The session with the AI flowed. You understood everything, you were gently corrected, you closed the laptop feeling it had gone well. Then you met a real person and froze. The feeling of learning used to be a fairly honest signal. AI has just made it dishonest.
You spend an hour with an AI tutor. The conversation flows. You understand everything it writes, it understands everything you write, and where you go wrong it corrects you gently. You close the laptop feeling that was a good session.
The following week you meet a real person, and you freeze.
What happened has a name, and the name is not something AI invented. But AI has made it happen faster, more smoothly, and to more people than ever before.
Cognitive psychology has known since the 1990s that people misread ease as mastery. Adam Alter and Daniel Oppenheimer call this processing fluency. Asher Koriat and Robert Bjork showed something worse: while you are actually studying, you predict what you will remember in a systematically wrong way — because the answer is sitting in front of you, so remembering it is deceptively easy.
That is exactly the highlighting-and-rereading trap this section described in its article on the testing effect. The illusion is not the new thing. The new thing is a tool engineered to make everything as smooth as possible.
A meta-analysis by Boning Lyu, Chun Lai and Jianing Guo found a medium-to-large effect for chatbots on second-language outcomes, strongest for speaking fluency, writing quality and motivation — though with substantial variation between studies. And there is one thing a chatbot does that humans find hard: it lowers speaking anxiety.
So this article is not saying AI is useless. It is saying something more uncomfortable: the same tool genuinely helps you and inflates your sense of how much it helped.
It removes retrieval. Asking is the opposite of remembering. Jeffrey Karpicke and Henry Roediger showed that the act of retrieval is itself the learning event. Every time you ask AI for the word instead of straining for it, you get the word and skip the learning.
It removes the desirable difficulty and the generation. Elizabeth and Robert Bjork's term: conditions that make you perform worse during practice often make you learn more in the long run — and AI is engineered to make you perform smoothly during practice. And by the generation effect, measured by Norman Slamecka and Peter Graf in 1978, reading a perfect AI sentence and nodding yes, that is what I meant feels just like having written it. It is not. You recognised it.
It removes the breakdown. This is the sharpest one. A real listener fails to understand you, and you are forced to say it another way — Merrill Swain called this pushed output. AI understands your broken input perfectly; it is designed to be robust to your errors. So you are never pushed. And as this section's article on fossilisation put it: an error that does not impede communication is an error nobody corrects, and it hardens.
It removes the clock. Fluency is automatisation under real-time pressure — the argument of the research on practice that Robert DeKeyser gathered. Typing to an AI has no clock at all: you can think for thirty seconds, send a perfect sentence, and still look fluent. A real conversation gives you about one.
Evan Risko and Sam Gilbert describe cognitive offloading: given a reliable external tool, the brain sensibly stops storing things itself. Betsy Sparrow, Jenny Liu and Daniel Wegner measured this with search engines — people remember where to look rather than what was there.
With a translator one tap away, your brain concludes, quite correctly, that this word is not worth filing. That sum is right for almost everything in life. It is fatal for exactly one thing: learning a language, where internal storage is the entire goal.
What the AI writes is fluent, so reading it is easy — and easy gets read as grasped. You understood it, and understanding feels a great deal like knowing. And here is the dangerous part: there is no failure signal. In an ordinary study session, struggle is the gauge. AI clears away the stumbling, and clears away the gauge along with it.
In 2026, a review by Sachin Kumar, Anna Mikayelyan and Olga Vorfolomeyeva gathered 41 studies of ChatGPT in classrooms from 2022 to early 2026 and gave the phenomenon its name: the fluency illusion. Their conclusion is worth copying out in full: efficiency, confidence and perceived understanding all rise; the evidence for durable learning and transfer remains mixed.
There is only one honest test: close the tab. Twenty-four hours later, with nothing open and nothing looked up, say the same thing you could say yesterday. Whatever survives is yours. Whatever evaporates was never yours yesterday either — it was the AI's, and you were standing next to it, watching.
