Critical Thinking vs. Automation: When AI Takes Away Your Judgment
AI takes away judgment when it replaces verification. What four studies show (Microsoft/CMU, PNAS, Anthropic, MIT), where they fall short and how to avoid it.
What does the evidence say?
None of the most cited studies measures design work. Here are four, with the size and type of each sample:
| Study | Sample and type | What it found | Main limit |
|---|---|---|---|
| Lee et al., Microsoft Research and Carnegie Mellon (CHI 2025) | Survey of 319 knowledge workers who use generative AI at work at least once a week; they shared 936 examples | More confidence in AI is associated with less critical thinking; more self-confidence, with more | Self-reported, and the sample skews young and tech-savvy |
| Bastani et al. (PNAS, 2025) | Field experiment with nearly a thousand high school students in math | With GPT-4, practice grades rose 48%; once removed, that group scored 17% lower than students who never had it | Math, not design |
| Shen and Tamkin, Anthropic (January 2026) | Randomized trial with 52 software engineers, mostly junior, learning a new library | With an AI assistant they scored 50% on the quiz; without it, 67% | Small sample, one task, comprehension measured right afterward |
| Kosmyna et al., MIT (arXiv preprint) | 54 participants writing essays; 18 in the final session; EEG measurements | The AI group showed lower brain connectivity and a weaker sense of owning their essay | Preprint, small sample, a single task |
On the first: six of its seven authors work at Microsoft, which sells Copilot, and the seventh is at Carnegie Mellon. The authors themselves admit participants sometimes confused “using AI took me less effort” with “I thought less critically”, and they suggest future studies measure critical thinking with observed tasks or think-aloud protocols, not only surveys.
On the MIT one: it is the most quoted in headlines and the weakest. A preprint with 54 people cannot support the claim that AI “damages the brain”. At most it shows that writing with a model felt less like your own work and activated fewer connections on that task.
The Bastani and Anthropic studies are covered in more detail in learning design with AI as a tutor and mentoring in the AI era. For the effect on people starting out, see being a junior designer in the AI era.
What exactly changes in the work?
Lee et al. do not describe thinking that disappears, but thinking that moves. According to participants, effort shifts from gathering information to verifying it, from solving the problem to integrating the AI’s response, and from executing the task to overseeing it. In other words, AI removes searching work and hands you reviewing work. If you skip the reviewing, you get what this article calls losing judgment.
The authors also note that AI can reduce critical engagement, especially on routine or lower-stakes tasks. Their argument is that without practice on everyday tasks abilities can deteriorate, and then they fail exactly when the hard case arrives. They recall the “irony of automation” that Bainbridge described in 1983: if the machine handles the routine and the person only deals with exceptions, the person loses the practice needed to handle them.
Where does judgment get lost?
It gets lost in concrete moments, which are a practical reading of the above rather than a measured finding:
- When you accept the first output. The answer is plausible, so it looks sufficient.
- On tasks that seem minor. If the task matters little, the review gets lighter.
- When you explain without the tool. If the AI chose and you only approved, the decision is not yours.
Two signs you delegated too much, as self-observation: you cannot explain a number or a claim without asking again, and you do not know how many alternatives existed because you took the only one you were given. If they show up, go back to a blank sheet, write the problem in your own words, and only then compare with the output.
What separates good automation from bad?
Good automation means the AI does one step and you keep judgment over the result. Bad automation means the result goes out without anyone questioning it. A practical test: if the output were wrong, would you notice within a minute? If not, that step should not be automated without a defined review.
Automation is the right call when the task is repetitive, the output can be verified quickly, and an error is cheap and reversible, like formatting a table or renaming files. There, protecting judgment does not mean reviewing every detail, but knowing what to check. To separate what AI can do alone from what needs supervision, see human-in-the-loop and hallucinations.
How do you keep your judgment without giving up AI?
- Write your hypothesis before asking for help, even in one sentence.
- Ask for counterarguments: “What could go wrong with this decision?”
- Define in advance the criteria you will use to accept or reject the output.
- Verify at least one claim against an external source, such as a real user or a system figure.
- Before you close the task, explain the decision without opening the tool.
A weekly practice, proposed by this wiki and not the result of a study: do a low-risk task without AI, with a timer running, and compare it with what the tool produces. Note three columns: what the AI did better, what you did better, and what neither resolved well. The third shows the gap the tool does not cover.
Example
Illustrative example: a designer uses AI to summarize twenty interviews. The summary says most users want an export button. Before accepting it, the designer counts how many interviews actually mention it. There are four, not eighteen. The summary had grouped vague comments about “getting my data out” as if they were a request for export. The missing criterion was a simple rule: a quote counts only if the user said it in those terms. The tool saved time, but verification prevented a wrong decision.
Common mistakes
- Trusting the tool more when the task seems simple.
- Accepting a summary without checking it against the source.
- Using AI to decide and then looking for arguments to justify the decision.
- Letting basic skills atrophy because “AI already does it”.
- Measuring productivity only by speed.
- Quoting a survey or a preprint as if it proved that AI damages thinking.
Sources
- Lee et al., The Impact of Generative AI on Critical Thinking (CHI 2025), Microsoft Research: survey of 319 workers, 936 examples, confidence in AI and in oneself, study limitations.
- The Register, Some workers already let AI do the thinking for them, Microsoft researchers find (February 2025): coverage of the study and the authors’ affiliations.
- Bastani et al., Generative AI without guardrails can harm learning (PNAS, 2025), IDEAS/RePEc record: experiment with nearly a thousand students, 48%, 127% and 17%.
- Anthropic, How AI assistance impacts the formation of coding skills (January 29, 2026): trial with 52 engineers, 50% versus 67%.
- Kosmyna et al., Your Brain on ChatGPT (arXiv, preprint): 54 participants, EEG, groups with AI, search engine and no tools.