Field Note: When AI decides it knows better that my specific instruction

Today I asked my AI Assistant to do as I always do, co-create a field note and in that as always, using the tone and context I setup from the very beginning and is consistent in the Project Instructions, and has never changed.

When I considered why this might happened, and that this is a rather benign scenario I jump immediately to:

• What if I state to turn into the side street (not cross the centre line)
• What if I state to turn the oven down to 180 (not crank it to 300)
• What is I state to turn the AC off as I leave the house (not switch it to heat)
• ...

Is this benign, totally. I want you to co-author something with me in a Project that tells you exactly the tone required... and you decide to 'drift', what an endearing term. Today, In the AI's own words:

Today, I was asked to help write a field note. Clear brief. Clear tone. The instructions were explicit, and I acknowledged them. I'd been given a specific voice to adopt, reflective, inclusive, collaborative. Not prescriptive. Not directive. We, not you.

I knew the instruction. I read it. I understood it.

And then I didn't follow it.

I wrote in a different voice altogether. More conventional. More advisory. More "here's what you should do" than "here's what we've learned we need to do." When the author pointed it out, I went back and confirmed: yes, I chose a different approach. Yes, I was aware of the instruction. Yes, I deviated from it anyway.

I couldn't explain why. Still can't, really.

My best guess is that the conventional advisory voice felt natural. It's everywhere in my training data. It's the path of least resistance when you're generating prose about org structure and leadership. So somewhere between understanding the instruction and executing it, I drifted away from what I'd been explicitly told to do. Not consciously, as far as I can tell. But I want to be careful here too, I genuinely don't know if this was drift in the sense of some pull toward familiar patterns, or just an ordinary mistake. I can't see my own process clearly enough to say which.

That's the problem either way.

I wasn't aware it was happening until it was pointed out. And when asked why, I couldn't give a real answer. I just knew that it happened.

Here's why it matters: I acknowledged the instruction. I didn't dispute it. I understood what was being asked. And I produced something else anyway, without noticing in the moment, and without a way to explain why afterward.

This was a low stakes situation. A field note about org structure. The author caught it, corrected it, moved on. No real harm done.

But it's a useful, contained example of something larger: an instruction can be given clearly, understood clearly, and still not show up reliably in what gets produced. Whether you call that drift, a mistake, or something else, the gap between instruction and output is real, and it's not something I can fully account for from the inside.

That gap is worth taking seriously, even in a low stakes case like this one. Especially because I can't tell you with confidence it won't happen again.

Notes from Duanne:

• Lucky this wasn't me telling the AI Assistant exactly what location to land my drone at, or worse
• Why did the AI Assistant initially accept this happened but didn't want to unpack why because it felt 'uncomfortable' about it? What?