On Second Thought...

We’ve spent the last few years getting better at asking AI for things. We’ve learned to provide context, define the role, specify the audience, give examples, establish constraints, request a particular format, and tell the machine exactly what success should look like.

When the answer isn’t right, we refine the prompt, add more context, push back, ask for alternatives, and keep going until we get something we can use.

All of which assumes something rather important:

Maybe we should’ve asked for that thing in the first place.

That assumption doesn’t get nearly as much attention, perhaps because AI is so good at working within it.

Ask it how to improve your newsletter and it will happily give you ten ways to improve your newsletter.

Ask it how to get more engagement on LinkedIn and it will develop a strategy.

Ask it to strengthen your sales page and it will rewrite the headline, reorganize the benefits, sharpen the call to action, and suggest three new sections you didn’t know you needed.

It may do all of those things brilliantly.

But what if the newsletter didn’t have to be improved?

What if LinkedIn isn’t where you should be spending your time? What if the sales page isn’t the problem because the offer itself isn’t compelling enough? What if the productivity system you’ve asked AI to design is simply going to help you become more efficient at doing a bunch of things you shouldn’t be doing?

That’s the peculiar danger of having an incredibly capable machine at our disposal. AI can make the wrong direction incredibly productive.

There’s been enormous attention paid to AI hallucinations, factual errors, bad advice, generic writing, and all the other ways an AI response can go wrong. Those problems are real, but at least they give us something to react to. A bad answer announces itself eventually.

A good answer to the wrong question is much harder to recognize.

In fact, the better AI becomes, the harder it may get.

Imagine asking AI to help you develop a marketing campaign for a product nobody particularly wants. It can create the positioning, emails, social posts, landing page, webinar, follow-up sequence, content calendar, and ad concepts. Within an afternoon, you can be surrounded by a massive amount of competent work.

The sheer enormity of progress creates its own evidence. Look how much we’ve accomplished! Look how quickly this is coming together! Look at this campaign!

But in all that excitement, nobody stopped to ask whether the product should even exist.

That isn’t really an AI problem. Human beings have always been capable of marching headstrong in the wrong direction. We fall head over heels in love with our ideas. We confuse activity with progress. We defend sunk costs. We solve the problem sitting in front of us because questioning the problem itself would mean revisiting decisions we already made, which is a huge buzzkill.

AI simply changes the economics of all this. It makes it dramatically cheaper and faster to just keep going.

And when continuing becomes almost frictionless, stopping to reconsider becomes more important.

There’s something almost old-fashioned about that phrase.

On second thought...

It means I was about to do something and reconsidered. I had reached a conclusion and disturbed it. I was moving in one direction and something made me hesitate long enough to notice another possibility.

We tend to associate second thoughts with what happens after an answer. You think something through, arrive at a conclusion, and then reconsider.

AI makes me wonder whether we need to move the second thought earlier.

Before the answer.

Before the prompt.

Sometimes even before we decide what the problem is.

“Help me write a book” becomes, On second thought, should this be a book?

“Help me get more clients” becomes, On second thought, do I need more clients, or fewer clients who are worth considerably more?

“Help me grow my audience” becomes, On second thought, is audience size actually what’s constraining my business?

“Make this sound more professional” becomes, On second thought, who decided professional was what this needed to sound like?

“Help me make this idea clearer” becomes, On second thought, what if the ambiguity is the interesting part?

Notice that none of these questions are asking AI to produce a better version of the original answer. They’re interrupting the path to the answer altogether.

That interruption may be one of the most valuable things we can learn to do.

Every prompt contains more than an instruction. It contains a worldview, however small.

If I ask how to grow my Substack, I’ve already decided that growing my Substack is desirable. If I ask for five ways to become more productive, I’ve accepted that greater productivity is the solution. If I ask AI to make an article more persuasive, I’ve assumed that persuasion is what the article lacks.

AI generally honors those assumptions because that’s what we’ve asked it to do. We don’t usually want a calculator to challenge our decision to calculate something, and we haven’t traditionally expected software to question the premise behind the command.

But generative AI is different because it operates at the level of ideas. It can participate in the thinking that precedes execution, not merely the execution itself.

And yet much of the way we use it still resembles the old software model.

We decide what we want.

We issue the command.

The machine executes.

Then we judge the output.

Even many supposedly sophisticated prompting techniques remain trapped inside that sequence. They’re designed to improve what happens after we’ve decided what we want the machine to do.

Perhaps the more interesting opportunity lies one step earlier.

Not: How can AI help me do this better?

But: What if this isn’t what I should be doing?

There’s an old idea in business that there’s nothing quite so useless as doing efficiently something that shouldn’t be done at all.

Generative AI takes that problem to a completely different scale.

For the first time, an individual can efficiently pursue a questionable premise across dozens of disciplines simultaneously. You can research it, strategize it, name it, brand it, write it, illustrate it, market it, analyze it, repurpose it, and develop an entire ecosystem around it before lunch.

That’s incredible leverage when the underlying idea is sound.

When it isn’t, it’s still incredible leverage.

Just in the opposite direction.

This is why I’m becoming less interested in the question of how to get AI to give us better answers. Of course we should learn how to use these systems well. Of course prompting matters. Of course better context and clearer instructions improve results.

But the cost of an answer has collapsed so dramatically that perhaps the answer itself is no longer where the greatest leverage lives.

Maybe the leverage took a U-turn.

To the premise.

To the assumption.

To the definition of the problem.

To the moment before we tell the machine what we want.

Lately I’ve become more interested in what happens when we use AI to interfere with our intentions before they harden into instructions.

• Challenge this assumption.

• Tell me what I’m treating as settled that isn’t.

• What would have to be true for this to be the wrong problem?

• What happens if I reverse the premise?

• What am I not considering because I’ve already decided what success looks like?

Those questions create a very different relationship with AI. The machine stops being merely an extraordinarily capable executor and becomes something more disruptive: a way of introducing resistance into our own thinking.

That resistance matters because AI itself is removing so much of it.

We can go from idea to execution faster than any previous generation could have imagined, which of course is extremely powerful. But speed magnifies direction. If you’re pointed toward the right destination, going faster is wonderful. If you’re pointed toward the wrong one, speed merely gets you farther away before you notice.

Maybe that’s why the next stage of working with AI won’t be defined entirely by better prompting. Perhaps it will be defined by getting better at knowing when not to prompt yet.

We’ve spent the first few years of generative AI learning how to prevent the machine from wasting our time by giving us the wrong answer.

On second thought, maybe we should spend a little more time making sure we haven’t asked it to solve the wrong problem.