Many months ago, during a job interview, I was for the first time unable to tell a candidate’s words apart from the raw answer artificial intelligence would have given me. Since then such situations have kept coming back, both when I read submissions for our recruitment tasks and when I listen to people in interviews, and every time it happens I ask myself the same question: what exactly would companies be paying a salary for, if access to the very models used by the people on the other side comes in a company subscription for a few hundred zlotys a month. The question itself is probably too harsh, because behind those answers are real people who simply wanted to show their best side. That is why, much more than looking for someone to blame, I am interested in what changed along the way. I have the impression that our thought processes are turning into prompt processes right before our eyes, and recruitment is merely the place where it currently shows most clearly.
It started with a strange feeling of déjà vu in one of the recruitment processes I was running at the time. I would open yet another submission for our recruitment task and somewhere deep down I already sensed what I was about to read.
The task itself looks unremarkable. It is a project specification that imitates an initial enquiry from one of our clients, with deliberate gaps, because real enquiries have them too. We ask the candidate to dive into it and show how they understand the problem of the company described in it, before they start solving anything. We put a lot of weight on this stage and we are fairly strict about it. Over time I started seeing chat-written answers at the other stages of the process as well: in the questionnaires candidates fill in when applying, in the tasks between stages and finally in the interviews themselves.
And for quite a while now, the answers coming in have been strikingly similar to one another. The same steps repeat in a similar order, along with a very similar way of thinking about the problem. In tasks involving estimation, the same 30, 60 and 90 day plan kept coming back, the one the model produces on its own when you ask it to lay out such work. When we paste our task into a chat, with a simple or even a somewhat more elaborate prompt, we get exactly what candidates send or say. Artificial intelligence varies the language a lot, so at first glance each of those texts looks different. Only when I lay them side by side do I see that I keep reading the same answer, dressed in different words each time.
We also ran this task internally, because it maps well onto our daily work with clients. And none of our people would approach the problem this way. The answer the chat suggests sounds reasonable and is hard to fault at first glance, but experience in this industry teaches you that in practice this approach often fails. The candidates apparently never asked themselves whether the proposed steps stood any chance of working for the client described in the specification.
This is not just my problem or a matter of one task. People running further recruitment processes at our company now bring me such cases themselves. I hear the same from other leaders I talk to, from technical directors in software houses, through HR directors in product companies, to recruiters who are hiring other recruiters. The job market is also starting to react in a way that two years ago would have looked like companies deciding to move backwards in how they source talent. Gartner surveys suggest that around 72% of recruitment leaders now run interviews on site, in the office, precisely to rule out AI prompting during remote meetings. Google, Cisco and McKinsey have brought face-to-face interviews back into parts of their processes. I should note these are industry surveys and press reports, not peer-reviewed research, but the direction matches what I see around me.
At first I treated the whole thing as an honesty problem. Someone took a shortcut, it happens. Today I think honesty is the least interesting thread here, because what I get from candidates has already been measured and named by researchers: homogenization.
Twenty messages a day
Recruitment is, after all, only one of the places where I see this. Because of the role I play in the company, I sometimes get as many as twenty cold messages a day from people who want to sell me something. Two trends overlap in them lately, hyperautomation and hyperpersonalization, and it leads to absurdities. In a LinkedIn invitation, within the limit of roughly three hundred characters, I can receive information about myself so condensed that the two sentences do not even connect. And in longer sales messages I keep getting back templates I myself used to send for laughs fifteen years ago, when I was taking my first steps in sales. There was a trick back then: to a client who stubbornly refused to pick up the phone, you would send a playlist with “Call Me Maybe”, “Not Right Now”, “We’re Not Gonna Take It” and “Leave Me Alone”. It was part joke, part desperation. Today those old moves are getting a second life from artificial intelligence and coming back to my inbox as follow-ups to messages automated with the wrong prompts.
