Outsource Your Thinking to AI, and It Will Replace You

Head of AI at Microsoft, Mustafa Suleyman, has just recently shared a terrifying prediction: within the next 12 to 18 months, most, or even all, of the white-collar work will be fully automated by artificial intelligence. Mr. Suleyman is not the first to share such an assumption (yes, I’m talking about Dario Amodei and his 5-year AI sentence to the world), and definitely not the last. But let me share a controversial take: AI, whether in 18 months or five years, won’t be able to replace white-collar workers. But if we stop thinking, it will.

And I’m not saying it won’t be integrated into most processes, or that it can’t help people with their workload. This is already happening. But replacing human thought, eliminating human input, is that one task AI is just not capable of (unless we as human beings replace our decision-making engine with AI, that is).

Why human thinking can’t and shouldn’t be outsourced

There are three specific reasons I keep coming back to whenever someone tells me AI is about to replace everyone. Each one is something human brains do as a matter of course, and each one is something AI, no matter how advanced, cannot reproduce.

AI can’t have the context that a human brain has

When a person makes a judgment, they are not, like a language model, just predicting the next word based on general probability. They are drawing on a full body of lived experience encoded across multiple systems at once: our personal memories, our learned habits, and the emotional stakes of those moments. This collective past allows the brain to constantly guess what will happen next, updating its view only when reality proves it wrong. Context, in other words, is the substrate every perception is built on.

A large language model works differently. Its weights come from a finite corpus of text it was trained on once and no longer has access to. It has no perception loop, no prediction error from a world it has to live in. The “context” available to it is whatever happens to be in the prompt window. It has never walked into the room before. So when we make a judgment that depends on knowing what is normal here, what came before, and what is at stake, we are running predictions against a lifetime of bound experience – personal and of generations that came before us. The model is pattern-matching across text databases. Two very different operations, and no scaling of context windows closes the gap.

Knowing the name is not the same as knowing something

This wisdom is coming from Nobel Prize-winning physicist Richard Feynman (or from his father, to be more specific). He believed that simply memorizing labels or definitions doesn’t mean knowing the topic. Real understanding comes from experience and comprehension.

Charlie Munger, an American businessman and investor, suggested a term for such surface-level learning — “Chauffeur Knowledge”. The anecdote behind this is about an actual chauffeur of physicist Max Planck, who heard his lecture on quantum mechanics so often that he learnt it by heart. He then offered to deliver it instead of Planck in Munich, which he successfully did. But when it came to questions from the audience, he, not surprisingly, couldn’t answer. And that’s the difference between shallow “chauffeur” knowledge and comprehensive, earned through years of work and struggle, Planck knowledge.

Today’s large language models are the most sophisticated chauffeurs we’ve ever built. They can recite quantum mechanics, summarize philosophy, draft a contract, and rhyme some words. But something that requires experience rather than recognizing patterns, like knowing why a brand position that worked in one market will inevitably fail in the next one, is unique to humans, and something that actually drives decisions.

But understanding why a brand position that worked in one market will inevitably fail in the next one can’t be gained by simple pattern recognition. Experience is unique to humans and is what actually drives decisions.

Still, the main difference between us and AI is not data processing — it’s ownership

An ability of people (or at least some of us) to take responsibility for the actions, own the consequences, the shame, learn from them, and do everything possible for the situation not to happen again, and therefore for us not to feel the agony of another failure — this is what makes our thinking and our decision-making so uniquely different from artificial intelligence.

AI doesn’t feel shame. Or guilt. It doesn’t lose sleep over a decision that turned out badly. It doesn’t carry the memory of a deal it lost, a hire it pushed for that didn’t work out, a strategy it defended in a board meeting that gradually destroyed value two years later. The model gives an answer, the answer turns out wrong, and the model moves on to the next prompt as if nothing happened (because for them, nothing has happened). There is no internal cost to being wrong, and so there is no pressure to be right next time – something that actually pushes humans to evolve and improve.

Ten years ago, I worked in venture capital, analysing early-stage startups, and we looked at autonomous driving deals. Even then, robo-taxi prototypes were statistically safer than human drivers. The technology worked. It still works in 2026. So why are our cities still not flooded with self-driving cars? Because the responsibility of an autonomous vehicle is not clear enough. When there is a crash, the legal system genuinely doesn’t know who to blame — the person who flipped the switch, the manufacturer, or the software developer. Until that question has an answer, we are stuck. The technology has run ahead of the accountability framework that should be there to absorb it.

How not to lose our minds (and jobs) in the age of AI

While I truly believe AI can’t replace human thinking, I can also see the pattern of people falling for the temptation to get a shortcut to the results. For the first time in history, you don’t really need a brain to get an answer. You just need to type a question. The friction that used to force us to think has been removed, which sounds great, until you realise the friction was the whole point. It was how the brain got built in the first place.

And here is where the job question also comes back in. AI won’t replace white-collar workers as a category. But it will replace the specific people who outsourced all their thinking to it. If you trade judgment, context, and ownership for fast answers from a chat window, you stop being the person who brings something to the table. At that point, your employer doesn’t really need you. They can talk to the model directly.

So, keeping your job in the age of AI and keeping your mind in the age of AI come down to the same question. And below are four habits you need to address it.

AI should increase curiosity

The effect AI has on your brain depends entirely on how you use it. If you are curious and willing to dig, AI becomes the best learning companion anyone has ever had. Better than Wikipedia, better than any book, because it can explain the same idea five different ways until one of them clicks, and follow you down any rabbit hole you want. Or you can prompt it to help you retain information better – like practicing spaced repetition until you truly get the concept.

