Artificial Intelligence in Education — Reading

Passage

A Enthusiasm for teaching machines is older than the computers that now run them. In 1958 the psychologist B. F. Skinner exhibited a mechanical device that presented a question, accepted an answer and advanced only when the answer was right. He predicted it would transform schooling within a decade. It did not, and the pattern has repeated with educational television, the personal computer, interactive whiteboards and online courses. Each arrived with a claim that instruction would be individualised and teaching transformed; each settled into a useful but unremarkable role. This history is worth holding in mind when assessing the current wave, not because the technology is equivalent, but because the claims are strikingly similar. B The clearest present gains are narrow. Adaptive practice systems, which adjust the difficulty of exercises to a learner's demonstrated level, work well in domains with unambiguous right answers — arithmetic, grammar drills, vocabulary retention. Meta-analyses find effects that are real but moderate, and comparable to those of other well-implemented interventions. Automated marking of short factual answers is now reliable enough for routine use, which returns teacher time to work that requires judgement. C Extended writing is where claims outrun evidence. Automated essay scoring has existed since the 1960s and correlates respectably with human markers, but the correlation is achieved partly through proxies — length, sentence variety, vocabulary range — that a student can learn to game without writing better. Systems have been shown to award high marks to sophisticated-sounding nonsense. The practical risk is not that the machine is wrong occasionally but that its criteria become the criteria, and students optimise for the measurable rather than the meaningful. D Generative models introduced a different problem, which schools initially framed as cheating. Detection tools were adopted quickly and proved unreliable, with false positives falling disproportionately on students writing in a second language, whose prose is often more formulaic. Several institutions withdrew them after wrongly accusing students. The more considered response has been to assess differently — in supervised conditions, or through drafts and discussion — which is expensive and has been adopted unevenly. E A less discussed risk concerns the effect on learning itself. Retrieval, struggle and productive failure are among the best-established mechanisms in the study of memory: material recalled with effort is retained far better than material re-read. A system that supplies a fluent answer on request removes precisely that effort. Early studies suggest students using such assistance perform well on the assisted task and worse on later unassisted tests than those who struggled through unaided — an effect long known in other contexts as the difference between performance and learning. F The equity argument runs in both directions and is genuinely unresolved. Free tools plausibly narrow gaps by giving students without tutors something approximating individual help. But benefit depends on knowing how to interrogate an answer rather than accept it, which correlates with the educational advantage the tools were supposed to offset. Access is also uneven: the capable models are increasingly paid. G The likely outcome resembles the earlier waves. Certain tasks — routine marking, differentiated practice, first-draft feedback — will be substantially automated, and that will matter. The relationship between a teacher who knows a student and that student will not be, because the difficulty in education was never mainly the delivery of information. Skinner's machine failed for that reason, and the reason has not changed. H A separate and more immediate concern has emerged around generative AI specifically: its effect on how students write. Tools capable of producing a competent essay on demand have forced a rapid reassessment of how written coursework can be used to assess learning at all, since a student's ability to submit fluent prose no longer reliably indicates that the student produced it, understood the source material, or could reconstruct the argument unaided. Some institutions have responded by shifting assessment back toward supervised, handwritten conditions reminiscent of an earlier era; others have tried to build the tools into the task itself, asking students to critique or improve an AI-generated draft rather than produce one from scratch, on the theory that evaluating an argument is a skill worth assessing even where composing one from nothing is no longer a meaningful test. Detection software marketed to identify AI-written text has so far proven unreliable in both directions, flagging genuine student writing as machine-generated often enough, and missing genuinely machine-generated text often enough, that several universities have abandoned it as a basis for disciplinary action. The debate is unresolved, and unlike most disputes in education technology, it cannot simply be deferred.

বাংলা অনুবাদ জমা দেওয়ার পর দেখা যাবে — আগে ইংরেজিতে বোঝার চেষ্টা করুন।