Two questions, both small, both pointing at something larger.
When did you last know someone’s phone number by heart? When did you last do a calculation in your head, without a calculator?
A scene from an ordinary office
At nine in the evening, Sadia — the name is changed — asked an AI to produce a report for the office. “Write the whole thing.” Ten minutes later it was done.
The next day her boss asked a single question: where did this number come from?
She could not answer.
That is the entire problem in one exchange, and it is worth being precise about what went wrong. The report was not necessarily bad. The number may well have been correct. What failed was that the work had been produced without passing through anyone who could account for it — and when the question arrived, it arrived at her, not at the model.
The name for this
In psychology this is described as cognitive offloading: shifting the burden of thinking onto a machine.
It is not inherently bad, and the writer of the original piece — a Bangladeshi tech blogger and content creator, writing for Prothom Alo — does not claim it is. Offloading is why we have notebooks, calendars and calculators. Nobody argues that remembering fifty phone numbers was a valuable skill worth preserving for its own sake.
The concern experts raise is about dose. Done past a certain point, the warning goes, your own judgement and the practice of original thought get less exercise. The capacity does not vanish overnight; it quietly stops being maintained.
Three rules worth adopting
The original column offers three, and they are more practical than most advice in this genre.
Think first, then ask. Form your own view before you put the question to the model. This costs a few minutes and changes what you are doing with the answer: you are now checking a draft against your own reasoning, rather than receiving a conclusion with nothing to compare it to.
Never forward an answer you have not understood or verified. Because if a question comes later, the responsibility is yours. This is the rule Sadia needed the night before, and it is the one most easily skipped at nine in the evening.
Work without AI at least one day a week — exactly as muscles need exercise.
The third is the one that will get ignored, and it is the one with the clearest logic behind it. If the worry is that a capacity decays through disuse, then the remedy has to be use. Reading about the risk does not address it.
The GPS test
The column offers an analogy that holds up well: people who know the route get into the least trouble, even when they have GPS.
Note that this is not an argument against GPS. Nobody navigates an unfamiliar city by memory out of principle. The point is narrower and more useful — the person who also knows roughly where they are is the one who notices when the directions are wrong, and who still arrives when the signal drops.
Applied to AI: the colleague who could have written the report themselves is the one who spots that a figure is implausible. That capability is invisible when everything works, and is the only thing standing between a confident error and a published one.
Where this lands in Bangladesh
This matters here for a specific reason. A great deal of Bangladeshi professional work — agency reports, client deliverables, academic submissions, government paperwork — is judged on the finished document, with nobody asking how it was produced. That is an environment in which full offloading carries no immediate penalty, right up until someone asks Sadia’s question in a meeting.
And for students the stakes are different again, because the output is not the point. A submitted assignment is not a product; it is evidence that a capability was built. Offloading it entirely produces the evidence without the capability, which is a trade that only looks favourable until an examination, an interview, or a first job.
AI can be an assistant to your brain, the column concludes, but not a replacement. What we have to decide is whether we are having it think for us, or taking its help to think.
Related: what happens when the judgement moves to a machine.