For someone who isn’t curious, AI does the opposite. It takes over the thinking they were supposed to be doing themselves. They get the output without the work, and the work was the whole point. Over time, that muscle atrophies. After the Industrial Revolution, when machines made physical labor less prevalent, we built gyms to give our bodies what they needed to challenge and maintain themselves. And that’s what we’ll eventually need: mental gyms to keep our brains fit.

In 2025, MIT Media Lab ran a study where 54 participants wrote essays over four months, split into three groups: ChatGPT, search engine, and brain only. While they were writing, the electroencephalogram was tracking their brain activity. In the end, the ChatGPT group consistently showed the lowest neural connectivity and produced essays that teachers described as “soulless”. The brain-only group, in turn, had the strongest brain activity throughout.

But the most telling part came in the final session, when the researchers swapped the groups. The brain-only people, once handed ChatGPT, outperformed everyone. They had built the thinking muscle first and then used AI on top of it. The other group could not catch up, because they had skipped the part where the muscle gets built.

To learn is a verb, so it can’t be passive

For learning to stick, theory has to be combined with practice. Reading about an idea is one form of input; doing something with that idea is another. Tapping, answering, applying, getting something wrong, and trying again — these activities pull in different parts of your brain than passive consumption does, and they keep you focused on the material. 

Socrates was already onto this thousands of years ago: he never simply handed his students the answer, but forced them to produce it (often a wrong one) first, and only then corrected it. And I myself experienced the benefits of this philosophy as a math major: we only got half credit for memorizing a theory, while the other half required proving it. This focus on fundamental understanding is what allowed us to better remember and apply the concepts.

Teaching a subject takes your knowledge to the next level, making it harder to fall into the psychological trap called the illusion of competence, and providing a better understanding of the topic. When you need to explain an idea clearly and do so several times in several different ways, there is no way for you not to understand and remember it yourself. Richard Feynman once noted with good reason: “If you can’t explain it simply, you don’t understand it.”

The numbers back this all up. In safety training studies, active learners retained 93.5% of information compared to 79% for passive learners (Engageli, 2025). The broader Learning Pyramid from the National Training Laboratories in Bethel, Maine, shows the same pattern at every level:

Data without reliable sources is gossip

Yes, people should question the answers they get from AI. They should ask for sources, and when there are no sources, they should be sceptical.

But there is a bigger problem behind all this, and it’s the fact that validating AI’s output is getting harder and harder. Models cite sources that don’t exist, they paraphrase real sources in ways that distort the meaning, they mix accurate information with invented information inside the same paragraph, and there is no clear way to tell which is which.

On top of that, more and more of the internet is itself AI-generated, which means models are now learning from the outputs of other models. The original human-written sources are getting buried under everything that has been auto-generated on top of them.

This is why I think reliable knowledge still has to come from people. If we can’t fully trust the model, and we can’t fully trust the sources behind the model, then human experience is the one thing left to rely on.

As Tony Robbins once said: “If you want to be successful, find someone who has achieved the results you want and copy what they do, and you’ll achieve the same results.” A person who has actually done the thing they’re teaching went through years of work, paid for their mistakes, and built their knowledge with real consequences attached — that is the kind of source that holds up.

Without context, the details are missing

Another habit that helps when learning with AI is to resist the temptation to jump straight to your specific question. Most people, when they want to know something, type the exact narrow question into a chatbot and take the answer at face value. The problem is that they miss the entire background in which the answer sits. A specific piece of information detached from its context is very easy to misinterpret.

I’m a history geek, so let me give a history example. The Gettysburg Address is a short speech Lincoln gave in 1863, on the site of one of the bloodiest battles of the American Civil War. Imagine you need to write an essay on it. You ask AI for an analysis and get a clean breakdown of the words, the structure, and the rhetorical devices.

But can you really write that essay without knowing anything about the Civil War itself? Probably not. You wouldn’t understand why this particular battle mattered, or why a speech given on that ground carried so much weight. You wouldn’t catch that the war had shifted from preserving the Union to ending slavery, and that the speech was Lincoln’s way of redefining what the country was fighting for. You wouldn’t see the political tension behind almost every line, because to you the lines would just be nice phrases about a sad event.

Without the context of the Civil War, the speech reads like a short eulogy. With it, you’re reading a turning point in how a nation understood itself. And that’s the kind of nuance AI will hand to you only if you already know enough to ask for it.

The same principle applies to almost anything you might use AI for. Start broad. Get the general picture. Then dive into the specifics. They will mean something once they have a context to land in.

Bottom line: AI is a tool, and tools need hands

Every major technological leap in human history triggered the same panic. Calculators were going to make us forget math. The internet was going to make us forget how to think. Search engines were going to kill our memory. And yet here we are, still calculating, still thinking, still remembering, just with better tools at hand.

AI is the next one in that line, and probably the most powerful one we’ve ever built. It can write, summarize, analyse, draft, translate, code, and do all of it faster than any of us. But it doesn’t want anything. It cannot care about the outcome. It cannot live with the consequences of being wrong. Those parts of the work still belong to us, and they are the parts that define whether the work is any good at the end of the day.

The people who thrive in the next decade will be the ones who stay curious, the ones who keep asking better questions, and the ones who use AI to learn faster and think deeper, instead of using it to skip the thinking altogether. They will be the ones bringing something to the table that no model can replicate: judgment, context, and the willingness to own a decision.

So no, AI will not replace you or make you lose your job. But the version of you that stopped thinking might.