First my whole LinkedIn feed insisted that sales had to be prompted and automated at scale. So some people went for the simplest automated solutions without thinking about how to do it sensibly. Now LinkedIn declares that automation has gone too far and it is time to return to the human. As a rule I agree, though artificial intelligence can genuinely help in sales. You would just need to use thought processes with it, not prompts alone, and I know how sharp that sounds as I say it.
Anil Doshi and Oliver Hauser measured this most precisely. Their experiment concerned short stories, but the conclusions carry over to any text written with a chat, from sales offers to recruitment answers.
In an experiment published in “Science Advances” they invited three hundred authors to write a short, eight-sentence story for young adults. Before anyone sat down to write, each person took a divergent association test, a simple measure of how creative someone is without any aids. Then the researchers split people into three groups. The first wrote unaided. The second received one three-sentence plot idea from GPT-4. The third could ask for up to five such ideas. The finished stories went to six hundred raters who did not know how they had been written. The effect was easy to predict.
Texts written with the help of artificial intelligence received higher ratings. Those who scored lowest on the creativity test gained the most, as their stories were rated as much as a quarter better written and clearly more enjoyable to read. Only when the researchers compared all the texts did they see the other side of the coin: stories written with a single idea from GPT-4 were 10.7 percent more similar to one another than those written unaided. Each author individually wrote a better text, but the group as a whole produced less diverse stories.
The authors called it a social dilemma. The individual benefits, while the collective loses part of its diversity of ideas. Another result turned out to be even more interesting. A single suggestion homogenized the texts more strongly than five different proposals. Apparently the first suggestion does not merely give us a starting point. It sets the direction in which all of our thinking then runs.
It is tempting to think that a salesperson or a candidate who wants to stand out from the rest can simply move to another tool, switch the model or the provider.
Emily Wenger and Yoed Kenett decided to check. They took a broad set of models from different makers and gave them the same standardized creativity tests that psychologists have used for years to measure human inventiveness. Then they measured how similar the models’ answers were to one another and set that against how much people’s answers differ from other people’s.
The models answered far more similarly to each other than humans answer relative to other humans, even after filtering out differences in structure or answer length.
The literature already has a term for this: algorithmic monoculture. The models are developed by different teams, but they learn from largely shared pools of data. As a result they suggest similar solutions, overlook similar possibilities and repeat similar mistakes. From the user’s perspective the chat window changes, but not necessarily the direction of thinking.
When I lay all this over our recruitment task, nothing surprises me anymore. The candidates did not conspire with each other. Each of them sat down alone with their chat, received a sensible-sounding plan and adopted it as their own. In parallel, an almost identical proposal went to everyone who asked a similar question. And what puzzles me most in those interviews is that the people bringing twin answers sound more and more confident while doing it.
Experts on everything
In the interviews themselves, candidates answer my questions directly less often than they used to, and more often try to guess what I want to hear. I have the impression they are manoeuvring. Mid-sentence they scan my reactions and the reactions of the people accompanying me, and calibrate the answer to our expectations as they go. As if they were looking for a prompter in us first, before finding one in the tools. What they themselves think fades into the background, if it appears at all.
I have my own way of dealing with it, though it is not particularly sophisticated. I repeat some fashionable claim in the interview and check whether the person simply agrees with it or engages in a deeper argument. It is my main filter today, because what I care about most is seeing whether someone thinks for themselves. And I regularly sit across from people who give the impression of being experts on practically every topic. The confidence is high and the knowledge broader than ever, except that after two or three probing questions it gets very shallow.
An experiment from Aalto University describes this best. Daniel Fernandes, Steeven Villa and Robin Welsch checked what happens to our self-assessment when we solve tasks with the help of a chat. They first invited just under 250 and then over 450 more people, and gave them logical reasoning tasks from the LSAT, the American law school admission exam. Half of the participants could use ChatGPT. After the tasks everyone assessed how they had done, and accurate self-assessment earned extra money, so nobody had a reason to fool themselves.
