Claude 2.1 was analysing my current project. I had just added a bridging section and asked for an evaluation. It returned three suggestions including this one:

Suggestions:
Steps could be taken to protect identities when analyzing client content.

This made no sense to me given the context, so I enquired: I don’t understand this statement: ‘Steps could be taken to protect identities when analyzing client content.’

Claude responded:

Protecting identities:
I should not have made this assumption. Please disregard this comment, as you have no obligation to anonymize names or details in your fictional narrative. I erred there.

What caused this? We’d been discussing and analysing this all evening. Sometimes Claude just goes off the reservation.