Produce first, then check. This is the single most important rule and the one most often run backwards. Write the sentence with what you have, badly if necessary, and only then show it to the AI. Do it the other way round and you are reading someone else's work. And tell it to test you, not to answer you: Quiz me on these twenty words; do not show the answers. Same tool, role changed, and what happens inside your head changes completely.
Put a clock on the speaking — answer within a few seconds, no revising; without time pressure it is a writing exercise. Turn off hover-translation while reading, and look words up after you have guessed. And give AI the part where it beats people: endless patience, endless examples, a place to try speaking where nobody can embarrass you. Give humans the part no machine has: unpredictability, and communicative breakdown.
The mechanisms above — retrieval, desirable difficulty, the generation effect, pushed output, automatisation — have decades of research behind them and are solid. The part that is specifically about AI is two or three years old, the studies disagree with one another a great deal, most measure the short term, and some measure perceived gains rather than gains. Read this for what it is: a mechanism argument with early support, not a closed finding.
The illusion here is not AI is useless. It is subtler: AI makes the feeling of learning available without the learning — and for as long as anyone can remember, that feeling was a fairly honest signal. Struggle used to be our gauge, and it has now been removed. So stop measuring by feel: close the tab, wait a day, and try to say it with nobody holding you up.
The same article, told in plain words — for younger readers, or for anyone who wants the point quickly.
You spend an hour with an AI tutor. It goes well. You understand everything it says. It understands everything you say. When you get something wrong, it fixes it kindly. You close the laptop feeling good about it.
Next week you talk to a real person, and nothing comes out.
So what went wrong? Nothing, exactly. The hour happened. It just did not do the thing it felt like it was doing.
Psychologists have known this for thirty years. When something is easy to read or easy to follow, your brain quietly decides you have learned it.
That is why highlighting a book feels great and teaches you almost nothing. The words start to look familiar. Familiar feels like known. They are not the same thing.
AI did not invent this trap. AI is just extremely good at making things easy.
Researchers put many studies together and found that AI chatbots really do help people learn languages, especially speaking and writing. And they do one thing better than humans: they are never embarrassing. If you are scared to open your mouth, a machine that never judges you is a real gift.
So this is not AI is bad. It is stranger than that. The same tool helps you some, and makes you feel it helped a lot.
There is one more thing, and it is not your fault. When a tool is always there, your brain stops bothering to remember. Scientists found this with search engines: people remember where to look instead of the answer itself.
With a translator one tap away, your brain decides this word is not worth keeping. That is a smart decision for almost everything in life. It is a disaster for one thing only: learning a language, where keeping it is the entire point.
Normally, when you are not learning, you can tell. You stumble. You get stuck. That struggle is your warning light.
AI removes the struggle — and takes the warning light with it. Everything feels smooth, so nothing tells you it did not work.
Close the tab. Wait one day. Then, with nothing open and nothing to look up, try to say the same thing again.
Whatever comes out is yours. Whatever has vanished was never yours — it belonged to the AI, and you were standing next to it, watching.
The ideas about remembering and struggling have been tested for decades. They are solid. The part about AI is only two or three years old, the studies disagree with each other, and most of them only look at the short term.
So this is a good explanation, not a finished answer. If better research later shows the problem is smaller than we thought, that would be good news.
Until then: the feeling of learning used to be a fairly honest signal. It is not any more. So do not measure by feel. Close the tab, wait a day, and see what is still there.
These articles summarize well-established research in learning science and linguistics. Key sources and further reading:
The question assumes focus is a muscle you train. The strongest evidence says the training barely carries over, and the...
No time, no talent, too old — those are explanations, not causes. What can be measured is duller: one sentence with a ti...
Not one event but four, running on clocks a billion-fold apart: a wave at 400 milliseconds, a word locked in overnight,...
