Something Laura Belmont and Matt Kohel said in our AI model governance webinar this week stuck with me, and I think it is worth passing along. They both highlighted the risks with using “reasonable” in AI contracts.
We see it all the time. Reasonable cooperation. Reasonable documentation. Reasonable notice. Commercially reasonable efforts. The word is intended to provide the parties with lots of flexibility. And it does work well in other kinds of contracts. But as we talked about it more in our recent webinar, it became clear that it isn’t as beneficial in AI contracts. It creates a vague standard that could be worse than leaving it out.
The context was our webinar on AI model governance provision. As we went through the three sample AI-drafted provisions, we noticed that “reasonable” kept showing up. In the cooperation provision, it sat next to “such information as vendor deems necessary.” That standard is too loose. It could be read to give the vendor a veto on what the customer needs to answer a regulator. If you are responding to a privacy authority on a 72-hour deadline, you can’t wait to find out if the vendor thinks your compliance requirement is reasonable.
Always focused on providing practical solutions, the consensus was it is much better in most instances to replace the word with something specific. Name the documents you need by type: model cards, bias audits, benchmarking, accuracy testing, data flow diagrams. Tie timelines and deliverables to regulatory requirements. In other words, find the specific standard that matters to your situation and include that.
Reasonable has a place in contracts. It’s just not a great choice when your AI compliance and governance depends on it.
