Sometimes it is better to just get out of the way and let AI get on with it, without your ‘help’.
I know most AI guidance says the opposite: steer it, shape it, stay in the loop (or my preferred version – guide the loop). I usually agree. But I’ve noticed there are times when that advice breaks down.
When you should intervene and be the human guiding the loop
When I use AI in my core work (to support a literature review, for example), I intervene constantly. It suggests search strategies; I rewrite half of them because I can see what’s missing. It proposes thematic framings; I reject the ones that flatten important distinctions. It surfaces candidate sources; I know which bodies of work it’s overlooked.
This works because I can diagnose why a result is weak, not just feel like it is. My interventions aren’t random personal preferences, but solid corrections grounded in knowledge of what good looks like in this domain.
Most people’s experience of AI looks something like this. You steer, the output improves, and the lesson seems obvious: stay involved.
And when you shouldn’t…
I’m currently living in Spain. Our boiler keeps tripping. I asked ChatGPT to help me diagnose it. I know just about enough about UK heating systems to have some opinions, so I started steering the AI’s troubleshooting: focus on the thermostat, check the pressure valve, try this, try that. Confident, specific interventions grounded in what I felt sure was transferable knowledge.
Except what I didn’t know is that Spain has a potencia system, a contracted electrical capacity limit that governs what your home can draw at any given moment – there is no UK equivalent so I had never heard of it.
Why is that important? Because it meant that every time I steered the AI towards my preferred explanations, I was narrowing the search space around an analogy from a system that didn’t transfer (and away from diagnosing the actual fault!). My interventions weren’t helpful. They weren’t even neutral. They were actively wrong.
Now at the time, none of this felt like contamination; it felt like I was being helpful. That’s what made it dangerous. The first (un-steered) responses had been closer to the truth. I made them worse by getting involved. Eventually, I re-read ChatGPT’s first reply and started from scratch asking it to walk me through how the system worked before giving me its diagnoses (if anyone is interested – the issue is that in Spain if you have low potencia, turn on 2-3 appliances at the same time and everything stops working!)
So, if I don’t know the domain, I should just trust AI?
At this point it might sound like I’m arguing that non-experts should hand everything to the AI and walk away. I’m not.
The real issue is not whether to intervene. It’s what to intervene on and how.
- When you have domain knowledge, steer the content: what’s included, how it’s framed, what really matters.
- When you don’t have domain knowledge, steer the process instead. Ask the AI to explain its reasoning. Ask it to surface its assumptions. Ask it what questions you should be asking it. Use it as a scaffold for learning the domain before trying to shape the output.
This problem pre-dates AI, as anyone who’s sat in a meeting where the most senior person confidently steered a decision they didn’t fully understand, will know. But AI makes the feedback loop faster and the consequences harder to ignore.
Deciding how best to guide the loop
Before you next intervene, three questions.
- Do I know what “good” looks like in this specific domain?
- Can I explain why this output is wrong, or do I just feel that it is?
- Do I understand the system well enough to know what I’m steering toward?
If yes: steer hard. Your judgement is the most valuable thing in the process.
If no: steer the learning, not the output. Ask for explanation before asking for changes. Let the AI surface the terrain before you start redrawing the map.
The uncomfortable truth is that most of life is in the messy middle ground of the two – partial knowledge, enough familiarity to feel confident, but not enough to be right.
So next time you think you are helping to nudge your favourite AI chatbot to a better result, but it keeps giving you slop, stop and think. Does your steering come from a place of expertise or should you take a step back until you understand things better. Maybe AI can help you do just that, but that’s another conversation!
