Most answers to this question are about tact. Do not correct someone in front of their colleagues, do not fill a grieving person’s silence, do not offer the solution before the problem has finished being described. All reasonable, all about the effect your words have on other people.

There is a narrower category with experimental evidence behind it, and it works differently. In these situations the cost of speaking falls on you, and specifically on something you already knew.

Describing a face makes it harder to recognise

The original finding is Jonathan Schooler and Tonya Engstler-Schooler’s, published in 1990. Participants watched a video of a crime, then either wrote a detailed description of the perpetrator’s face or did an unrelated task, then tried to pick the person out of a line-up. The ones who had written the description performed worse. In two of their studies the gap was 22 and 25 percentage points.

The effect was named verbal overshadowing, on the theory that producing a verbal account interferes with the visual memory it is meant to describe. Faces are recognised as wholes and described in parts, and the description appears to get in the way of the recognition.

It is the kind of result that could easily have been a fluke of one laboratory in one decade. It was not left that way.

Why the replication is the interesting part

In 2014, a Registered Replication Report published in Perspectives on Psychological Science put the effect through a test very few classic findings have faced. The protocol was agreed in advance, the analysis was specified before any data came in, and the study ran in dozens of laboratories at once.

The first round, in 31 laboratories, followed the original timing: describe the face, then a twenty-minute filler task, then the line-up. The pooled effect was a drop of 4.01 percentage points, with a confidence interval running from 0.87 to 7.15. Real, in the predicted direction, and much smaller than the original.

The second round, in 22 laboratories, moved the description so that it came immediately before the line-up rather than twenty minutes before. The pooled effect there was a drop of 16.31 percentage points.

So the effect exists, it is smaller than first reported, and how much smaller depends on how recently you did the describing.

That pattern is worth more than a cleaner result would have been. An effect that survives pre-registered testing across dozens of independent laboratories, with a specific moderator identified, is on considerably firmer ground than most things quoted in articles of this kind. It has also been shrunk to its real size, which is what the process is for.

The practical version is narrow and genuinely useful. If you are ever asked to describe someone you saw, and you will later be asked to identify them, the describing is not free.

Explaining a preference can move it away from the thing you liked

The second case is about judgement rather than memory, and it comes from the same researcher working with Timothy Wilson.

Their 1991 paper in the Journal of Personality and Social Psychology, Thinking Too Much, ran two studies. In the first, 49 students tasted five strawberry jams and rated them. Half were simply asked to rate. The other half were told to analyse why they felt the way they did about each one.

The jams had already been ranked by Consumer Reports’ trained tasters, which gave the researchers an outside standard to compare against.

The students who just tasted and rated correlated with the expert panel at 0.55. The students who analysed their reasons correlated at 0.11.

Their preferences had not become random. They had shifted towards the attributes that are easy to put into words, such as sweetness and tartness, and away from whatever the trained tasters were responding to. The reasons they generated predicted their own final ratings almost perfectly, at 0.92. The explanation was doing the choosing.

The second study, with 230 students, applied the same manipulation to course selection. Students asked to write out reasons signed up for highly rated courses at a lower rate than students who were not, and the difference showed up partly in what they actually enrolled in that semester.

What this does not license

A good deal, and it is worth being exact, because this research is easy to bend into advice it does not support.

These are specific tasks: recognising a face you saw once, and evaluating options where some of the relevant information is hard to articulate. Neither result says that thinking about decisions is bad, that reasons are useless, or that first instincts are generally better. Where the important attributes of a choice are easy to state, analysing them should not hurt, and the studies do not test the many situations where deliberation obviously helps.

The samples were undergraduates at American universities, which is the standing limitation on a great deal of this literature.

Most importantly, none of this is about emotional conversations. Nothing here suggests that talking about how you feel damages anything, and the research on that question is a separate body of work with different findings. Reading verbal overshadowing as a case for keeping things to yourself would be a misuse of it.

What the two lines of research share is a specific and limited claim: some of what you know is held in a form that words do not carry well, and the act of translating it can overwrite the original.

That is a small category. It is also the only category where saying nothing is not a courtesy to somebody else but a way of protecting your own accuracy, which makes it worth knowing about separately from all the ordinary reasons for keeping quiet.