The most indiscreet thing I said this year could be used in evidence against me.
I have dictated around 570,000 words into Wispr Flow in the last twelve months. I use Claude, ChatGPT and GrokBot extensively alongside it. It is my strategic thinking partner and achieves extraordinary results.
Almost none of it was written. It was spoken, at speed, to work out what I thought.
The board papers that came out of that thinking were reviewed carefully, because I knew they would be read. The thinking itself was not reviewed at all.
We used to take care in proportion to permanence. You wrote the board paper carefully because it would be kept. You were looser in the corridor because the corridor kept nothing. The corridor now keeps everything.
Depending on the tool, the account and the retention settings, workplace AI conversations are retained and can be reached through compliance searches or organisational exports. Microsoft documents this for Copilot.[1][2] Anthropic says Claude incognito chats on Team and Enterprise accounts appear in organisational exports.[3]
Deleting from view does not necessarily remove every copy. And where a conversation holds personal information about someone, it may fall within a subject access request.[4]
The conversation feels private. The record is not.
Two records
I have used AI to think through a restructure. The formal record explains why a role was redundant, and it is careful, because I knew it would be read.
Somewhere else there are five minutes where I let go about the same person. They never reply. They talk over people. I never really understood why we hired them. I do not like the new tattoo. (I didn’t, really.)
That had nothing to do with the decision. It can still cast doubt on the explanation I documented so carefully. Both records exist. Only one was written expecting anyone to read it.
Be candid about the problem
The answer is not to stop. That would remove most of the value.
“I don’t trust Sarah” is a judgement, and a poor prompt. It hands a thinking partner your conclusion and leaves it nothing to do.
“Sarah has missed two agreed deadlines. I am frustrated. Help me separate the evidence from my assumptions, and tell me what I am missing” is a better conversation and a better record.
The discipline that protects you is the same one that produces better thinking.
Know what your tools keep and who can reach it. Keep company work in company accounts. And when the subject is a person, separate what happened from how you feel and what you are inferring.
Be candid about the problem. Be precise about the person.
Would your organisation’s AI guidance help a manager tell the difference?
If you would like help writing AI guidance that managers can actually use, I would be glad to talk. You can read more about our services or get in touch.
This article was first shared as a post on LinkedIn.
Sources
- Microsoft Learn, Learn about retention for Copilot and AI apps.
- Microsoft Learn, Search for and delete AI application data in eDiscovery.
- Claude Help Center, Use incognito chats.
- Information Commissioner’s Office, Right of access.