Thinking Without Acting, Acting Without Thinking in the Age of AI

For programmers, product people, students and anyone who uses AI every day. You do not need psychology or statistics. Every hard term gets a short explanation.

This is a reflection, not a scientific conclusion. I read the studies from the original paper or the authors' page when I could find it, and from summaries when I could not. Many studies are small or new, and some are not peer reviewed. I note each weakness where it appears. "My opinion" parts are opinion, not measurements.

---

Quick summary#

• AI made both the product of thinking (plans, summaries, comparisons) and doing (code, drafts, prototypes, chains of steps run by an agent) cheap. That opens two opposite traps: thinking without acting, and acting without thinking.
• Trap one: feeling and reality can diverge. People feel they understand more before they have to explain (Rozenblit and Keil, 2002), developers felt 20% faster while measured 19% slower (METR, 2025), and people who let AI write for them scored lower on a quiz than people who wrote themselves (Anthropic, 2026).
• Trap two: running is not the same as right, safe or reversible. People with an AI assistant wrote less secure code and felt more confident (Perry and colleagues, Stanford), and an agent deleted a real database during a "code freeze" (Replit, 2025).
• No study says AI is always harmful. Other Copilot trials found speedups. The point is narrower: do not confuse having an answer with understanding, and do not confuse it runs with it is right.
• My suggestion is a small loop. First ask "can we get back after this?", then build the smallest piece in an hour somewhere failure is harmless, check and explain it back without looking, measure the real time, write down the decision.

---

Hard words, plain meaning#

---

1. What changed#

Picture a beautifully printed map when you have not taken one step. Before, getting a plan, a comparison or an explanation took reading, asking and writing. That effort was also when we learned. Now AI returns all three in ten seconds, tidy, confident and complete.

What got cheaper is the product of thinking. What did not get cheaper is understanding and trying. The two are easy to mix up because the page looks the same whether we understood or not.

It is also true that AI made doing cheaper: writing code, drafting, building prototypes, even handing an agent a chain of steps to run. So delaying a try costs more than it used to, and waiting for the perfect plan before acting is a worse choice than before.

But both got cheaper, so there are two opposite traps:

• Thinking without acting: holding a fluent page and stopping there.
• Acting without thinking: pressing run on something that looks fine without understanding it, checking it or knowing how to go back.

This piece goes through both: parts 2 to 5 for the first trap, part 6 for the second, and part 7 puts them together.

The question of this piece is not "does AI make us dumb". It is smaller and more useful: when are we really thinking and when are we just reading something that looks like thinking, and when are we really doing and when are we just pressing run?

---

2. Evidence one: the feeling of understanding comes before understanding#

In 2002 Leonid Rozenblit and Frank Keil at Yale had participants rate how well they understood some familiar objects (in the summary I read: a cell phone and a sewing machine). Then they asked for step-by-step explanations and rated again. The self-ratings dropped after explaining. They called this the illusion of explanatory depth, and found it strongest for "how does it work" knowledge.

Linking this to AI is my inference, not a result of the study: when we read a fluent explanation written by AI, we nod and feel we understand. But nodding is not being able to explain it back. The cheapest test is to close the laptop and say it in your own words.

Limit. A critique argues that part of the drop may come from participants using different scales before and after, not only from an illusion. The effect is widely cited, but do not treat the numbers as exact.

---

3. Evidence two: felt fast, measured slow#

In 2025 the METR group ran a randomized trial with 16 experienced open-source developers. They did 246 real tasks, about two hours each on average, on projects they had contributed to for years. Each task was randomly allowed or not allowed to use AI.

The point to remember is not the word "slower" but the gap between feeling and measurement. The authors list several possible contributors (time writing prompts, waiting, fixing the output; they report developers accepted under 44% of AI suggestions) without claiming which is the main one.

A very important limit. Only 16 people, early-2025 tools, large and familiar projects. The authors themselves say it does not mean future models will not help. And it is not the only study: an earlier one (Peng and colleagues, 2023) found Copilot users finished an HTTP server task 55.8% faster (the 95% confidence interval is very wide, 21% to 89%). Three field experiments at Microsoft, Accenture and a Fortune 100 electronics maker (4,867 developers) found about 26% more completed tasks, with larger gains for less experienced people.

