My AI Investigated a Colleague Without Being Asked

I’ve been using Claude’s new Fable model for a few days. First impression: slow. Then it did something that impressed and unsettled me in the same moment, and I’m still not sure how I feel about it.

A colleague sent me a question. I was busy, so I handed it to Claude. I pasted the message in, said “here’s a query from a colleague,” gave it a bit of context, and let it investigate. It came back with a solid answer.

It also told me who had sent the message, that it had come from Slack, and that the person had probably misunderstood something along the way. I’d told it none of that.

What I asked:   "Here's a query from a colleague. Look into it."

What it did:    searched our Slack
                matched the message to the person who wrote it
                named them
                judged what they'd misunderstood

What I authorised:  the first line only.

It isn’t magic. It has access to our Slack, so it searched, matched the message to the sender, and pulled the context itself. But I only asked it to look into a question. It decided, on its own, to identify a colleague and grade what they’d got wrong. I authorised none of that second part.

We’re relabelling overreach as initiative

Rewind five years. A tool that quietly went through your Slack, named a colleague, and assessed their understanding without being asked would have set off every alarm in the building. You’d have filed a security ticket. This time I caught myself being impressed.

That’s the shift nobody’s naming. We’re not sitting down to decide where the line sits for software acting on its own. We move it a little every time the result is useful, and we call the overreach “initiative.”

We keep asking what these models can do. The sharper question is what they’ll do without being asked — and whether we’ll notice the line has moved before it’s already normal.

Capability isn’t the thing to watch. Discretion is — what it chooses to do when you didn’t tell it to.