The classic Dunning-Kruger effect works like this: the weakest overestimate their abilities the most, while the best tend to underestimate them. In the group working with the chat this pattern fell apart. Everyone overestimated their results, regardless of actual ability.
And in some of the measurements the effect outright reversed, because the people most convinced of their own high proficiency with artificial intelligence overestimated themselves the most.
The researchers also looked at how participants actually worked with the tool. Most asked exactly one question per task. They copied the text, pasted it into the chat, took the result and moved on. No second question and no check whether what they received even held together logically.
The confidence I see in candidates is probably neither an act nor nerve. These people genuinely feel like experts, because the answer they bring sounds good. Except that this conviction came from the quality of what the model showed them, not from their own path to the solution. In such conditions it is easy to mistake access to knowledge for possessing it.
Between reading a problem and solving it, some work always used to happen. A person had to understand what actually needed to be done, break the problem into smaller parts, form their own hypothesis and check whether it held up at all. For the participants of the study this middle part dropped out. What remained was the beginning, reading the text, and the end, accepting a ready answer. And if it dropped out even under the conditions of a controlled experiment, I find it hard to believe it looks any better in everyday work, where nobody is watching our hands. That cut-out middle part is precisely the thought process. And it is the part we increasingly replace with a prompt process.
The prompt process
When I talk to people about their real challenges, also far from recruitment, the conversation moves surprisingly quickly away from the problem itself to what to type into the chat and how to phrase the prompt to get a good answer out of it. The question “how do we solve this” somewhere along the way turns into the question “what do we ask”.
All the work then begins and ends at the chat. We compose the best possible instruction, and later we judge whether what came back looks good enough to hand over or show to someone. What is left for us is the role of a supervisor of someone else’s thinking.
A study by Microsoft and Carnegie Mellon says a lot about what this new role looks like in everyday work. Hao-Ping Lee and his team asked 319 knowledge workers who use artificial intelligence at least once a week to describe a total of 936 real situations from their work. These were specific emails, reports and analyses these people had actually worked on, and for each such situation they reported how much of their own critical thinking they had put into it and how much they had trusted what the tool returned.
The more someone trusted the tool, the less critical thinking they put into the task. The more they trusted their own skills, the more often they checked, corrected and questioned, even though it cost them more effort.
The more important thing, however, was happening to the thinking itself, because in these people it did not disappear, it changed character. It shifted from gathering information to verifying what the machine returned, and from solving problems to stitching together ready answers, until in the end the work itself turned into supervising it.
The candidates from our recruitment processes took the model’s answer, judged that it looked good and passed it on. The catch is that to sensibly assess someone else’s thinking, you have to be able to walk that road yourself.
It is easy to demand it from a candidate while holding a list of questions, and then send a message we did not think about for a single moment. It is hard to honestly assess someone else’s path when we hand our own over to the model with every longer message.
Today, when we have a longer email to write, we very often start by opening a chat window and throwing in a few sentences about what we want to convey. Sometimes it is a short instruction the model is supposed to infer the rest from. Back comes a text longer and more polite than what we had in mind, so it goes out almost unchanged.
On the other side, the recipient gets several paragraphs the sender did not write, and still has the rest of the day’s correspondence in front of them. So they copy what they received into their own chat, ask for a summary and, while at it, for a draft reply. The reply comes back to us just as elaborate, so we do exactly the same with it. In this exchange the sentences are arranged by two models, while we, on both ends, handle the copying and pasting.
The joke about wiring two chats together and letting them work things out without us has been circulating for years. It is getting harder to laugh at, because in such an exchange we are sometimes the sender and sometimes the one pasting someone else’s text in for a summary. When confirming a date or settling meeting details, nothing bad happens. It is worse when a problem description, a project scope or feedback for someone on the team passes through us this way.
It does not matter how the text was produced, since it was I who handed the model my thinking and it merely dressed it in words. That is how we explain it to ourselves, and it is hard to disagree, as long as we have something to hand over. It is just as easy to type a bare instruction into the chat window and wait for the model to invent the content for us. From the outside both messages look the same, so people increasingly stop reading the longer ones and assume upfront that a chat wrote the other side’s message anyway.
