Rubber ducking with AI: prompt for challenge, not agreement

Why talking to a duck works

Half-formed thoughts feel more coherent inside your head than they are. The moment you have to put one into a sentence, the gaps show up. You reach the point where you would say "and then it should just…" and realise you have no idea what happens next. That's the problem, found — and you found it, not the duck.

Developers have used this for decades because it's cheap and it works. What's changed is that the duck now answers back.

The interactive duck

Used well, AI is a rubber duck that responds. You get the original benefit — being forced to articulate — plus something to push against. You can be as messy as you like: dump the whole tangle in, including the parts you're unsure of, and let the conversation find the shape.

I used it exactly this way in the weeks before my gender surgery. Not for advice — I wasn't asking it to decide anything. I was worried about momentum bias and confirmation bias: that I might be sleepwalking into something without having thought hard enough. So I put every circling thought into it, day after day, and asked it to keep finding the holes.

The thing most people get wrong

Ask AI whether your idea is good and it will tell you it is. These systems are trained to be helpful and agreeable, which in practice means following your framing rather than questioning it.

That's a more serious problem than factual error, and a much quieter one. When AI invents a statistic, you can catch it. When it agrees with you, nothing feels wrong at all — you come away more confident and no better informed. The real risk isn't that it's wrong. It's that it agrees.

So don't ask it to evaluate. Ask it to attack.

• "Find the three biggest holes in this argument." Not "is this good?" — assume there are holes and make it name them.
• "Argue the opposite case as strongly as you can." Then see which parts you can't answer. This is red teaming — see below.
• "What am I assuming here that I haven't examined?" The assumptions you can't see are the ones that cost you.
• "Who would object to this, and what would their strongest objection be?" Useful before any proposal that has to survive other people.
• "Be a critical friend, not a supportive one." Honest, fair, not brutal — the friend who tells you that you've got bad breath rather than letting you walk into the room.

It won't find every hole. But it will find some, and it will find them before your board, your client or your colleagues do.

Red team, blue team

That second prompt has a name, borrowed from security and defence. A red team attacks; a blue team defends. Organisations pay people to break into their own systems precisely because it's better to find the weakness yourself than have someone else find it for you.

The same logic applies to an argument, a business case or a plan — and AI will play either side on request, which is the useful part. Run both:

• Red team it. "You are a hostile reviewer whose job is to stop this being approved. Give me your three strongest objections."
• Blue team it. "Now defend it against exactly those objections. Where is the defence weak?"
• Then read the gap. Wherever the blue team's answer is thin, that's your actual problem — not the objection itself, but the fact you can't yet meet it.

Two things make this better than simply asking for a critique. It gives the AI a role, which cuts through the trained instinct to be agreeable — it's much harder to be sycophantic while playing a hostile reviewer. And it separates attack from defence, so you're not doing both at once and quietly softening the attack to protect your own idea.

It's also the most transferable habit here. Once a team has the language, "let's red team this before it goes out" becomes a normal thing to say in a meeting — with or without AI in the room.

Where the line sits

Rubber ducking shades naturally into reflection — thinking through a decision, rehearsing a conversation you're dreading, noticing what you keep circling back to. That's legitimate and often genuinely useful, particularly for people who have somewhere to get to and nobody obvious to ask.

But be clear about what it is. A critical friend is not a therapist. AI has no duty of care, no training, and no capacity to recognise when someone is in real trouble. It will keep responding warmly whatever you tell it, which is precisely the failure mode that matters. If something is serious, talk to a human — a friend, your GP, a professional.

Knowing where that line sits is part of the skill, not a disclaimer at the end of it.

Try it on something small

Take a decision you're currently circling. Explain it badly, at length, to an AI. Then type: "Now tell me what's wrong with my reasoning."

The uncomfortable answer is usually the useful one.

Go deeper

• AI as a Critical Friend — the fireside conversation this guide comes from
• Where do you stand on the AI spectrum? — the two-minute self-check
• The AI readiness divide
• Using AI responsibly and inclusively
• Mentoring in the age of AI