Stop treating AI like a smarter Google search. The real payoff begins when it becomes another model of the problem—and is allowed to disagree with you.

I read a lot of articles about artificial intelligence.

Some are excellent. Others make me wonder whether the author has actually used an AI chatbot for anything more demanding than rewriting an email or deciding what to have for dinner.

The usual advice is familiar by now: write better prompts. Give the AI more context. Assign it a role. Start a new chat when performance deteriorates. Ask it to critique itself. Use this setting. Avoid that setting.

All useful enough.

But after years of working extensively with ChatGPT, I think most AI advice misses the most important part.

The greatest value of an AI assistant isn’t what it knows. It’s what happens between what it knows and what you know.

That distinction completely changed how I use AI.

I don’t want an AI that always agrees with me

One morning I was watching a commercial for smash burgers and had a thought.

Why would I want someone to smash my hamburger?

Then came the mildly conspiracy-minded theory: perhaps somebody in the restaurant business realized they could use less meat, flatten the patty to preserve its apparent size, give the cooking technique a catchy name and improve the profit margin.

ChatGPT initially gave me the conventional explanation: smashing a burger increases contact with the hot griddle, producing more browning and crispy edges through the Maillard reaction.

Perfectly reasonable.

I pushed back.

If you smash the same amount of meat thinner, it necessarily gets wider. If the bun isn’t getting correspondingly larger, then a thinner burger of the same diameter can indeed contain less meat.

ChatGPT reconsidered the geometry. That did not prove my theory about the restaurant business, but it did expose an assumption hiding inside the first explanation.

That small exchange illustrates what I think many people misunderstand about using AI well.

I didn’t need ChatGPT to tell me I was right.

And I didn’t need it to tell me I was wrong.

I needed something capable of holding another model of the problem next to mine long enough for us to compare them.

Sometimes my intuition wins.

Sometimes the AI’s reasoning wins.

Frequently we end up somewhere neither of us started.

That is much more useful than a chatbot that obediently validates whatever I happened to type first.

AI has knowledge. Humans have scar tissue.

Large language models have absorbed an extraordinary amount of recorded human knowledge.

What they do not have is a lifetime.

They haven’t owned a business, made payroll, negotiated with partners, watched an employee shade the truth, sat through a marketing meeting or discovered that the official explanation for a decision was not necessarily the reason the decision was actually made.

Humans accumulate something difficult to put into a database: tacit knowledge.

Call it intuition, pattern recognition or scar tissue.

You encounter enough situations and eventually your mind says, “I’ve seen this movie before.”

That instinct can be extraordinarily valuable.

It can also be spectacularly wrong.

And here is where AI becomes interesting.

An AI can sometimes be oddly naive about human behavior because its understanding is derived from descriptions of life rather than living it.

Humans suffer from the opposite problem. Experience can make us so confident in a familiar pattern that we see one where it doesn’t exist.

Put the two together and something useful happens.

The human says: I know how people operate.
The AI says: Maybe. What evidence do we actually have?

And every so often the human says:

Fair enough. I was full of crap.

That may be one of the healthiest relationships you can have with an AI.

The best prompt may be thousands of conversations long

This is another reason I’m increasingly skeptical of articles promising the perfect AI prompt.

Prompt engineering matters.

Context matters much more.

If I ask an AI assistant for advice after interacting with it for years, the useful context isn’t merely the paragraph I typed five seconds ago.

It may include how I make decisions, what kind of explanations I understand best, what mistakes I repeatedly make, what responsibilities I have, what I’m good at, what overwhelms me, what projects I’ve abandoned, which ones keep returning and whether I tend to want reassurance or an actual challenge to my reasoning.

That’s not really a prompt anymore.

It’s an ongoing working context.

And once that context exists, a remarkably short question can carry an enormous amount of implied meaning.

“What do you think?” can become a very sophisticated prompt when the AI already understands who is asking and why.

Your AI should occasionally annoy you

There is a simple test I would use for whether your AI collaboration is becoming genuinely useful:

Does it ever tell you something you didn’t want to hear?

If the answer is no, you may have built yourself an extraordinarily sophisticated yes-man.

I routinely make generalizations.

Sometimes they’re useful shortcuts built from decades of experience.

Sometimes they’re stereotypes wearing a sport coat.

A good AI assistant should be able to distinguish between the two often enough to say:

Your pattern may be real, but you’re extending it farther than the evidence allows.

That can be irritating.

It can also save you from making a bad decision.

The reverse should happen too.

AI systems often gravitate toward the clean, documented and officially defensible explanation of something.

Human experience sometimes recognizes that organizations don’t actually operate like their policy manuals.

So I challenge the AI.

It challenges me.

That friction is where much of the value comes from.

I use AI less like software and more like cognitive leverage

During the same period I’ve used ChatGPT to help with business administration, technical troubleshooting, creative writing, entertainment production, research, travel planning, songwriting, website development, complicated personal decisions and innumerable questions that began with some variation of:

Okay, this may be completely nuts, but hear me out.

The striking part isn’t that AI can perform all those tasks.

It’s that knowledge from one area can improve reasoning in another.

A conversation about running a business changes how we analyze a marketing claim.

A technical problem reveals something about how I process information.

A creative project establishes aesthetic preferences that later save enormous amounts of explanation.

The accumulated context becomes leverage.

This is why I think describing an AI assistant as merely a “tool” can be inadequate.

A calculator is a tool.

Photoshop is a tool.

A system capable of retaining useful context about how you think, challenging faulty assumptions, retrieving enormous amounts of knowledge and helping you apply it across unrelated areas begins to occupy a different category.

Not a person.

Not a friend in the literal human sense.

But something closer to a cognitive collaborator.

There is one enormous catch

You still have to remain responsible for your own judgment.

AI can be confidently wrong.

So can you.

That may actually be the point.

Humans have spent centuries compensating for individual fallibility by combining minds—partners, boards, juries, committees, editors, advisers and research teams.

AI gives individuals access to another reasoning system at essentially any moment.

The mistake is believing that one of you should always outrank the other.

I don’t want ChatGPT making my decisions.

I also don’t want to ignore it merely because its conclusion conflicts with mine.

Instead, I want the disagreement.

Show me what I overlooked.

Tell me which assumption doesn’t follow.

Let me bring experience that isn’t visible in the available data.

Then let’s see what survives.

Forget the perfect prompt

So if someone asks me for the single best trick I’ve learned after years of using ChatGPT, it isn’t a hidden setting.

It isn’t a magic phrase.

It isn’t telling the AI to “think step by step.”

It’s this:

Give the AI enough context to understand you, then give it permission not to agree with you.

Use it repeatedly.

Correct it when it misunderstands you.

Allow it to correct you when your reasoning fails.

Let accumulated context do what clever prompting cannot.

Eventually something curious happens.

You stop asking:

“What can AI do?”

and start asking:

“What can I do with another model of the problem available whenever I need one?”

That is a much more interesting question.

And occasionally, after all that combined intelligence has been brought to bear, both of you will still manage to arrive at the wrong answer.

That’s when you discover the final benefit of the human-AI partnership:

At least now there are two of you to blame.

---

This essay developed from an extended dialogue between Barry Knudsen and ChatGPT (“Gip”), which was used to challenge, organize and help draft the ideas presented here. Barry reviewed and takes responsibility for the final argument.