Arvind Narayanan is the the AI critic you may want to follow for two reasons:

01. He is not trying to be sensational by taking extreme positions

02. He is a first rate academic

His work tries to define the boundaries of what can be useful and where you have to be careful in a pragmatic way.

He has published a solid paper outlining these boundaries in legal space.

01. Information processing tasks have high clarity and high observability and are best use cases to start (categorizing requests for legal help, e-discovery)

02. Creativity, reasoning tasks are a range (spotting errors in legal filings are easier, preparing legal filings harder)

03. Predictive tasks are fraught with challenges (legal judgment predictions)

Lots of legal departments are evaluating Generative AI - and this is a good paper for them. In fact, we are doing some solid value accretive work in contract extraction space - that helps turn unstructured data into structured information for downstream applications. And our experience has been consistent with these observations.