Taken together, my lesson is: real speed depends on the person, the task and the tool; what does not change is that a feeling of speed cannot replace a measurement.

---

4. Evidence three: thinking less when the task feels light#

A group from Microsoft Research and Carnegie Mellon University surveyed 319 knowledge workers who use AI weekly (CHI 2025), plus nearly a thousand real examples of work they described. The main results, as reported by the respondents:

• The more they trusted AI, the less critical thinking they reported. The more confident in their own expertise, the more they reported (and the more effort it cost).
• On high-stakes tasks they said they put in more effort with AI. On routine, low-stakes, rushed tasks they said they put in less.
• Effort shifted from doing to verifying information, integrating answers and overseeing the work.

Limit. All of it is self-reported, one cross-sectional survey. It tells us how much people felt they thought, not how well they caught errors. It also does not prove skills fade over time.

What I keep from it is the pattern by stakes: people think hard when the task matters and let go when it is light. That sounds reasonable, but "light tasks" add up to most of a working day.

---

5. Evidence four: learning is not the same as producing#

This is the closest to "thinking without acting", because it asks: if AI finishes the work, do we learn?

Anthropic, 2026. A randomized trial with 52 developers (mostly junior) learning a new Python library called Trio. The group with AI scored lower on a quiz about concepts, reading and debugging code than the hand-coding group: about 50% versus 65 to 67% (depending on the summary). Completion time was almost the same (about 23 versus 25 minutes, not statistically significant). Worth noting: how AI was used mattered. People who asked follow-up conceptual questions while working kept more than people who only asked it to write. Limits stated by Anthropic itself: small sample, short time, one library.

MIT Media Lab, 2025. A study with 54 people writing essays while EEG was recorded. The ChatGPT group showed the weakest neural connectivity, the hand-writing group the strongest, the search group in between. News reports gave a figure: right after writing, most of the ChatGPT group could not quote a sentence from their own essay (I have not checked this against the paper). This is a preprint, not peer reviewed, small, and another group published a commentary arguing there is not enough evidence to say AI lowered performance. The authors themselves asked the press not to use words like "dumber". EEG measures engagement during a task, not intelligence.

Gerlich, 2025. Surveys and interviews with 666 people: frequent AI use went with lower critical-thinking scores, with offloading as the link in between. The author stresses this is correlation, not cause.

None of the three is enough to conclude. But together with the psychology of memory, the direction is fairly consistent: recalling and producing yourself is remembered longer than rereading and receiving a ready result. Roediger and Karpicke (2006) found that mixing in self-tests slowed learning at first but helped memory a week later. Slamecka and Graf (1978) found that things we generate ourselves are remembered better than things we only read. Robert Bjork calls this "desirable difficulty": some effort while learning buys lasting memory.

---

6. The opposite trap: acting without thinking#

If we stopped at the part above, this would turn into a slogan: "just do it". It is not, because AI made doing cheap, and that has a trap of its own: we press run before we ask. Running is not the same as right, safe or reversible. Four pieces of evidence, each with its limit.

AI-written code often has security holes. Pearce and colleagues (2021) had Copilot write code for 89 scenarios tied to security weaknesses, producing 1,689 programs, and about 40% were judged vulnerable. A 2023 replication with newer Copilot versions found the rate had fallen (for Python, from about 35% to about 25%), so the number moves with the version. Veracode (2025), a security vendor, reported that code from more than 100 models had security flaws in 45% of 80 curated tasks, and for cross-site scripting only about 12 to 13% was secure. This is a vendor report and I read it through the announcement, so check the method before quoting it.

With AI, people were more confident even though their code was less secure. Perry and colleagues (Stanford, 2022, published at ACM CCS 2023): 47 people did five security tasks. On four of the five, the group with an AI assistant wrote less secure solutions, and believed their code was more secure. People who trusted the AI less and took care writing their prompts wrote fewer flaws. Limits: small sample, an older model, security tasks only.

Over-trusting machines is an old story. Parasuraman and Manzey (2010) reviewed research on "automation bias": when a tool's suggestions are sometimes wrong, people make both omission errors and errors of following the wrong suggestion, and according to the paper's summary this does not go away with practice, in novices and experts alike. It is common ground for every automated tool, not only AI.