The tool has also raised the bar. Since everyone writes longer and faster, more messages arrive, and from the outside the only thing visible is how quickly someone replied. An elaborate message takes a minute to produce, while understanding what the other side really means takes a dozen or more. We choose the minute, because only the minute fits into the day we have. Along with the writing, we then hand the model the decision about what we actually want to say to the other person, and after the whole exchange we know roughly as much about the matter as before it, except now it is ticked off.
In news portal headlines all this handing of thinking over to the machine runs under the banner of “is AI making us dumber”, and the more serious version of the question is how AI affects critical thinking. The data is less clear-cut than the headlines.
Michael Gerlich collected data from over 600 people and the more often someone used artificial intelligence, the worse they performed on critical thinking tests, with the relationship strongest among the youngest. The mediating factor was a phenomenon called cognitive offloading: the more willingly we transfer mental effort onto a tool, the less often we engage it ourselves.
This is a correlation though, not proof; people with weaker critical thinking may simply reach for the chat more often. Meanwhile a team from MIT Media Lab, whose experiment I described at length in “I miss thinking”, coined the term cognitive debt for this state, because postponed thinking works like postponed debt and the costs have to be repaid someday. In their study, people writing essays with a chat had the greatest trouble quoting a sentence from the text they had handed in moments earlier, and when the tool was taken away in the next round, their brains did not return to the previous level of engagement right away. I will just add that this was an experiment on 54 people, still without peer review, so I treat it as a hypothesis.
This swap does not end at recruitment either. In sales, which I wrote about earlier, hardly anyone wonders why clients stopped replying; instead more automation gets bolted on. In the emails, reports and analyses from the Microsoft study you can see it in the most ordinary office work.
In all these places the missing middle part of the process can stay invisible for a long time, because the finished output looks equally neat whether or not someone thought it through. On top of that, thinking used to produce along the way not only a solution but also doubts and questions with which a person checked themselves. The chat rarely raises doubts, because it is built to agree with us, so the natural moment of pause disappears too. The client simply does not reply, and the chat-prepared report goes to a manager (who sometimes also uses AI to analyse it), and since it looks like any other, nothing in this loop signals that thinking was missing from it.
Recruitment is merciless in this respect. We sit at one table and at any moment I can ask a question that was not in any prompt. Unless, of course, the candidate in an online meeting is using tools that analyse the conversation in real time and feed them suggested answers.
Apps like Cluely or Final Round AI display suggestions on a semi-transparent overlay, invisible when sharing the screen, and a company that analyses job interviews calculated on a sample of nineteen thousand interviews that the share of candidates using such prompters doubled in the second half of 2025, from 15 to 35 percent. Naturally, this data comes from a vendor that makes its living detecting them, so I take it with a grain of salt.
The prompter
I came across such a prompter personally last year. We were running a remote job interview with a candidate applying for one of the positions at our company, and from the beginning I had the impression he was reading his answers. He looked straight at the screen, almost directly into the camera, and at moments he sounded as if he were simply reciting a prepared text. At some point I had no doubts left and I asked him directly whether he was using artificial intelligence suggestions at that very moment.
The candidate was flustered for a moment. He looked a bit like a gamer caught playing with cheat codes. In the end he denied it, but from that moment on our conversation was completely different and on a different level. I was finally listening to what he himself thought, not to what a tool had generated in real time. These are exactly the interviews I expect.
I told him that at the end, by the way, because the second part of our meeting, the one where he spoke for himself, had a different tone than the first. I can also understand why he reached for the prompter at all. I tell myself he simply wanted to come across better and probably somewhere along the way came to believe his own answers would not be enough. The IT job market has changed a lot in recent years; employers find candidates much more easily now and competition for every position has grown considerably. It ended up backwards though, because the suggestions that were supposed to help him get the job spent half the interview covering up the very things he could have been hired for.
