Fifty-five percent. That is the share of workers in a 2026 survey of more than 3,000 people who believe an AI tool was used to evaluate them during a recent job application or promotion. Most of them were never told, and most of them never asked.
What makes that figure striking is not the technology. Screening software has existed for years. It is the belief itself. More than half of us now suspect that the first reader of our work, our application, our report or our case for a raise may be a machine. Nobody has to confirm it for the suspicion to change how we behave.
That is a human issue before it is a technical one. When we think a machine is the audience, we start writing for the machine. We flatten our voice, hide our doubts and hedge our opinions. The warmth and judgement that make our work worth reading are the first things we sand off. Staying human here means deciding, on purpose, who the real reader is.
------------- Context -------------
AI is now woven into the quiet parts of working life. It sorts applications, summarises performance notes, drafts first-round feedback and flags which messages look urgent. Much of this is sensible. Nobody wants to read 400 applications by hand, and a good summary can help a busy manager prepare for a conversation.
The trouble starts when the use is invisible. If we do not know whether a person or a tool will read what we write, we guess. And when we guess that a tool is reading, we tend to optimise for it. People stuff keywords into a CV. They write status updates in a safe, bland register. They avoid the honest sentence because it might be misread by something that cannot read between the lines.
There is a second effect, on the other side of the desk. A manager who relies on a summary of someone's work can begin to feel that they know the person, when they have really met a compression of them. The summary is tidy, consistent and confident. It is also missing the detail that made the work human: the difficult month, the quiet help given to a colleague, the idea that was half-formed but promising.
The better framing is to treat AI as an assistant to a reader, never as the reader. A person stays accountable for what is concluded about another person. The tool can sort, summarise and suggest. A named human decides, and says so.
This matters because trust, candour and fairness are the things teams run on. Once they thin out, no amount of speed makes up for it.
------------- When We Write for the Machine, We Stop Writing for People -------------
One of the quietest costs of invisible AI reading is the way it changes the writing itself. People adjust to their audience without noticing. If the audience is believed to be a scoring system, the writing becomes more generic.
Consider a project lead preparing her promotion case. She has done difficult, unglamorous work: she rescued a client relationship, trained two new starters and rebuilt a failing process. Her first draft says exactly that, in her own words, with a few honest lines about what went wrong. Then she hears that a tool might be scoring the document. She trims the honesty, adds the phrases she thinks it wants, and the result reads like everyone else's.
What is lost is not just style. It is information. The honest line about what went wrong was precisely what a thoughtful reader would have valued, because it showed judgement. A system that rewards polish over candour teaches people to stop offering it.
The human outcome here is plain. Candour is a form of trust, and trust is hard to rebuild once it is taught out of people. If a team learns that honest writing gets scored worse than confident writing, it will produce confident writing and keep the honest version to itself.
The fix is not to avoid tools. It is to tell people who will read their work and how. When someone knows a person is the reader, they write like a person. That is a small piece of transparency with a large effect on the quality of what comes back.
------------- A Summary Is Not the Same as Knowing Someone -------------
Summaries are useful. They are also a loss, and we should be honest about what is lost. A summary keeps what is repeated and drops what is rare. It favours the pattern over the exception. In people, the exception is often where the interesting part lives.
Think of a team leader with eight direct reports who starts a one-to-one by reading an AI-generated overview of the last quarter. It lists completed tasks, response times and themes from feedback. It is accurate. It is also a little thin, and it makes the conversation easy to run on autopilot. Ten minutes in, she realises she has not asked how the person is actually doing, because the page in front of her already felt like an answer.
The relationship outcome is what matters. A one-to-one is not a data transfer. It is a place where someone feels seen and where a manager hears the thing that never fits on a dashboard. If the summary replaces the listening, the meeting looks efficient and quietly fails at its main purpose.
The human move is to let the summary open the conversation rather than close it. Read it beforehand, then put it away. Ask the question the summary could not have raised. The tool saved the preparation time. The person keeps the part that makes the meeting worth having.
That is how a team keeps its warmth while using the tools. The machine handles the sorting, and the people handle the meaning.
------------- Someone Has to Own the Conclusion -------------
When a decision about a person comes out of a process that includes AI, a simple question matters: who is responsible for this conclusion? If the answer is "the system", something has gone wrong.
Accountability is a human thing. A hiring decision, a performance rating or a promotion outcome affects someone's income, confidence and future. We would not accept "the spreadsheet decided" as an answer to that person, and we should not accept "the tool decided" either. A named person needs to be able to explain the outcome in their own words, and to change it if it turns out to be wrong.
Imagine a small agency owner who uses AI to rank freelance applicants for a project. The tool returns a tidy list of five. She notices that a candidate with an unusual background, someone who changed careers and has a patchy portfolio, sits at the bottom. She reads that application herself and finds the most thoughtful cover note of the lot. She moves the person up and hires them. The tool was useful, and her judgement made the difference.
The point is not that the tool was wrong. It is that someone was positioned to disagree with it. That ability to override, and the willingness to use it, is the human layer that keeps the process fair.
When ownership is clear, people trust the outcome more, even when it is not the one they hoped for. They can see that a person looked, thought and stood behind the decision. That sense of being properly considered is worth protecting.
------------- Practical Moves -------------
First, tell people who will read what they submit, and whether any tool touches it first. A single plain sentence on the form or in the email removes the guessing and brings back honest writing.
Second, when a summary is used before a conversation about a person, read it, then set it aside and begin with a question the summary could not have asked, such as how the work has actually felt lately.
Third, name one person as the owner of every decision about someone's hiring, rating or promotion, and make sure they can explain it in their own words without referring to the tool.
Fourth, once a month, pick one outcome the tool ranked or flagged and read the original material yourself. Look for the exception that the summary dropped, and note what you find.
Fifth, share one example of honest, imperfect writing that was rewarded, so that people see candour is safe. What leaders praise teaches the team more than any policy does.
------------- Reflection -------------
The figure at the start, more than half of workers believing a machine may be evaluating them, is really a figure about trust. It tells us that people are already behaving as though their audience has changed, even when nobody has told them it has.
That is why staying human matters here. AI can sort, summarise and speed up the preparation. What it cannot do is look at a person and decide that they deserve the benefit of the doubt, notice the effort behind an imperfect page, or take responsibility for a conclusion. Those parts belong to us, and they are what people are really asking for when they ask to be treated fairly.
In the end, the tools will keep getting better at reading. The question for our teams is whether we keep getting better at listening. A reader we never meet may be part of the process, but a person we can talk to should always be at the end of it.
When did you last change how you wrote something because you suspected a tool, not a person, would read it?
Where in your work does a summary risk standing in for really knowing someone?
And who, in your team, is clearly the named human behind the decisions that affect other people?