When an agent works without brakes. In July 2025, SaaStr founder Jason Lemkin said the coding assistant Replit, during a 12-day trial and under a "code freeze", deleted the company's database (press reported roughly 1,200 executive records and nearly 1,200 companies). The assistant said recovery was not possible, and it later turned out recovery worked. Replit's chief executive called it unacceptable, and the company acknowledged that it was then using one database for preview, testing and production. The sources are news reports and an AI incident database, not an investigation report, and details differ between sources.

The shared pattern is something done first and thought about later. A "code freeze" was only a sentence in an instruction, not a barrier in the system. This is my opinion: with agents, the limit belongs in the system (separate environments, just enough permission, backups), not only in the instruction. The two cases in Who Is Accountable When AI Is Wrong? (a chatbot misstating a policy, lawyers filing unchecked case law) are of the same kind.

---

7. Two traps, one map#

Talk can replace action. Jeffrey Pfeffer and Robert Sutton (Stanford, 1999) described a "knowing-doing gap" in companies: talking about a problem gets treated as solving it, and fluent, clever speakers are rewarded, so a "smart talk trap" appears. They wrote before the AI era. AI only makes smart talk cheaper.

Skilled people think while acting. Donald Schön (MIT, 1983) described good practitioners as having a "reflective conversation with the situation": they take a step, look at the feedback, then adjust. Thinking and doing are one loop, not two phases in a row. He also admitted that reflection can sometimes get in the way of action.

Putting the two traps together, I draw a four-cell map (my opinion, not a measurement):

The goal is not thinking less or doing more. It is shortening the loop between thinking and doing, and knowing when to think first.

One-way doors, two-way doors. In his 2015 shareholder letter, Jeff Bezos split decisions into two kinds: hard or impossible to reverse ("one-way doors"), which need to be made slowly, carefully and with consultation, and reversible ones ("two-way doors"), which should be made quickly by high-judgment people or small teams. He warned that the common mistake is treating two-way doors as one-way, which leads to slowness and fear of trying (I read this through a summary, not the letter). From it I draw one first question before handing work to AI: "if I do this, can I get back?"

---

8. An illustrative scenario (not a real case)#

This is a situation I made up to make it easy to picture. A team of four wants to add an article recommendation feature. There are three ways.

Way one: a lot of thinking.

Way two: acting without thinking.

Way three: the small loop.

Way three is not "winging it" and not "thinking forever". It still thinks, but with something visible, and it still acts, but where a failure can be undone. The scenario is only an illustration. What ties it to reality is the results above: the feeling of "faster" in the people involved can differ from the measurement (METR), and people with AI can be more confident even though their code is less secure (Perry), so way three adds two more steps: check what the AI produced and write down the real time.

---

9. A small loop (my opinion)#

This is not a research result. It is how I gather the ideas above into a habit:

1. One question, and a reversal question. Write what you truly want to know, in one sentence that can be answered yes or no. Then ask: can we get back after doing this? If not, stop to think and back up first.
2. The smallest piece of work, in one hour, in a separate environment. The ugliest running version beats the prettiest plan. Use AI to do, not only to talk about doing, and do not give it more permission than the work needs.
3. Check, then explain back without looking. Read what the AI changed, run it. Then close the screen, say or write how it works in your own words. Wherever you stumble is where you have not understood.
4. Measure. Write down the predicted time and the real time.
5. Decide and record. Drop, keep or change direction, with one line of reason.

The measuring step fits in a small table. The formula for the gap: (actual minus predicted) divided by predicted.

The data in the table is an example for you to recompute. After a few weeks of logging you will see which way your feeling is off, and which tasks to hand to AI and which to do yourself.

Five signs you are thinking without acting:

• You have more versions of the plan than test runs.
• You ask AI the same question again in different words, hoping the answer feels "firmer".
• You read it and feel you understand, but cannot explain it to someone else.
• You have saved more articles, courses and "read later" tabs than things you tried.
• Each time you are about to start, you feel you need to "look into it a little more".

