Havruta: The Oldest Form of Rigour, Rebuilt for AI

We confused the vehicle for the destination. An essay was never valuable because of the finished document. It was valuable because of the journey to it: the reading, the wrestling, the structuring, the deciding what to cut. The artefact was simply proof that the thinking had happened.

Now the artefact is free. A student can produce a clean, referenced, well-organised essay in the time it takes to make a coffee, and it proves nothing, because the thinking it used to stand for may or may not have occurred. The document looks identical either way. The honest response to that is to change what we assess. The reflex instead was to reach for detection.

Two claims, and both have to be made carefully, because this is a field where the headlines run far ahead of the data and an academic reader will notice.

The first claim is that using AI to avoid thinking costs you the thinking. The most solid evidence is a peer-reviewed study of more than three hundred knowledge workers, which found that the more a person trusted the AI, the less of their own critical judgement they brought to the task (Lee et al., 2025). It rests on self-report, so treat it as a strong signal rather than a final verdict, and be wary of the more dramatic “AI rots your brain” headlines, which so far come from small, unreviewed work. The careful claim does not need them, because it sits on a century of settled learning science. We remember less when we expect a machine to hold the information for us (Sparrow et al., 2011). We learn through effort, not ease: retrieving an answer outlasts re-reading it (Roediger and Karpicke, 2006); explaining our own reasoning to ourselves deepens understanding (Chi et al., 1989); the difficulties that feel unpleasant in the moment are often the ones that build durable learning (Bjork and Bjork, 2011); and students made to wrestle with a problem before being shown the method understand it better for the struggle (Kapur, 2008). Struggle is not a flaw in learning. It is the engine. Take it out and the learning goes with it.

The second claim is the one that turns a complaint into a method, and it comes from two studies that settle the argument when you read them together. In a large field experiment, school students were given an AI assistant for mathematics practice. The plain version, the one that simply gives answers, lifted their performance while they used it and then left them worse off on a later exam they sat without it (Bastani et al., 2025). They had leaned on a crutch and never built the muscle. A second version of the very same tool, redesigned to withhold the answer and make the student reason first, removed the harm. Separately, a randomised trial at Harvard gave students an AI tutor deliberately built to question rather than answer, and they learned more than twice as much, in less time, than in a strong, well-run active-learning class (Kestin et al., 2025).

Same technology. Opposite outcomes. The only variable is whether the machine makes the human think first.

Two caveats, because they make the argument honest rather than weaker. These are results about careful design, not about chatbots in general, and one excellent tutor in one course is a promising signal, not a settled science. The wider literature is sober about exactly how good tutoring gets: a careful review found well-built tutoring systems reach an effect size of about 0.76, close to the 0.79 of expert human tutors, and far short of the two standard deviations sometimes claimed for tutoring (VanLehn, 2011). The point is not that AI is magic. The point is that the same tool helps or harms depending on a single design decision, and across most of higher education we are making the wrong one.

Havruta is one rule, held with discipline. Configure the machine to do what a good study partner does and a vending machine never will: ask before it answers.

In practice, the tool does not produce the essay or the analysis on demand. First it turns the questions back on the student. Say the task is an essay on whether raising the minimum wage reduces employment. A Havruta-configured assistant does not write it. It asks: what is your thesis, in one sentence? What is the strongest piece of evidence for it? What is the best argument against you, and why does it fail? Only once the student has answered, in their own words, does the machine help: tightening the structure, pressing on a weak claim, offering a counter-example to test the case. The student does the thinking. The machine makes the thinking harder, and better.

This is not a gentler AI. It is a more demanding one, and it is demanding in exactly the way the learning science says matters. It forces the student to retrieve, to generate, and to explain their reasoning before they can move on (Chi et al., 1989; Roediger and Karpicke, 2006). It puts the desirable difficulty back in, on purpose. And it works on the one material a machine cannot fabricate, because it is the one thing the machine does not have: the student's own reasoning. We call the moment the tool turns the questions back the Flip, the point at which the machine stops answering the human and starts questioning them.

Because the space is crowded with adjacent ideas, it helps to be precise about the boundaries.

It is not “co-intelligence”, the popular notion that you should simply work alongside AI as a clever colleague. That is a posture, and a good one, but it is not a method. Havruta is a discipline with a hard rule: the machine questions first.

It is not an AI examiner that grills a student after they hand work in. That questions after the fact, to catch them out. Havruta questions before, as the price of producing anything at all.

It is not a generic Socratic chatbot that gently leads a student toward an answer the machine already holds. Havruta does the opposite. It extracts what the machine cannot hold: the student's own reasoning, which exists nowhere else.

And it is not detection by another name. It does not police a finished file. It changes what gets produced, and how it is judged.

There is nothing nostalgic in any of this. For most of its history the university examined people by making them think out loud, in dialogue, under pressure, and defend what they said. We replaced that with the silent, solitary, easily gamed document, and now the document can be generated in seconds, and we are surprised it tells us nothing.

The way forward is not to detect the machine. It is to put the dialogue back, with the machine on the other side of the table, asking the questions. That is the oldest form of rigour, rebuilt for AI.