have to think through every issue on the first prompt. Keep asking; AI is patient. (Your employer might be less patient if it’s paying per-prompt or per-token, but most likely your care is worth its weight in tokens.)

Bad Data and Playbooks – Plus Sycophancy

You should also recognize that hallucination is not the only AI mistake to watch for. Errors infest both training data and the information that language models (large and small) retrieve from online and elsewhere: retrieval data. And we dig the potential hole deeper when we use playbooks or fine-tuning data with more errors – a more common problem than you might think.

See content credentials

Unfortunately, that problem may hit contracting particularly hard: harder than activities like summarizing the news, for instance. LLMs track patterns, so commonly-repeated ideas – the bulk of information under the bell curve – often shape their views. That may work for news that fits the bell curve model, with the crazier rightwing and leftwing stories on the edges. (Of course, the mainstream stories can still be wrong.) But for contracts, the available data probably looks more like a pyramid, with the really good documents and advice at the top and the bulk at the base. The AI may have no way to distinguish the top from the bottom, so if it follows commonly repeated patterns, the bottom of the pyramid will influence it more. A legal/contracts AI might do better, if it’s fine-tuned on good contracting sources, but even that doesn’t solve the problem.

AI accuracy issues get worse if we users consistently make the same mistake, because the AI may try to “please” us. It may give us outputs that reinforce our views: a problem called sycophancy. So we’ve got to keep up our own knowledge.

Divine Scripture and Best Practices

Of course, all these problem get worse when the AI delivers a bad output with all the certainty of divine scripture. Language models sometimes admit uncertainty, but nowhere near enough.

Take advantage of AI’s issue-spotting power and its speed at drafting or adapting terms. Take advantage of its other abilities, like searching through 5,000 contracts in a minute. And take advantage of the fact that, unlike you, the AI never gets tired, bored, or angry at the other side. Then question, test, re-question, cast doubt, and reassess.

* This great analogy appeared in a mere LinkedIn post, though the professor’s made similar points elsewhere. E. Mollick, LinkedIn post (Aug. 28, 2023).