The exercise and depression result is weaker than the headlines said

In January 2026 Cochrane published an updated review of exercise for depression, pooling 73 randomised trials, which allocate people to a treatment or a comparison group by chance, covering at least 4,985 adults. The coverage that followed settled on a line: exercise rivals therapy and medication. The review's own numbers support a more careful sentence, and the gap between the two is worth walking through, because it is a good illustration of how a finding gets firmer as it travels.

Against no treatment or a control condition, exercise produced a standardised mean difference of 0.67. That measure expresses the gap between the groups in units of the spread of scores within them, so 0.67 is a moderate to large improvement in depressive symptoms, pooled across 57 trials and 2,189 people. That is the headline result. Cochrane rated its certainty as low, meaning the true effect may be substantially different from that estimate.

How we know

The most informative number in the review is not the headline one. It is what happens when you keep only the well conducted trials. Cochrane repeated the main analysis using just the seven trials at low risk of bias, meaning those least exposed to a systematic distortion of the result. Across 447 people, the effect fell from 0.67 to 0.46, and its confidence interval, the range within which the true effect can reasonably be expected to lie, ran from 0.88 down to 0.04, which is to say it very nearly touched zero.

That pattern, where an effect shrinks as trial quality rises, is one of the most reliable signals in evidence synthesis. It usually means some of the apparent benefit in the larger pool came from how the weaker trials were run rather than from the treatment. It does not mean the effect is absent. It means the honest estimate is smaller and less certain than the pooled figure suggests.

Risk of bias here has specific meanings: whether the allocation to groups was properly concealed, whether everyone randomised was analysed in the group they were assigned to, and whether the person assessing the outcome knew which group a participant was in. That last one is hard in exercise trials, because you cannot blind someone to whether they have been exercising, and it is a structural weakness of this whole literature rather than a failing of any one study.

Where the headline came from

The rivals therapy framing comes from two comparisons. Against psychological therapy, exercise produced a difference of 0.03, essentially nothing, at moderate certainty, from ten trials covering 414 people. Against antidepressants, the difference was 0.11 in favour of exercise, again close to nothing, at low certainty, from five trials covering 330 people.

A null result is not the same as an equivalence result, and this is where the reporting slipped. Finding no significant difference between two treatments in a small set of trials mostly tells you the trials were not large enough to detect a difference if one exists. Demonstrating that two treatments are genuinely equivalent requires a study designed for that purpose, with a pre-specified margin, and these were not those studies.

The long term picture is thinner still. Nine trials covering 405 people followed participants after treatment ended, and produced 0.53 with an interval running from 1.11 to minus 0.06. That interval crosses zero, so the review cannot say the benefit persists. Cochrane rated that result very low certainty.

This is not a new pattern

Cochrane has been updating this review since long before the current version, and the previous edition, published in 2013, tells a strikingly similar story. It included 39 trials covering 2,326 people. Against no treatment or a control, the 35 trials with usable data, covering 1,356 people, produced a standardised mean difference of 0.62, with an interval from 0.81 to 0.42, which the authors described as a moderate clinical effect.

Then it did the same quality restriction. Keeping only the six trials with concealed allocation, intention to treat analysis and blinded outcome assessment, covering 464 people, the pooled result fell to 0.18 with an interval from 0.47 to minus 0.11, and was not statistically significant.

So the shape of the finding has held for over a decade: a moderate effect overall, a markedly smaller one in the best conducted trials. The 2026 update nearly doubles the trial count, 73 against the 2013 edition's 39, and the well conducted subgroup now reaches significance where it previously did not, which is real progress. But the underlying problem is unchanged. After more than twenty years of trials, the honest summary still rests on a handful of studies built well enough to trust.

Why it matters

None of this argues against exercise. The direction of every comparison favours it or is neutral, no comparison suggests harm, and a moderate benefit against doing nothing survives even in the better trials, if smaller. For a treatment that is cheap, widely available and carries other benefits, that is a reasonable evidence base.

What it argues against is the specific claim that got amplified. Cochrane itself hedged: the review's conclusions describe the evidence as low certainty and call for better trials. The confident version appeared downstream, in the coverage rather than the research. That distinction matters, because a reader who acts on rivals therapy is acting on a sentence nobody in the study wrote.

It also matters for what happens next. Exercise for depression has been studied for decades and the certainty is still low, which is a statement about how the trials were built rather than about exercise. The review's most useful contribution may be its account of what a better trial would look like: larger, properly concealed, analysed as randomised, and followed long enough to see whether anything lasts.