Why your own judgement will not do

You already have an opinion about how you communicate. This lesson is about why it cannot be used.

The cleanest demonstration comes from a study of 238 government executives, average age 46.6, paired with somebody who disagreed with them and put through a five-round online conversation. Afterwards, each person rated how receptive they had been, and each partner rated how receptive the other had seemed.

Self-ratings ran higher than partner ratings, which is unremarkable. What matters is the next result. The participant's own rating of their receptiveness predicted their partner's rating at β = 0.13, p = .08, and their general dispositional self-rating predicted it at β = −0.05, p = .32. Neither reached significance.

Meanwhile a language algorithm, reading only the words they had actually typed, predicted the partner's rating at β = 0.29, p < .001.

And the consequence: the partner's algorithmically measured receptiveness predicted their willingness to collaborate further (β = 0.23, p < .001). The partner's self-reported receptiveness predicted nothing at all (β = 0.05, p = .42) (Yeomans, Minson, Collins, Chen & Gino, 2020).

Read that carefully. What you did was legible in your own words and predicted a real downstream outcome. What you thought you did was not connected to either.

A second, independent demonstration comes from selection interviews. Interview anxiety correlates between −.15 and −.49 with self-rated performance and between −.07 and −.28 with interviewer-rated performance (McCarthy & Goffin, 2004). In the largest field study available, 8,343 real applicants across 373 companies in 93 countries, the relationship between anxiety and observed performance was γ = −.03: statistically detectable and practically nothing (McCarthy et al., 2021).

There is a specific version of this that will affect how you use every exercise in this course, and it is worth naming now.

In the receptiveness studies, writers were asked to predict how their message would land. They systematically overestimated how willing raters would be to collaborate with them: predicted 4.81 against an actual 4.45, t(754) = 5.5, p < .001. And they completely failed to anticipate that receptive language would make them more persuasive. Their own prediction of that effect was β = 0.05, p = .506, against an actual effect of β = 0.24, p < .001.

So people are wrong in two directions at once. They think they are landing better than they are, and they do not know which of the things they do is working.

This is why the course is built on recordings and on somebody else's report rather than on reflection. Not because reflection is worthless, but because it has been measured against an external criterion in this exact domain and found not to track it.

One honest qualification. None of this says you know nothing. You hold information nobody else does, particularly about which specific moments you found difficult and which questions you had no answer for. That is a fact about the content and it is worth recording. What does not survive is the global judgement: how it went, how you came across, whether you were clear. Those are the ones to distrust.