That was not the plan. We paid 349 strangers to write in the air — every letter, digit and a few shapes, three times each, phone held like a remote — to teach the wand to read air-writing. 52,191 letters later, we had age, sex and handedness attached to each one, so we looked.
The recogniser, tested the honest way
Leave one writer out, train on the rest, see how the model does on someone it has never met. Done that way, the wand reads left-handers as well as right-handers (68.5% vs 68.2% of letters, cold start), women as well as men (68.0% vs 67.4%), and the over-60s as well as anyone. Whole words come out exactly right 85% of the time for a first-time stranger, and 93% after a short personalisation pass — twelve writers, twelve improved.
The one group it struggled with was the 18-to-24s: 55% of letters, against about 70% for everyone else. Same on a second run, with a later model. So we went looking at what they did.
One in five of them wrote micro-letters. Twenty centimetres of hand travel per letter, or less — postage-stamp size, thumb-typing scale — where everyone else draws a letter the size of a dinner plate. The model was taught on dinner plates. It reads a postage stamp about 8% of the time. Take the micro-writers out and the rest of the under-25s still sit a bit lower: smaller letters, more hesitations, a little less smooth, and they followed the on-screen arrows less closely than their elders.
I’m not going to claim a generation has forgotten how to write; twenty-three people is a lead, not a verdict. But it does look as though the hand that grew up on a glass screen makes small, quick, thumb-scale marks — and a machine trained on the rest of us has to learn that too.
Men and women write the same in the air
Nine ways of measuring a stroke — time, size, speed, slant, roundness, smoothness, how quickly the hand sets off, which way the circles go, how much the hand wobbles at rest — nine ties. Give a model all of it and it guesses a writer’s sex right 58% of the time. A coin manages 50. The old stereotype — rounder, neater, more careful — did not turn up anywhere we could measure. The one thing that did differ is the pitch of the tiny wobble in a hand holding the phone still: a touch lower in women, the same size in both. The likeliest reason is mechanical — the same phone is a bigger load for a smaller hand.
Older hands don’t shake more. They slow down — after 50.
A letter takes 1.7 seconds from 18 to about 50, then a fifth of a second longer per decade. The wobble of a hand holding still — its size and its pitch — doesn’t change at all from 18 to 78. Neither do the size, slant or roundness of the letters.
Which way do you draw a circle?
85% of right-handers’ circles run anticlockwise — the figure Google’s Quick, Draw! found for the US on a screen. A third of left-handers’ circles run clockwise. Odds ×2.9, the same on iPhone and Android and in every recruitment wave. Letters with a bowl — a, d, g, q, e, 6 — run anticlockwise for almost everyone, left-handed or not: the classroom wins for a letter; the free circle is where the hand shows.
How we did it
Participants were adults recruited and paid on Prolific, who consented to their phone's motion data being recorded and used for research; we report only aggregates. Motion was recorded by the phone's own accelerometer and gyroscope at 60 Hz. The direction, slant and rotation senses were not assumed — they were calibrated from prompts to draw arrows and clockwise/anticlockwise circles, separately per platform, and checked against letters whose direction is fixed by their shape. Hypotheses, glyph sets and tests were written down before the analysis ran; effects are reported with confidence intervals, false-discovery control, and a replication in each platform and in each recruitment wave. Recognition numbers are leave-one-subject-out; personalisation numbers come from twelve writers who returned a day later. Limitations: an online-panel sample; thirteen people over 65 and twenty-three under 25; two phone platforms of many models; hand size unmeasured; the phone's hand not recorded.
What this is not
It is not graphology — reading character from loops has been tested for a century and it does not work, and we make no claim about anyone's personality. It is not a test for anything medical. And it is not a way to tell who you are: a model can spot a left-hander only three times in four, guesses sex little better than a coin, and guesses age no better than guessing the average.
Fancy a go? Your phone can stand in for the wand: phone-demo.the-wand.ai/write
— John