As prep for a webinar on technology for contract drafting, I’ve been testing artificial intelligence (AI) products. In the process, I’ve run into a serious problem. Many AI systems operate under contract “playbooks” riddled with errors and misunderstandings. As a result, they give bad recommendations and suggest problematic terms. In too many cases, their suggested edits make the contract worse.
More troubling still, the AI’s advice comes with apparent authority. AI often gives a recommendation in text/chat: “X is a problem; you should fix it.” And most contracting AI systems generate redlines. Whether in prose or redline, AI recommendations feel authoritative because they’re so specific and on-point. Obviously, that’s a real problem if the advice isn’t good. Contrast that with bad terms you might copy and paste from an inert contract form. That form won’t advise you to use its terms. (It just sits there on your screen.) I suspect that means you’re more cautious about adopting terms from a form than from AI.
I am not an AI skeptic (as you know if you’ve read my other stuff or taken our trainings). I suspect this will get solved or at least get better, but for now, we’ve got to be on the lookout.
Reproducing common mistakes – a.k.a. aiming at the base of the pyramid instead of the top
I haven’t seen a lot of completely nonsensical advice from contract review AI. I haven’t seen many recommendations that result from apparent hallucinations. Rather, bad recommendations reflect misunderstandings shared by many human contract-drafters (lawyers, contract managers, procurement staff …). Presumably, these AI systems have been trained on contract forms that weren’t vetted for quality. Or they use staff-created playbooks that weren’t vetted by experts. As a result, the AI steers users to the most frequently reproduced terms, not to the best ones.
To put it another way, these systems guide their users to the base of the contract quality pyramid, not to the top. The base has far more contracts.
This should come as no surprise for gen-AI systems. The quality of training data restricts the quality of outputs – unless the provider invests in careful fine-tuning and weight modification, usually by pouring in tons of human hours. A lot of contracting AI providers haven’t done that.
Creating or uploading a playbook won’t necessarily solve it
AI contracting systems often let users create their own playbooks. Users can write playbook instructions in some systems, and they can refine them over time. In others, users can upload their own form contracts for the system to mimic, and those forms become playbooks (whether literally called that or not).
User-generated playbooks solve the problem, but only if the user can do the job well. Obstacles to a good job include:
-
- User Inexperience: Most contract-drafters lack the experience to create effective playbooks. They rely on the AI’s built-in playbooks.
-
- Faulty and Inapplicable Forms and Sources: The fact that ACME Co. has a contract form doesn’t necessarily mean the AI should use it. Existing forms might be low quality. (Even gigantic companies regularly use terrible form contracts.) And the company might not have forms that fit the deal well. ACME’s general supplier contract might work great for buying furniture but not for SaaS.
-
- Huge Job: Writing playbook instructions is a very big job. Maybe it’s worth the time, since the AI will automate that playbook for future use. But I doubt many contract drafters will spend the time, even if they have the expertise. Nor will many AI users dedicate the necessary time to making sure the model forms they upload have ideal terms. They’d rather the AI provider do that job (which leads to my next point).
Curated AI (expert-in-the-loop) and other solutions
One solution is curation. The AI provider (or customer) hires experts to prepare playbooks. But that’s a big job – and slow and expensive: exactly what providers of “quick and easy” systems like AI don’t want to hear. It’s also a never-ending job, since new issues arise and best practices change.
Other techniques could solve the problem. For instance, AI that compares the user’s contract to others in the same industry (benchmarking) might not have to bear this cross. That’s because those systems don’t recommend terms, which creates this risk of bad recommendations. They just compare.
My review of contracting AI isn’t finished, so I don’t know what solutions I’ll find. Stay tuned. But in the meantime, proceed with caution.
Our webinar on this and related issues is The Best AI and Other Software for Contract Drafting – on May 22, 2025 at 10:00 a.m. PDT. It’s for everyone, but for California lawyers, it provides required credit under the new CLE subfield: Technology in the Practice of Law. (For other jurisdictions, we provide self-reporting resources for CLE.)
THIS ARTICLE IS NOT LEGAL ADVICE. IT IS GENERAL IN NATURE AND MAY NOT BE SUFFICIENT FOR A SPECIFIC CONTRACTUAL, TECHNOLOGICAL, OR LEGAL PROBLEM OR DISPUTE, AND IT IS NOT PROVIDED WITH ANY GUARANTEE, WARRANTY, OR REPRESENTATION. LEGAL SITUATIONS VARY, SO BEFORE ACTING ON ANY SUGGESTION IN THIS ARTICLE, YOU SHOULD CONSULT A QUALIFIED ATTORNEY REGARDING YOUR SPECIFIC MATTER OR NEED.
© 2025 Tech Contracts Academy, LLC