The leap AI can’t make is what graduate programmes don’t teach

Machines have mastered pattern-finding and are conquering formal proof. What they lack is the abductive leap that produces new explanations. Graduate programmes train everything except that leap, and one corner of the academy already knows how to teach it.

In January, a position paper from Google DeepMind asked whether a modern AI, given everything Albert Einstein knew in 1907, could have invented general relativity. The author, Tom Zahavy, concludes that it could not.

His reason matters more than his conclusion. Machines have mastered induction, the finding of patterns in data. They are rapidly mastering deduction, the unfolding of consequences from premises. What they lack is abduction: the generation of a new explanation for a surprising observation.

Einstein’s happiest thought, the falling observer who feels no gravity, was not compressed from data. There was no data. It was a leap.

The creative spark

The DeepMind paper has travelled widely, mostly as a comfort. The creative spark, readers conclude, remains ours.

The rector of Bocconi University Francesco Billari argued in University World News in April that machines uniquely make human skills more decisive: creativity, judgement, intuition. He is right. I made a version of the same argument in AACSB Insights in July: the capabilities that matter are the ones AI cannot perform.

But lists of that kind, mine included, stop one step short of a design question: which cognitive act, exactly, and can it be trained?

The DeepMind paper answers the first half. If abduction is the one act machines cannot perform, graduate professional education now holds a specification of what it should be building.

By that specification, most programmes are training everything except the thing that matters.

What programmes reward is the machines’ territory

Look at what a typical MBA or executive masters programme rewards. Recognising which framework applies to a case is induction from worked examples. Deriving the implications once the framework is chosen is deduction.

Both are exactly the operations the machines are conquering. What programmes rarely train, and almost never assess, is the move in between: noticing that the case does not fit, holding the discomfort, and generating the explanation nobody supplied.

Eight years of teaching working professionals across nine Executive MBA cohorts at CFVG (Centre Franco-Vietnamien de Formation à la Gestion) in Vietnam have demonstrated to me the same pattern in every intake. Managers can recite frameworks fluently and cannot deploy them under pressure on a common Tuesday afternoon.

Deployment is not recitation. It is abduction: deciding what this surprising situation is a case of.

Generative AI has made the gap more dangerous, not less. In May I argued that students under time pressure hand the planning, monitoring and evaluation of their own thinking to the AI model.

The consequence for invention is immediate. Outsource those functions and the surprise that triggers abduction never registers. The essay is adequate. The leap never happens.

Sociology proceduralised the leap six decades ago

Here is the part of the story higher education has forgotten it already knows.

When Barney Glaser and Anselm Strauss built grounded theory in 1967, they were proceduralising exactly the move the DeepMind paper says machines cannot make: generating new explanatory categories from surprising observations. Constant comparison. Theoretical sampling. The discipline of the memo.

None of it waits for inspiration. The procedures force the leap on a schedule.

Sociologists have argued since, most sharply Stefan Timmermans and Iddo Tavory in 2012, that grounded theory’s engine was never induction at all but abduction, and that the leap runs on a prepared mind: a researcher who carries explicit prior positions into the data, so that surprise has something to strike against.

The jump, in other words, is not a gift. It is a practice. Priors have been built.

What training ‘the jump’ looks like for a manager

What would that practice look like for a manager rather than a field researcher?

Write your position down before you consult any tool. Otherwise, the machine’s first coherent frame becomes yours. Before a meaningful decision, record what you expect, what would prove you wrong, and the date you will check. Once a week, in prose, compare the week you lived against the principles you hold. State where each principle would mislead you.

Readers of my May commentary in University World News will recognise these disciplines. I have since organised them, with two companions, as The Learner’s Discerning Practice Model.

The label matters less than the mechanism. Each practice manufactures the two preconditions of abduction that the DeepMind paper identifies and grounded theory institutionalised: explicit priors, and a structured collision with surprise.

The disciplines were drawn from three decades of commercial leadership and the reflective practice tradition, not from sociology. The convergence with grounded theory is one of shared ancestry, not derivation. It is the convergence that persuades me that the mechanism is real.

Which acts does the programme reserve for the student?

The question dominating curriculum committees this year is how to integrate AI into teaching. It is the wrong first question.

The right one is which cognitive acts the programme deliberately reserves for the student, and how it trains them.

A programme can hand every inductive and deductive task to the machines and lose nothing that matters, provided it builds, somewhere in the week, a protected area where students state positions before consulting, predict before deciding, and write the comparison in their own words with a date attached.

That space costs no budget and no software. It costs conviction, and a propensity to assess something harder to grade than framework recognition.

The DeepMind paper ends by proposing that machines may one day acquire world models rich enough to simulate their way to new axioms. Perhaps. But the timeline of that speculation is not the timeline of the students in front of us.

They will spend their careers in the territory between what machines do well and what machines cannot yet do at all. The width of that territory will differ for each of them. It will be set by whether anyone trained their capacity to jump.

Sociology showed the capacity is trainable. The machines have now shown us why we must.

Dr Phuong Nguyen (MBA, PhD) draws on three decades of commercial leadership and talent development, along with the wider literature on adult and professional learning, in his teaching at CFVG (Centre Franco-Vietnamien de Formation à la Gestion) in Vietnam.

This article is a commentary. Commentary articles are the opinions of the author and do not necessarily reflect the views of University World News.