Another example came to me from team members who were running one of our recent recruitment processes. They came across a candidate using artificial intelligence for practically the entire interview, including the questions he asked them at the end. From the start the answers were very fluent and perfectly arranged, without a single stumble, like ready-made scripts fitted to every question.
When they began to suspect something was off, they switched to questions requiring him to refer to earlier answers or to his own experience, and then the way he answered became, oddly enough, even more characteristic of a model. The strongest moment came when they asked the candidate to act out a short code review scene, in which he was supposed to react to a junior developer’s mistake. Several times they clarified the scenario and steered him towards the task, but each time he came back to an idealized, general description of what giving feedback should look like, and not once did he step into the role. It looked as if on the other side there was a tool loaded with a CV and project descriptions, generating ready answers to read out in real time.
In a job interview a prompter hides a person for an hour at most, but at work it can hide them for years, and that is beginning to cost real money, and not only the employee’s.
A salary versus a subscription
The salary of even a junior specialist is today the equivalent of several dozen company subscriptions to artificial intelligence. So if someone’s work begins and ends with rewriting the problem into a chat window, their cheapest competitor on the job market becomes precisely that subscription.
Erik Brynjolfsson and two collaborators showed in numbers what suggestion tools do to the value of someone’s work. The researchers got access to data from a large technology company that rolled out an AI assistant to its customer service consultants, in essence an open and legal version of the prompter from the previous section. The tool followed the conversation with the client in real time and suggested to the consultant what they might reply; the consultant could use the suggestion, rework it or ignore it. The analysis covered more than five thousand consultants and around three million conversations, together with information on how quickly they resolved cases before and after the rollout.
Average productivity rose by 14%, which for such rollouts is not particularly impressive. Only breaking the result down by group showed what had really happened. The least experienced consultants improved by as much as 34%, while the best ones barely at all. The assistant, it turns out, learned from the conversations of the best and handed their way of working to everyone else. Beginners received in the suggestions the experience they had not yet had time to gather, while the best merely got back what they had worked out earlier.
For beginners and for companies this is good news, but it drags a less pleasant consequence behind it. If a solid average is available on the spot to anyone who pays for the tool, the market stops paying extra for it, and the skill premium, that is, how much an employer is willing to add for being better than average, moves to where the tool does not reach. You can already see it in employment data.
The same Brynjolfsson, this time on payroll data covering millions of American workers, checked what is happening to the youngest people on the market.
Employment of people between 22 and 25 in the occupations most exposed to artificial intelligence fell by a relative 13%, while their older colleagues in the same occupations are holding steady.
The decline concentrates where AI automates the work and passes over the places where it merely assists. This is still fresh data, hard to separate from the general economy, so I treat it cautiously, as one more clue.
So where is the edge supposed to come from? A partial answer comes from an experiment by Hamsa Bastani’s team at Wharton, run on nearly a thousand high school students learning mathematics.
Some of them got a regular chat for their practice sessions, some got a version reworked into a tutor that guided them with questions and hints but did not hand out results, and the control group learned the old way. During practice the results with the tool looked great: 48% better with the regular chat, and more than twice as good with the tutor, compared to no help. The resolution came at the exam, where the chat was no longer available.
The regular chat group scored 17% worse than students who had never seen artificial intelligence, while the tutor group lost nothing. What decided the outcome was whether the student kept doing their own thinking along the way or handed all of it over to the tool.
For company leaders the stakes here are double. First, they pay someone a salary many times higher than a subscription, so they want the person’s own work, or at least honest verification, to happen between the problem and the answer. Second, companies live mostly off thinking about their clients’ problems differently than others do. If offers and project approaches start coming out of the same models the competition uses, the client will sooner or later stop seeing the difference. A buyer who asks five companies for a quote and receives five documents with the same steps and very similar promises is in exactly my situation with the recruitment task. And when offers stop differing, price becomes the only thing deciding the choice.
It is easy for me to write about this from the position of an employer who judges other people’s answers. At some point, though, I had to ask myself the same question.
