Page 6 of 12 · Pillar two

Notice

Read the room before you enter it. This is the pillar with the largest gap between what training moves and what it does not, and the design is built around admitting that.

The knowing-and-doing gap

Gates et al. (2017) meta-analysed group social-skills programmes and found gains that were almost entirely about knowing the skills (g = 1.15) rather than doing them (g = 0.28), with teacher ratings showing nothing at all. The 2024 randomised trial of the leading programme in 12–18-year-olds is typical: knowledge effect very large, parent-rated social responsiveness modest, teacher ratings flat, no blinded observation.

So we teach the rules, because teaching rules works and is cheap to grade - and we measure behaviour separately, expecting it to move much less. Nothing on this page is sold as making anyone socially skilled.

One negative note belongs here too. The Social Thinking framework has been argued to meet the definition of pseudoscience: many products, few controlled studies, and a rebuttal resting on "layers of evidence" rather than trials. Its vocabulary is fine as plain language. We do not cite it as evidence and we do not copy it.

Walk in

Moderate

A thirty-second scene of two to six people. Then four questions, one at a time: what is the topic; who looks open to being joined; where is the pause; what would you say.

Parameters
Three scenes per session; a pass means all four right; scenes escalate from two people to six across the programme.
Scored
Each step separately. The first three have keyed answers written by a person when the scene was made. The fourth is typed and rated against the scene's own facts.
Why this shape
These are the four entry steps from the PEERS programme - watch and listen, move closer, wait for a pause, mention the topic - with the matching assessment of whether you were accepted, and a script for leaving gracefully if you were not. They are the single most useful concrete artefact anyone has published for the walking-into-a-room problem.

Read the scene

Strong as a measureModerate as training

A short clip, then one question with four options: what does she feel, what did he mean, was that a joke.

Parameters
15 to 25 items per set, with a separate held-out set reserved for measurement so in-app gains are not memorisation.
Scored
Accuracy, and error type. Every wrong option is coded in advance as over-reading, under-reading, or taking it literally, so the feedback can be specific without being about you: "you read more intent into that than was there."
Why this shape
The research instruments in this format have good internal consistency (about 0.86) and retest in adolescents, are not fooled by socially desirable answering, and have adolescent norms showing teenagers score 6–11% below adults, with sarcasm and lie detection maturing latest. They are the template for a machine-gradable scene quiz.

One more look

Moderate

An ambiguous situation with a few clues. Two buttons: decide now, or look for more. Each look reveals another clue and costs a little.

Parameters
Reward is accuracy multiplied by restraint; both rash decisions and endless gathering lose points.
Scored
Clues taken before committing, against correctness.
Why this shape
Borrowed from the jumping-to-conclusions drills in Social Cognition and Interaction Training, whose three-phase structure - emotions, then figuring out situations, then integration - is worth copying even though the adolescent pilot was underpowered. The material maps directly onto an impulsive misread.

Your turn

Moderate as a measureWeak as training

Real dialogue audio. Tap when the speaker is about to finish.

Parameters
A tap more than 200 ms before the true end counts as an interruption.
Scored
The whole distribution of your taps, not just the hits - offset in milliseconds from the true turn end.
Why this shape
Across languages the typical gap between speakers is about 200 ms, far shorter than the roughly 600 ms it takes to plan a word. People therefore predict turn ends from tone and grammar rather than reacting to them, and voice-activity models show tone alone is enough. That makes prediction trainable in principle and perfectly gradable in practice. No trial has tested it as training, hence the grade.

What changed

Weak

Two versions of a social scene, or a three-second clip. Who moved, who stopped talking, what changed.

Parameters
A ten-second cap. Social changes and object changes are tracked separately.
Scored
Accuracy by change type, and time to first look at a social element.
Why this shape
It builds on change-blindness research, where people preferentially notice social interactions. No trial has used it as training, so it is honestly graded as a plausible attention drill rather than a treatment.

Tone and face drills

Strong nearWeak far

The same sentence in different tones: pick the emotion, then pick the acoustic cue that gave it away. Separately, faces and voices at graded intensity across six to eight emotions.

Parameters
Adaptive intensity, ten minutes, three times a week, four to six weeks. Tone drills start with strong cues and fade them.
Scored
Accuracy by emotion and by intensity level.
Why this shape
Two meta-analyses find medium pooled effects on emotion recognition, and the pattern repeats everywhere: gains appear on close generalisation - new faces, same task - and not on distant generalisation to real social behaviour. These are warm-ups, labelled as warm-ups.
Next: the Connect exercises →

Version 133 · Privacy · Terms of use · Appearance