Can You Train Working Memory? Part Three: When Training Makes You Worse

Abstract

Almost the entire cognitive training literature asks one question: how much does training help? Part One found a small effect, mostly located in the measurement. Part Two found that expectation contributes about as much as the training does.

This part covers a question the field rarely asks. A 2023 study in Memory & Cognition tested whether training on one working memory task can make you worse at learning the next one — and found that it can.

The effect runs in one direction only, which is what makes it informative rather than merely discouraging.

1.Training as Skill Acquisition

Nan Ni at Kyoto University, with Susan Gathercole and Dennis Norris at the MRC Cognition and Brain Sciences Unit in Cambridge and Satoru Saito at Kyoto, work within a framework Gathercole and colleagues proposed in 2019 that reframes what training does.

On that account, training does not expand a capacity the way exercise expands a muscle. It builds cognitive routines — task-specific procedures for handling a particular kind of problem, acquired much as any skill is acquired.

This reframing explains Part One's results without any appeal to bad faith. If training builds a routine rather than a capacity, then transfer should occur exactly where the routine applies and fail everywhere else. Which is what the meta-analytic literature finds.

It also generates a prediction that the capacity account does not. If routines are skills, then an established routine can interfere with acquiring an incompatible one — the way a habit from one instrument gets in the way of learning another. Training should sometimes hurt.

2.The Experiments

Two experiments, each with two training phases. In Experiment 1, participants trained for three sessions on one of three tasks: backward digit span, backward circle span, or a colour change-detection task serving as the active control. All three groups then trained for three sessions on backward letter span.

Backward digit span and backward letter span are both verbal: recall a sequence of verbal items in reverse. Backward circle span is spatial: recall a sequence of locations in reverse. The colour task requires no serial order at all, which is what makes it the appropriate control.

The question was what phase one did to phase two. And the answer was that it depended entirely on which phase one you had done.

The group that had trained on backward digit span — verbal, like the target — outperformed the control group on backward letter span. Ordinary positive transfer within a domain.

The group that had trained on backward circle span — spatial — performed significantly worse on backward letter span than the group that had done the colour task. Not merely no benefit. Worse than not having trained at all.

3.The Asymmetry

Experiment 2 reversed the order. Participants trained first on backward square span, backward letter span, or the colour control, then all trained on backward circle span. Its hypotheses and analysis plan were pre-registered.

If the interference were simply a matter of two incompatible task types colliding, the damage should run both ways. It did not. Verbal training before spatial training produced no harm at all — if anything a weak positive transfer.

So spatial training impairs subsequent verbal training, while verbal training does not impair subsequent spatial training. The effect has a direction.

The authors' reading is that the two tasks call for different optimal routines, and that the interference sits in the acquisition process rather than in performance itself. The routines developed for backward circle span get in the way of developing the right ones for backward letter span; the reverse combination leaves the learner free to build what the new task needs.

A directional effect is much harder to explain away than a symmetric one. General interference, fatigue or boredom would degrade both orders equally. Something specific to what was learned is doing this.

4.What This Is and Is Not

The scale is small: three sessions per phase rather than the fifteen or twenty of a full training study, chosen because performance gains in this paradigm are concentrated in the first few sessions. The transfer being measured is from one training phase to the next, not the one-shot before-and-after transfer of a conventional trial.

Nothing here shows that using a brain-training app makes you worse at anything in daily life. It shows that under a specific pairing of laboratory tasks, prior training reliably impaired later learning, and that the impairment was direction-dependent.

What it does establish is that the sign of a training effect is not guaranteed. In a field where every study measures how much benefit occurred, a design that could detect harm found harm.

That should change how the null results in Part One are read. An aggregate effect near zero has always been interpreted as training doing nothing much. It is also consistent with training helping in some pairings and hurting in others, averaging out.

5.Where Three Parts Leave It

Working memory improves with training by a standardised mean difference of 0.18 on untrained tasks, and by 1.15 when measured with tasks resembling the training. Fluid intelligence does not improve, and the people whose working memory improved most did not gain more of it.

Roughly as much short-term benefit can be produced by telling participants the training will work, and on the intelligence measures specifically, expectation moved scores while training did not.

And training is not uniformly beneficial: in at least one well-controlled pairing, having trained made subsequent learning measurably worse.

The most economical account of all three is the routine framework. Training teaches you a procedure. Procedures transfer to problems they fit, fail to transfer to problems they do not, and occasionally get in the way. That is an unremarkable description of learning any skill — and it is a very different product from the one that has been sold.

Editorial Comment

MindHeaven® has no cognitive training product and this series recommends none. Nothing in it should be read as advice to stop doing something you enjoy: the finding above concerns learning rates on laboratory span tasks, not the value of mental activity.

We wrote three parts on a category we do not sell into because the reasoning generalises. An effect that shrinks by a factor of six when measured with an untrained task; a placebo contribution the size of the treatment; a possibility of harm that nobody looked for until a theory predicted it. Each is a question worth asking of any intervention, including ours.

The third is the one we would most easily miss. Supplement research asks how much a compound helps and almost never asks whether it could leave someone worse off on an outcome nobody measured. That is a gap in the field, not a claim about any particular product, and we would rather name it than wait to be asked.