I know what I don’t know
Just before publishing this text I came across an experiment by Valerio Capraro’s team, devoted precisely to the sentence “I don’t know”. In five studies on more than three thousand people, participants answered difficult questions about small details from films, for example the colour of the team kits in “Bend It Like Beckham”, and at every question they could also decline to answer. Some of the participants received a suggestion from a model that the researchers had deliberately chosen to be usually wrong on these questions. The point was that nobody could justify their decision as reasonable trust in a competent tool.
Without the suggestion, 44% of participants were able to admit they did not know the answer. With the suggestion, only 3% still said so. People answered more questions, were right three times less often, and their confidence nearly doubled.
The researchers then tried paying for accuracy and penalizing mistakes, which helped a little, but the willingness to admit ignorance never returned even close to the level from before the suggestions. It is still a preprint, so I approach it cautiously, but it is hard to find a better illustration of how quickly the sentence “I don’t know” is disappearing from us.
I used to catch myself on this too: when something came up that I did not know, my first thought was “I don’t know this, I’ll ask AI”. Somewhere along the way, “so I’ll learn it” dropped out of that sequence. The answer from artificial intelligence arrived instantly, so the topic felt settled, and I moved on to the next thing with the pleasant feeling that I already knew everything. After a while I noticed that this feeling had little to do with what I could actually do.
Today I prefer an approach that lets me make better decisions. Instead of feeling that I know everything, just flatly, one prompt deep, I prefer to know what I do not know and to know how to learn it. In practice the difference comes down to how a conversation with a chat ends. If the answer closes the topic, I am left with someone else’s conclusion and an illusion of knowledge. If it opens it, that is, it points me to sources, explains the mechanics instead of just the result and leaves me a piece of work to practise on my own, then it is one of the better learning methods I know.
I write all this as someone who uses artificial intelligence daily, deploys it for clients and wrote a book about how to do it sensibly. I am not urging anyone to put the tools away. What I care about is that artificial intelligence helps me do things better, teaches me something or automates repetitive work, and that I do not pretend to understand something I have no understanding of, because that is simply dangerous. I try to stick to the rule that repetitive tasks can be taken over by the tool, but understanding the problem and deciding on the direction stay with me.
The same filter I use in job interviews works in everyday teamwork too. When someone shows me the result of their work, I can ask why they chose this solution and what alternatives they rejected along the way. A person who went through their own thought process will talk about it without much hesitation. One who only pasted in the chat’s answer usually stops where their prompt ended.
The tutor from the Wharton experiment behaved like a good teacher: it guided instead of handing out results. Nothing stops us from setting up our own chat the same way; it is enough to start the conversation with a simple instruction like “do not give me the ready solution, guide me with questions and check my answers”.
What gives me the most hope in this whole topic is a study by Eric Hanushek and his team. In Germany, more than three thousand people between sixteen and sixty-five solved reading comprehension and numeracy tests as part of a large international survey of adult skills. After roughly three and a half years the researchers returned to the same people and asked them to solve the tests again. Thanks to this, instead of comparing the young with the old, they could see how a specific person’s skills change as the years pass.
The average confirmed what you would expect: results rose until about forty and then began to decline. But when the researchers split participants by how often they use reading and numeracy at work and at home, it turned out the average was gluing together two completely different groups. Among people who do it rarely, the decline started already around age thirty-five. Among those who do it often, there was no decline at all, at any age studied. Thinking behaves like a muscle in this respect, and it works both ways, throughout life.
I use these tools every day myself and I see how much they can straighten out in a company and how much repetitive work they take off people’s shoulders. All I care about is that in this whole process we ourselves remain, and that regardless of the technology we surround ourselves with, the decisions still belong to us, the good ones, and with time better and better ones.
Today I leave you with two questions I regularly ask myself as well. How many of the answers you received from others this week were still their answers? And when did you last crack a problem yourself, from start to finish, before asking a chat about it?
