AI in Education and Learning

Education is one of the oldest testing grounds for artificial intelligence. Long before chatbots, researchers were building software that watched a student work through a problem and adjusted the next step. Today that lineage sits inside apps millions of learners open every day: practice that gets harder or easier as you go, hints that arrive at the right moment, and captions or read-aloud tools that open a lesson to someone who could not otherwise reach it.

The stakes are high because education shapes opportunity. A tool that genuinely helps a struggling student is a real gift; a tool that widens the gap between well-resourced and under-resourced schools, or that quietly teaches students to lean on a machine instead of thinking, does lasting harm. This lesson looks at where AI helps, where the evidence is thinner than the marketing, and why the teacher's role does not disappear.

Tutoring and adaptive learning

An intelligent tutoring system tries to do what a good human tutor does: track what a learner understands, notice mistakes, and offer the next problem or hint accordingly. Adaptive-learning platforms use the same idea at scale, routing each student along a path that fits their current level rather than marching a whole class in lockstep.

Feedback and grading support

AI can mark multiple-choice work instantly and can draft comments on writing or flag likely errors in code or math. Used well, this frees a teacher's time for the harder, human parts of feedback. Used carelessly, automated scoring can be gamed, can misjudge unusual but correct answers, and can bake in bias from its training data.

Accessibility

Some of the clearest wins are in access: live captions for deaf students, text-to-speech and speech-to-text for learners with dyslexia or motor differences, and translation that lets a newcomer follow a lesson. Here AI removes barriers rather than replacing instruction.

Picture a student stuck on fractions. An adaptive platform notices she keeps missing problems that involve unlike denominators, so it stops advancing, serves a short explanation, and gives three targeted practice items. When she succeeds, it moves her on. Her teacher opens a dashboard, sees the same pattern across six students, and pulls those six for a five-minute small-group reteach. The software handled patient, repetitive practice; the teacher handled diagnosis, encouragement, and the decision about what to do next. That division of labour, machine for practice, human for judgment, is where these tools tend to help most.

Now imagine a school that buys the same platform and treats it as a substitute for teaching. Students are parked in front of screens for an hour a day with little adult contact. Some race ahead, but others click through hints without thinking, learn to farm the system for answers, and disengage. Meanwhile a neighbouring school with newer devices and reliable internet gets far more from the identical software. The tool did not fail on its own; the way it was deployed turned a possible help into a source of boredom and a widened equity gap.

In 1984 the educational psychologist Benjamin Bloom published what he called the "2 Sigma Problem." His research groups had found that students given one-to-one tutoring performed dramatically better than students in ordinary classrooms, by roughly two standard deviations in his studies. Bloom framed this as a challenge: could group instruction ever match what a personal tutor achieves? That question helped motivate decades of work on intelligent tutoring systems, including the Cognitive Tutor developed from cognitive-science research at Carnegie Mellon and later commercialised by Carnegie Learning for mathematics.

The honest summary is mixed. Reviews of intelligent tutoring systems, such as a 2016 meta-analysis by Kulik and Fletcher, report that well-designed tutors can produce meaningful gains, though typically smaller than Bloom's headline two-sigma figure and varying widely by system and setting. The lesson is neither hype nor dismissal: carefully built, well-studied tutoring can help, but no software has closed Bloom's gap, and results depend heavily on design and how the tool is used.

• "AI tutors already match a human teacher." The evidence shows real but modest and uneven gains, not a replacement for skilled teaching.
• "More screen time means more learning." Unsupervised drill can breed disengagement and hint-farming rather than understanding.
• "Technology automatically levels the playing field." Without devices, connectivity, and support, the same tool can widen gaps between schools.
• "Automated grading is objective." Scoring models can misjudge unusual answers and inherit bias from their training data.

You are shown five claims from ed-tech product pages, such as "personalised learning that replaces the tutor" or "instant, unbiased grading." For each, decide whether it is a realistic, evidence-backed benefit or an overstatement, and name what evidence you would ask for before trusting it.

Think Like a Maester: Ask what a tool actually does for a learner, and who is still needed in the room, before you believe it can teach on its own.

1. What does an intelligent tutoring system try to track, and how does it use that information?
2. Name two ways AI can genuinely improve accessibility for students.
3. What was Bloom's "2 Sigma Problem," and why did it matter for tutoring research?
4. Why can the same adaptive platform help one school and widen inequality at another?
5. Give one reason automated grading should not be treated as fully objective.

AI shows up in education as intelligent tutoring and adaptive practice, feedback and grading support, and accessibility tools that remove real barriers. The strongest historical thread runs from Bloom's 1984 finding that one-to-one tutoring works remarkably well, through decades of intelligent tutoring systems like the Cognitive Tutor, to modern reviews showing real but modest and uneven gains. The technology helps most when it handles patient practice while a teacher keeps diagnosis, motivation, and judgment. Its dangers, over-reliance, unequal access, and misplaced trust in automated scoring, are just as real. Used to support skilled teachers rather than replace them, AI can make learning more responsive; mistaken for a teacher, it disappoints.

• Prompting: Getting Good Results
• AI in Creative Industries