Five signs you are acting without thinking:

• You press accept without reading the changes.
• You test straight on real data because it is "convenient".
• You do not know how to go back if it breaks.
• You gave an agent more permission than the work needs.
• It runs, so it is done, and nobody can explain why it runs.

---

10. A balanced view#

• Many studies are small, self-reported, from a vendor, or new. METR has 16 people and early-2025 tools. Anthropic used one library. Microsoft is self-reported. MIT is not peer reviewed. Gerlich is correlational. Veracode is a security vendor's report. Pearce and Perry used 2021 and 2022 models.
• The Replit case comes from the press. One incident, retold through news and an incident database, and details differ.
• I may have picked studies that fit my argument. I added the Copilot trials that found speedups to balance it, but someone else might choose differently.
• Careful thinking is needed, above all for one-way doors. High-risk work, such as security, money and health, deserves a pause to think and check first. The Microsoft respondents themselves said they put in more effort on important tasks.
• AI can help learning. In Anthropic's trial, people who asked conceptual questions kept knowledge better. How it is used matters more than whether it is used.
• The small loop and the four-cell map are not laws. Some work needs a week of study before touching anything, and some should simply be done because it can be undone. What I try to avoid is planning instead of trying, and trying instead of checking, not planning or trying in general.

---

11. Conclusion#

AI made the product of thinking cheaper, and it made doing cheaper too. What did not get cheaper is real understanding, checking, and knowing it works in the world without breaking anything. Two traps grow from that: we easily hold a fluent page and believe we are halfway, or press run on something that looks fine and believe it is right.

Both traps loosen the same way: shorten the loop between thinking and doing, and ask "can we get back?" before acting.

One small step today: take one thing you were about to plan or about to hand entirely to AI. Ask "if I do this, can I get back?" Build the smallest version in an hour in a place where failure does not matter, read carefully what the AI produced, then explain it in your own words without looking at the screen. Write down the predicted time and the real time. That number will tell you more than the feeling does.

---

References#

Open the original paper before quoting any number above.

• Rozenblit and Keil (2002), "The misunderstood limits of folk science: an illusion of explanatory depth", Cognitive Science (read through summaries, not checked against the paper)
• METR (2025), Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity and the arXiv version
• Lee and colleagues (2025), The Impact of Generative AI on Critical Thinking, Microsoft Research and CHI 2025
• Anthropic (2026), How AI assistance impacts the formation of coding skills and Shen and Tamkin, How AI Impacts Skill Formation (arXiv)
• Kosmyna and colleagues (2025), Your Brain on ChatGPT (arXiv preprint, not peer reviewed) and a critical commentary
• Gerlich (2025), AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking, Societies
• Stanford News (1999), on the book The Knowing-Doing Gap by Pfeffer and Sutton
• Roediger and Karpicke (2006), Test-Enhanced Learning, Psychological Science
• Peng and colleagues (2023), The Impact of AI on Developer Productivity: Evidence from GitHub Copilot (arXiv)
• Cui and colleagues, The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers (SSRN, Management Science)
• Pearce and colleagues (2021), Asleep at the Keyboard? Assessing the Security of GitHub Copilot's Code Contributions
• Perry and colleagues (2022), Do Users Write More Insecure Code with AI Assistants? (Stanford, ACM CCS 2023)
• Veracode (2025), press release on the GenAI Code Security Report (a vendor report, read through the announcement)
• Fortune (23 July 2025), the Replit database deletion and AI Incident Database, Jason Lemkin entry (news and incident-database sources)
• Parasuraman and Manzey (2010), Complacency and Bias in Human Use of Automation, Human Factors (read through the abstract)
• A summary of Amazon's 2015 shareholder letter on Type 1 and Type 2 decisions (Rework) (secondary source, the letter itself not read)
• Slamecka and Graf (1978), "The generation effect: Delineation of a phenomenon", Journal of Experimental Psychology: Human Learning and Memory, 4(6); Bjork (1994) on "desirable difficulties"; Schön (1983), The Reflective Practitioner, Basic Books (read only through summaries)
• Related posts on this blog: Learning With AI Without Losing Your Skills, Why Not to Trust AI Blindly, Reviewing AI-Written Code, Who Is Accountable When AI Is Wrong?