Will AI replace you?

Almost everyone is using AI now. Among university students the figure reached 92% this year, up from 66% the year before, and among professional developers it is 84%, with just over half of them reaching for it every single day. So the first thing to be clear about is that access to AI is no longer an advantage of any kind, because the person next to you in the interview pool has exactly the same access you do. The only variable left is whether you use it in a way that makes you sharper or a way that quietly makes you worse, and in my opinion that single difference is about to decide who gets hired and who spends a year wondering why they cannot.

So will it actually replace you? The honest answer is no, but the comforting version of that answer is wrong, and the difference is the whole point. AI will not replace engineers (at least not yet), but it will absolutely replace the specific people whose only real skill is doing what the model already does for free. If your value is producing the standard solution to a known problem, you are no longer competing with the other candidates, you are competing with a tool that does your one trick instantly and never asks for a salary, and that is a competition you lose by definition. What survives is the part the model cannot do for you: deciding what to build, reasoning through something genuinely unfamiliar, and noticing when the confident answer on the screen is quietly wrong. Training yourself to be a slower version of the thing that is already faster than you is the losing move, and yet it is exactly what most people are now doing.

Watch how people actually use these tools and they sort into three groups of very unequal size. The first and largest treats it as a crutch, so the moment a problem resists them they paste it in, take the answer, and move on feeling productive, having skipped the precise struggle the exercise existed to create. This is the group the data should genuinely worry, because the damage is now measured rather than theoretical: in a controlled study from Anthropic, developers who leaned on AI to learn a new library scored 17% lower on comprehension afterward, 50% against 67% for the people who worked through it themselves, and the gap was widest on debugging, which is the skill interviews lean on hardest. They are not standing still while they do this, they are actively eroding the ability they showed up to build, and because the code runs, nothing tells them it is happening.

The second group, much smaller, refuses to touch it at all, usually out of some mixture of principle and fear, and while I understand the instinct it is still a mistake, because used correctly this is the highest-leverage tool the field has produced in a generation, and turning it down wholesale means surrendering real speed to the people who learned to wield it. The third group, smaller again and the entire reason this matters, uses it correctly, which means they never hand it the part that builds the skill and they lean on it hard for everything that does not. That group is small enough to be rounding error, and that scarcity is the problem the rest of this is about.

Here is the finding that should make all of us humble about which group we are actually in. METR ran a proper randomized controlled trial in 2025 on experienced open-source developers working inside their own repositories, and the ones allowed to use AI took 19% longer to finish, while believing they had worked 20% faster. Sit with that gap, because it is the most important pair of numbers in this whole discussion: people who are genuinely strong, on their home ground, were slowed down by the tool and could not feel it. If correct use is rare even among experts on familiar territory, it is close to nonexistent among students on unfamiliar territory, which describes most people reading this. The point is not that AI is useless, because in the right hands on the right task it is enormous. The point is that the sensation of speed is not evidence of anything, and almost everyone is trusting that sensation completely.

The popular story is that AI lifts everyone equally, and it is simply false. People who were already strong are the only ones positioned to extract the leverage, because they can tell when the model is right, discard the eighty percent of its output that is mediocre, and spend the saved time on the parts that are genuinely mechanical, so for them it is a real multiplier sitting on top of a real foundation. People who were already struggling get the inverse, because they cannot evaluate what the model hands them, so they accept its mistakes, ship its confident nonsense, and skip the exact reps that would have closed the distance, which means the tool that was supposed to help them catch up instead pushes them further behind. The trust numbers show the wider context, since developer trust in AI accuracy has fallen to 29% from 40% a year earlier, so even the people using it most are increasingly aware it cannot be taken at face value, and evaluating it is itself a skill that only the already-skilled possess. The distance between the top and the bottom is widening quickly, and the uncomfortable part is that the tool everyone assumed would level the field is doing the opposite, paying out strictly in proportion to the skill you already brought to it.

Companies can see all of this, often more clearly than the candidates can, and it is the real reason the interview itself has moved. When every applicant can produce a clean solution to a known LeetCode problem, because a model produced it for them, that round stops telling the company anything at all, so they have shifted the test to where a model in the next tab cannot quietly answer on your behalf. Meta and Google have added AI-assisted rounds with deliberately harder problems, Stripe runs its loop on a real production codebase, and roughly 90% of startups now run at least one round built specifically to resist the tool. None of this is companies being cruel or nostalgic, it is companies doing exactly what you should be doing, which is refusing to measure the skill the machine already has and going looking for the one it does not.

Underneath every rule about correct use sits one discipline that matters more than the rest, which is that you have to know, precisely and honestly, what these models can and cannot do yet. Most people never build that map, so they end up either trusting the tool everywhere or distrusting it everywhere, and both are just different ways of letting it run you. The moment you understand its limits exactly, where it is reliable and where it confidently lies, you become the one controlling it instead of the other way around, and that understanding is never handed to you by using it more, it is earned the same way every real skill is, by doing the work yourself often enough to recognize when the machine is wrong.

From there the correct path is easy to state and hard to hold: never let the model do the part that builds the skill, and lean on it for everything that does not. When you practice, solve the problem yourself first, all the way to genuine depletion, and only then bring in the model, and when you do, use it to find the hole in your reasoning rather than to replace the reasoning, asking it where your approach breaks instead of asking it for a finished approach. That is the entire design idea behind the AI on WhiteBox, and it is why it behaves differently from a raw ChatGPT window: Tsuki is built to push you back toward your own thinking instead of dropping the solution in your lap, and White Room remembers what you are weak at and keeps steering you back into it rather than around it. Used that way, AI becomes the tutor that explains the idea you were missing after you have wrestled with it, not the answer key that lets you skip the wrestling, and the gap between those two uses is the gap between the top group and the bottom one.

Put the numbers back together and the picture is hard to look away from. 92% of students are using these tools, the single most common use is having concepts explained to them rather than working those concepts out, and the measured effect of learning this way is worse comprehension, so the median student is right now using the most powerful tool available in the precise manner that hollows out the skill the interview is shifting to reward. They are optimizing hard for the wrong target and feeling productive the entire time. The way out is not to quit the tool, and it is not to grind blindly without it, it is to be deliberate about the line between the two uses: do the reasoning yourself, every single time, until you have actually earned the answer, and then let AI sharpen and extend what you already understand. In my opinion the gap opening right now, between the few who hold that discipline and the many who do not, will be the defining divide of this hiring era, and which side of it you land on is still, for the moment, entirely your choice.