Skip to content
Go back

AI at Work: People Who Use It Are Judged Lazier (There's a Study)

Illustration with the title: AI doesn't replace you, it empowers you

In 2025 four researchers published in PNAS a series of experiments on over 4,400 people asking a very specific question: what do others think of you when you disclose that you used AI to do your work?

The result is consistent across all four studies. People who disclose using AI tools are rated as lazier, less competent, less diligent. And in the hiring experiment, evaluators who didn’t use AI themselves were less likely to hire candidates who disclosed using it. The sign flipped only when the evaluator was a regular user too.

So the question “how do I make the most of AI at work” has a component almost no guide addresses: it isn’t only about what you get from the tool, but what happens when people find out you used it.

The practical consequence: AI gets used in secret

This social penalty has a measurable consequence. In Microsoft and LinkedIn’s 2024 Work Trend Index, a corporate survey that should be read as such rather than as an experiment, 78% of people using AI at work said they brought their own tools rather than company-provided ones. And just over half admitted being reluctant to say they had used AI for their most important tasks.

Together those two figures describe a pattern anyone working in an average company will recognise: the tool is everywhere and officially doesn’t exist.

The consequences aren’t only reputational. If usage is underground, there are no guidelines on what can be pasted into a chat, nobody checks critical outputs, and skills don’t spread between colleagues because nobody admits having anything to teach.

The pendulum: first we trust too little, then too much

There is a literature predating generative AI that describes what’s happening rather well, and it’s worth knowing because it avoids two opposite mistakes.

On one side, Berkeley Dietvorst, Joseph Simmons and Cade Massey documented in 2015 what they called algorithm aversion: after seeing a statistical model make an error, people stop using it and revert to their own judgement, even when the model remains more accurate than they are. A human error is forgiven; a machine error isn’t.

On the other, Jennifer Logg, Julia Minson and Don Moore found in 2019 the opposite phenomenon, which they called algorithm appreciation: given identical advice, people follow it more if they believe it came from an algorithm rather than a person.

The two results don’t contradict each other; they depend on how much you know about the task. Together they explain the pendulum you see in offices today: categorical refusal until it works, total delegation the moment it seems to.

The second extreme has a technical name and a substantial literature. Raja Parasuraman and Dietrich Manzey, in a 2010 review in Human Factors, call automation complacency the reduction in monitoring that occurs once an automated system has proved reliable. Linda Skitka and colleagues had already shown in 1999 that automation produces two distinct error types: omission errors (you miss what the system didn’t flag) and commission errors (you follow a wrong flag against the available evidence).

Both become more likely the more reliable the system is. Not less.

The ironies of automation

The most useful piece of this literature is forty years old. In 1983 Lisanne Bainbridge published in Automatica a short paper titled Ironies of Automation describing a mechanism today’s rhetoric ignores entirely.

The argument runs like this. When you automate a job you automate the easy, regular parts, because those are the automatable ones. What’s left for the human operator is only the exceptions, that is, the hardest cases. But the competence needed to handle hard cases was built precisely by doing the easy ones, repeatedly. Remove the ordinary practice and you erode the capacity to face the extraordinary, and you erode it in exactly the people asked to step in when the system breaks.

This is the most serious objection to the most repeated sentence of recent years, the one where AI frees you from low-value tasks to leave you the high-value ones. To the extent that the high-value task was learned by doing the low-value one, the promise contains a problem.

Automating the first draft is convenient. The cost is that, after a while, you no longer recognise a bad first draft.

How to use it without losing your craft

  1. Decide in advance what you will never delegate. Write down two or three activities you’ll keep doing by hand because your professional judgement rests on them. It’s the direct countermeasure to Bainbridge’s irony.
  2. Separate the first draft from the result. AI is useful for getting you from a blank page to something you can criticise. The value you add is the criticism, not the draft.
  3. Verify every verifiable fact. Names, figures, citations, legal references. That’s the category where the model is most convincing and least reliable; why is explained in generative AI.
  4. Don’t paste confidential data into unapproved tools. That’s the concrete risk of underground use. If your company has no clear policy, asking for one is a professional act, not an admission of guilt.
  5. Be transparent about the process, not the tool. In light of the PNAS study, saying “I used ChatGPT” and saying “I did a first pass, then verified the sources and rewrote it” communicate very different things, and the second is also more informative.
  6. Measure your actual times. Perceived time saved is unreliable. Two weeks of notes are worth more than any impression.
  7. Invest in the part that isn’t levelling out. Understanding context, saying no, deciding under ambiguity. More on that in critical thinking.

My take: the penalty will pass, the risk of deskilling won’t

I think the social penalty documented by the PNAS study is a phase, and that it will fade once usage becomes the majority among evaluators. The finding that evaluators who were users didn’t penalise already points that way. But in the meantime it’s real, and it’s worth keeping in mind when writing an application or delivering work to a client.

I also find the most common advice, “use AI for the boring tasks”, more ambiguous than it sounds. Some boring tasks really are just boring. Others are boring and the only place experience accumulates. Telling the two apart is work nobody can do for you, because it depends on your trade.

Here I’m going beyond the data, so take it as a hunch: in ten years the difference between professionals won’t be how much AI they use, but how well they function when the tool gives a wrong answer with perfect confidence.

Frequently asked questions

Should I tell my manager I use AI? It depends on company policy, which should be respected. On perception, the PNAS study suggests how you say it matters: describing the process and the verification is different from just naming the tool.

Does using AI make me less competent? The research doesn’t say that. Bainbridge’s argument says something subtler: if you stop practising the ordinary tasks, your ability to handle extraordinary ones atrophies. That’s a manageable risk, not a destiny.

Why do I sometimes trust the output too much? That’s automation complacency, documented long before language models. It grows with the system’s perceived reliability, which makes it most dangerous precisely when the tool is good.

Will AI replace my role? The question deserves more than a paragraph. I covered it in AI and work, with the Frey and OECD figures that measure different things.

Where do I actually start? With one recurring, well-defined, verifiable task, for two weeks, measuring the time. The operational guidance is in how to use ChatGPT at work and writing effective prompts.

Sources

This article was written following the process described in how we write. If you find a figure reported wrongly, write to me: I’ll correct it and note the correction at the bottom of the article.

Founder of UpSkillLab, with a Master's in Digital Marketing from the University of Florence. Every article starts from public sources cited in the text and is drafted with the help of AI tools, under my review and editorial responsibility: no purchased content, nothing published without human checking. Read the editorial method or report an error.


Share this post:

See all articles in AI at Work

Previous Post
Generative AI: Experts Get Slower, and Don't Notice
Next Post
No-Code: 88% of Spreadsheets Contain Errors, and That Matters