AI is unusually good at helping you move in whatever direction you point it. That’s useful when the direction is right. Less so when it isn’t.

Before, friction did some of the filtering for us. Most ideas required enough time, money or technical ability that they never got very far. Some good ones died this way too.

That friction is disappearing. You can take a half-formed idea and have research, positioning, a prototype and a launch plan around it before you’ve seriously asked whether the idea deserves any of them.

So judgement has to do more of the filtering now.

And judgement is prone to self-deception. It’s also uneven. You can have a good eye for how something should look and no idea whether anyone wants it.

AI doesn’t fix this automatically. It gives your judgement leverage.

Bad judgement used to produce mediocre work slowly. Now it can produce mediocre work quickly, and the output looks professional the whole way down.

Part of the problem is that the model takes some of its posture from you.

“I think this is a great idea. Help me figure out how to launch it.”

and

“Assume I’m emotionally attached to this idea. What would make it fail, and what evidence should change my mind?”

can turn the same machine into two very different collaborators.

The first can give you customer segments, positioning, features and a launch plan. Twenty minutes later your idea looks remarkably like a company.

The second might tell you that the problem isn’t painful enough, there are easier substitutes, distribution will be expensive, or that nobody has much reason to pay.

Neither response is necessarily wrong. You asked different questions.

This is where a mirror is the useful metaphor. An LLM isn’t literally a mirror. It knows things you don’t and can notice things you’ve missed. But when the answer depends on judgement rather than fact, it often starts from the frame you give it.

If you go looking for encouragement, there is plenty available.

That makes sycophancy partly a model problem, but not entirely. Once you know how these systems behave, you can usually ask for resistance. You can tell them to attack your assumptions, argue the opposite case, look for disconfirming evidence and tell you what you probably don’t want to hear.

The more uncomfortable problem is that sometimes we don’t want that.

I keep instructions on by default telling the model to push back and stress-test whatever I throw at it. It helps. I’ve killed ideas because of conversations that exposed something I hadn’t properly considered.

But I still rush into things without examining what’s driving me. And when the machine starts digging into my reasoning, I sometimes shut it down.

Building is more fun while all the possibilities are still alive.

I suspect this is the more interesting failure mode. AI can become a very sophisticated way of giving yourself permission. Not because it forces agreement on you, but because you quickly learn which questions produce the answer you were hoping for.

Once an idea survives the harder interrogation, though, another problem appears.

There’s a difference between asking whether something could work and asking whether you actually want to do it.

The first question can be attacked. Who wants this? What already exists? What would have to be true? What evidence would make the idea weaker? There may not be a definitive answer, but reality gives you something to push against.

The second question is different.

An idea can be viable and still be the wrong thing to spend the next few years on. There is no amount of market research that can tell you whether you want the life attached to it.

AI can help expose your reasoning there too. It can point out contradictions, opportunity costs and motives you may be hiding behind. But at some point there is nothing left to optimise.

You have to choose.

That’s why I don’t think the real promise of AI is better judgement. It is leverage on whatever judgement you already bring to it.

As execution gets cheaper, that distinction matters more.

The difficult part is becoming less about turning an idea into something real.

It’s deciding which ideas deserve to become real at all.