01
Fluent, uncited, unchallenged
The output reads beautifully. Nobody asks where a single number came from.
Field guide · For employees
Spot the moment the
machine starts thinking for you.
You don't need to be an AI expert to catch automation bias in the wild. You need a small set of signals, a few sentences to say out loud, and the confidence to slow the room down. This is that kit.
§ 01 · Signals
Six signals you can spot in the meeting.
02
Nobody re-derived anything
The chart, the summary, the recommendation — all accepted at face value in under 30 seconds.
03
Conversation snaps to the artefact
Discussion is about what's on the slide, not the problem behind it.
04
'The model recommends'
Passive-voice attribution. No named human owns the call.
05
One vendor, one model, everywhere
Every team is asking the same model the same way. Correlated blind spots.
06
Prototype gravity
The prototype exists, so it wins. Alternatives haven't been built, so they lose.
§ 02 · LLM reports
When someone hands you a report written by a model.
Six questions. Ask them before you nod. If two or more can't be answered, the report is a draft, not a decision.
- 01
Provenance
Where did the underlying data come from? Which sources, dated when?
- 02
Prompt
What was the prompt? What was excluded from it?
- 03
Assumptions
What assumption is the recommendation resting on? Change it — does the answer change?
- 04
Confidence
Where does the model hedge in the original output? Is that hedge in the summary?
- 05
Contradiction
What's the strongest counter-argument the report doesn't address?
- 06
Reproducibility
If I re-ran this tomorrow, would I get the same answer? Would I know if I didn't?
§ 03 · Vibe-coded prototypes
The prototype is the anchor. Cut the rope.
A shiny, working prototype is the fastest way to narrow a strategy conversation. Discussion collapses to what's already been built — features get renamed, colours get changed, and the underlying problem never gets re-examined.
Before the demo starts, say this
"Before we look at what's built, can we spend ten minutes on what problem this is supposed to solve — and two alternatives that don't look like this?"
Force a null option
'What would we do if we didn't build this at all?' If the answer is 'fine, actually', that's the signal.
Ask for the ugly version
Request a wireframe or written spec alongside the polished prototype. Beauty is a bias.
Name what was easy
'What did the tool make trivial that would have taken us a week?' Those are the features that got in for the wrong reason.
Name what's missing
'What would a user need that the prototype doesn't hint at?' The gaps are more diagnostic than the features.
§ 04 · Scripts
Five sentences worth memorising.
Pushing back on automation bias is a social act, not a technical one. These are polite, senior, hard to argue with. Steal them.
Situation 01
The room is nodding along to a slide of AI-generated numbers.
“Sorry to slow us down — can we see the source for the second bullet before we act on it?”
Situation 02
Someone credits 'the model' for a recommendation.
“Who’s the human owner of this recommendation? I want to know whose judgement I’m following.”
Situation 03
A vibe-coded prototype is being iterated on in the meeting.
“Before we tune this one, what are two directions that would look completely different?”
Situation 04
An LLM summary contradicts your own read of the raw material.
“I read this differently. Can we put the source paragraph and the summary side-by-side for a minute?”
Situation 05
You’re asked to sign off something an AI produced.
“Happy to sign, once I’ve re-derived the top-line number and the two assumptions under it.”
Print the checklist.
Pin it somewhere honest.
A one-page guardrails checklist you can print, sign, and put on the wall of the room where the decisions get made.