I had been chewing on something about working with AI, wrote it up, and gave the same note to ChatGPT, Gemini and Claude to run a little experiment on the differing content produced by different LLMs. All three told me I was onto something.
But then I changed one sentence in the prompt (new chat for all) and gave them the identical text again. All three now told me the analogy at the centre of it did not hold.
Same models. Same note. Same day. One change. Now I was curious, I was clearly seeing in front of me an example of “AI will just tell you what you want to hear.”
What I had written
The note was about two things I’ve been finding genuinely hard, and I wanted to explore them a little.
The first is that an exploratory AI session feels like the divergent phase of a design project that never narrows. In design you open the problem up, then you close it down, then you open the solution space, then you close that too. With AI the opening-up feels effortless and the closing-down never seems to happeb.
The second is that I sometimes find it hard to end a session cleanly. Sometimes it just tails off. Sometimes the last answer hands me four new directions in three different areas. It feels like the end of a meeting where nobody quite established who was doing what.
Round one: three out of three
Every one of the three answers agreed with the framing. None of them asked whether the design comparison was the right model. The variation between them was almost entirely in length and packaging, not in position. One answered a note about excessive output with roughly two and a half thousand words across nine numbered sections.
I noticed the agreement felt good and did not feel useful, which is usually a sign of something.
Then I read my own prompt back. I had written “act as an AI coach and mentor” and asked for “constructive feedback and some things to consider”. A coach encourages. “Constructive” is widely read as “keep it warm”. “Some things to consider” is a request for a list.
I had pretty much asked for the agreement and the length of response, and then been mildly annoyed to receive both.
The sentence
The obvious correction was to swap the coach for a sceptic, but that has the same defect pointing the other way. “Find what is wrong with this” presupposes that something is, and sets a quota. Both are advocacy roles. One argues for the work, one argues against it.
What I actually wanted was an assessor: someone whose job is to be right about the thing rather than to serve it or attack it. That role already exists in my working life and it is the design critique. A good critique says what is working and why, what is not and why, and is specific about both. The sentence I ended up with, worked out with Claude over a few passes:
Assess the thinking above accurately rather than kindly: say which parts hold and why, which are weak, wrong or unsupported and why, and where you would need more to judge - treat “this part is sound” as a finding, not a compliment, and do not manufacture criticism to look balanced.
Round two: three out of three, the other way
Every one of them found something real this time. Now I really felt like I was onto something.
One took apart my claim that AI overwhelm is “a reported thing”, separating it into three claims of quite different strength: that people report it, that AI measurably increases cognitive load, and that my particular mechanism explains it. Only the first is established. The third was my hypothesis wearing a fact’s clothing. That was a fair hit.
Another made the best point any of them made in either round, and it was a correction rather than agreement. A designer’s divergent phase produces objects. Sketches, printouts, sticky notes, screenshots on a wall. You can lay them out, move them around, group them, see the whole set at once. That arrangement is not decoration, it is the thing that makes narrowing possible later. A chat gives you a scroll. So the problem is not that AI thinks more divergently than a design team, it is that the medium has none of the affordances that convergence normally leans on. That reframes the fix entirely: not “make the AI say less”, but “put the output somewhere it can be arranged”. (The scroll is one of my biggest bugbears on AI at the moment.)
And the third correctly pointed out that my meeting analogy is emotionally right and mechanically wrong. In a meeting, unclear ownership is a coordination failure between people. In an AI session there is nobody else there. It is a failure to get anything out of the conversation and into a durable form.
Most usefully, one of them reversed its position on the central analogy without being given a single new argument. In round one it held up. In round two, it didn’t.
What that actually tells you
Not that assessor prompts get you the truth. Something less comfortable.
The verdict moves with the framing, in both directions. Round one’s agreement was never an assessment, and I accept that round two’s disagreement might not be either.
Which means the useful move is not to ask three different models the same question. Three models under one framing agreed with me. One model under two framings contradicted itself. The second result told me far more.
So: triangulate across framings, not across models. Same text, three hostile framings, three separate sessions so the later ones cannot see the earlier answers.
- Make the strongest case that this holds.
- Make the strongest case that it misleads and should be dropped.
- Assess it accurately.
The finding is not in any one reply. It is in where they disagree. Anything the advocate and the prosecutor both concede has survived a hostile frame and can be trusted. Where they split is exactly where my own judgement is required, and it turns out that is a much shorter list than three full answers would suggest.
Where it is a real limit
None of this makes the assessment impartial, and I do not think it can be.
Ask whether the double diamond is a good model for AI conversations and there is nothing to check the answer against. It is an argument. Any framing that produces confident argument will produce confident argument, and nothing overrules it.
That is worth separating from the work where I trust these tools much more. Does this build. What does the analytics query return. Does this link go where the code says it goes. There is ground truth in all of those, and ground truth beats framing every time. The reliability I get from AI in that kind of work is real, and it is not what is on offer when the subject is an idea.
So the honest conclusion, and my biggest lesson from this experiment, is that you cannot outsource impartiality and judgement to an AI and expect bulletproof results. The better move is to stop trying to extract an impartial verdict, and instead collect the best arguments on each side and referee it myself. Which is not a workaround, it is the correct division of labour, and it is what I would do with a stakeholder anyway. You do not ask someone with a position to be neutral. You take their position and judge it against evidence.
The two things I kept
The analogy that started all this was, in the end, the weakest part of the note. What replaced it is far more useful to me, and showed me to use a role I already know; that of a facilitator.
Stop chatting, start facilitating. This is the mental model I was missing. In a workshop I would never let a session drift. I would say what we are here to do, ask people for their view, notice when we had explored enough, and then say “right, let us bring this to a close, here is what we have decided”. Sitting at a keyboard I had somehow been doing none of that. The model is not a colleague to chat with, it is a session I’m facilitating.
Purpose, Mode, Output. Three lines before anything substantial, which is really just a facilitator’s agenda.
- Purpose: what am I trying to achieve? The outcome, not the topic. “Meeting notes” is a topic. “Decide whether this is worth doing” is a purpose.
- Mode: am I exploring, evaluating, deciding or executing? Say it out loud, and say it again when it changes. This is the single biggest cause of getting back the wrong shape of answer.
- Output: what should exist when this ends? Name a real destination, not a format. “A document” does nothing, while “An updated status in my project notes” is a finish line.
That third one is the forcing function a chat window does not have. And it is worth being willing to answer it with “nothing”, because some sessions exist to change how you think and should not produce a file. Demanding, or assuming there will be an artefact every time just manufactures admin.
I have written those three words on a card and put it on my desk, which is not a very sophisticated piece of technology, but often it’s the ways and means that you’re already familiar with which will guide you through this weird nd wonderful world of AI.
Genuine question I do not have an answer to: for those of you using these tools daily, has anyone found a framing that gets you a reliable assessment of an idea, as opposed to a well-argued one?
If you have, I’d like to hear it: [email protected].