Can You Train Working Memory? Part Two: What Expectation Contributes

Abstract

Cognitive training cannot be blinded. There is no way to make a memory game and a control game feel identical while ensuring only one contains the active ingredient, which means every trial in this field carries an unmeasured expectation effect.

A 2022 study in the Proceedings of the National Academy of Sciences measured it directly, by deliberately inducing positive expectations in half its participants and negative expectations in the other half, crossed with real or control training.

Twenty sessions of working memory training and a sentence about what to expect produced effects of roughly the same size. On the intelligence measures, only the sentence mattered.

1.The Problem This Study Was Built For

In a drug trial, expectations are controlled by making the treatment and the placebo indistinguishable. Jocelyn Parong, Aaron Seitz, Susanne Jaeggi and Shawn Green begin by pointing out that this is structurally impossible in behavioural training.

Whatever the active ingredient of a training programme is, it necessarily changes how the programme looks and feels. A memory task and a trivia quiz cannot be made perceptually identical without removing the thing that distinguishes them.

The field's usual response is to choose control activities that generate similar expectations, then check whether the matching worked. Parong and colleagues took the opposite approach: rather than minimising the difference in expectations and hoping it was small, they maximised it, to find out how large an effect expectation alone could produce.

2.The Design

A two-by-two factorial: working memory training on an adaptive n-back task versus an active control task of general knowledge trivia, crossed with a positive expectation condition and a negative one.

Expectations were induced two ways at once. Explicitly, participants were told the training would improve — or would worsen — their performance on the tasks they had just done. Associatively, halfway through, they completed a fake assessment that had been secretly made easier or harder than the original, so that they experienced evidence matching what they had been told.

That second manipulation is the one that makes this study unusual. It borrows directly from placebo analgesia research, where the strongest effects come from combining a verbal suggestion with an experience that appears to confirm it.

Participants completed twenty sessions of twenty minutes across several weeks, with a ten-task battery before, after, and again after being fully debriefed about the deception. The study was pre-registered and used separate research teams to deliver the expectation manipulation and to run the testing, so that neither participants nor testers knew which expectation condition anyone was in.

Of 263 people recruited, 193 completed all sessions and 125 provided complete data for the main analysis — roughly thirty per cell. Attrition did not differ by condition.

3.What Happened

At the immediate post-test, both manipulations worked. Those who received real training outperformed those who did the trivia task, with a partial eta squared of 0.16. Those given positive expectations outperformed those given negative expectations, with a partial eta squared of 0.14, regardless of which training they had done.

Twenty sessions of adaptive working memory training bought approximately what a sentence and a rigged practice test bought.

There was no interaction between the two. Their effects were additive: the best-performing group was trained and optimistic, the worst was untrained and pessimistic, and expectation contributed the same amount either way.

Then the participants were told the truth, and the battery was run once more. At that delayed test the training effect not only survived but grew, to a partial eta squared of 0.20. The expectation effect was no longer significant.

Two things follow. Expectation is real and large enough to account for much of what short training trials report — and it is also the part that does not last, while whatever training does appears to persist.

4.The Finding Inside the Finding

Tor Wager and Heejung Jung of Dartmouth, writing a commentary on this paper in the same journal, drew attention to a pattern in the task-by-task breakdown that the original authors passed over.

The two fluid intelligence measures — Raven's Advanced Progressive Matrices and the University of California Matrix Reasoning Test — showed an effect of the expectation manipulation and no significant effect of training.

Wager and Jung call this a provocative suggestion that widely used measures of cognitive ability can be influenced by participants' beliefs, and it is difficult to improve on that phrasing. The tests that the entire far-transfer claim depends on responded to being told they would respond.

Read alongside Part One, where fluid intelligence did not move across 2,729 participants, the picture is consistent: matrix reasoning scores are movable, but not by memory training.

Wager and Jung add a second observation. The expectation effect was numerically about twice as large in the control training condition as in the real training condition — an underadditive pattern suggesting either ingredient may be sufficient on its own, though they are careful to note the study was not powered to test that.

5.How Firm Is This

Firmer than most in the field, and not firm enough to be the last word. The pre-registration, the double-unaware structure and the twenty-session length are all genuine strengths.

Against that: the primary analyses tested directional hypotheses, so the expectation effect at p = 0.044 would not survive a two-tailed test. Fewer than half the recruited participants entered the final analysis. And Wager and Jung note that with about thirty people per cell, the task-by-task breakdown — including the fluid intelligence pattern above — is underpowered and needs replication.

One exploratory finding is worth recording despite that. The people most susceptible to the manipulation were those highest in growth mindset — the belief that intelligence is malleable and improves with effort. The belief that makes someone buy brain training is the belief that makes the suggestion work.

Editorial Comment

We have no product in this category and nothing to gain from either result. We are covering it because the structural problem it describes is one we are subject to.

A supplement can at least be made perceptually identical to a placebo, which is more than a training programme can manage. But the same expectation machinery operates on someone who has paid for a bottle, read the label and decided to feel the difference — and it is measurable at roughly the size of a real intervention effect.

The honest inference is not that everything is placebo. Training beat control here, and the training effect was the one still standing after the deception was revealed. The inference is that in any unblinded or imperfectly blinded intervention, a substantial part of the short-term effect belongs to expectation, and only a follow-up after the expectation has been dismantled can tell you which part.

Part Three covers the possibility this literature almost never raises: that training can leave you worse off than not training.