You do it every day. The tool holds the thread, the context, the thing you were looking for — and what that leaves of your focus has never been measured. Not once.
Fifteen minutes to measure it. Five a day to train it back.
Free · no signup needed · works in your browser
Fig. 1 — the shape practice curves take. Your own two numbers come from your baseline and your monthly retest.
The tool holds the files, finds the bug, drafts the paragraph. Hand information to an external store and your own recall of it drops — that is cognitive offloading, replicated long before anyone had an agent in their editor.
666
people asked how much they lean on AI
The more someone used AI tools, the lower they scored on a critical-thinking test — with cognitive offloading as the path between the two. The youngest group leaned hardest and scored lowest.
A cross-sectional correlation: which way the arrow points is unknown. Peer-reviewed, with a published correction.
Gerlich, 202583%
could not quote their own essay
Fifty-four people wrote essays with an LLM, with a search engine, or unaided. The LLM group showed the weakest brain connectivity of the three. Asked to quote one sentence from the essay they had just finished, 83% of them could not.
Preprint, not peer-reviewed, and formally contested — hold it that loosely.
Kosmyna et al., 2025 · MIT Media LabOrder
the finding nobody quotes
Participants who worked unaided first and brought the assistant in afterwards showed higher connectivity, not lower. The sequence changed the outcome. Which is to say: this is not a verdict on the tool, it is a question about how you use it.
Same preprint, fourth session, 18 participants. Thin — but it points the other way.
Stanković et al., 2026 — the published critique of that preprintAnd we put the reps back — five minutes a day on the exact skills the tool now does for you.
1
Fifteen minutes, five standard tasks, one number per skill.
2
Two tasks, five minutes. The difficulty climbs when you do.
3
Once a month, on forms you never trained. The honest one.
No account needed to start · account keeps the streak
Millisecond timing, a fresh random sequence every run, and every parameter published on the methodology page.
Sustained attention
Press for every digit, hold on the 3. Catches the moment focus slips.
Inhibitory control
Name the ink, not the word. Measures what conflict costs you.
Selective attention
Follow the middle arrow while four others disagree.
Working memory
Hold a sequence, then reverse it. Working memory, in digits.
Processing speed
Ninety seconds against a key. Raw throughput, nothing else.
Fifteen minutes, and you have a number for all five.
Quoted as published. An effect size of 0.2 is small, 0.5 medium, 0.8 large.
0.45
how far practice moves the skill you practise
Across the meta-analyses of cognitive training, practising a skill produces a solid, repeatable gain on that skill and on tasks built the same way.
Sala et al., 2019 (second-order meta-analysis)0.87
still there two years later
In the largest training trial ever run — 2 832 adults, ten sessions — speed-of-processing training held an effect size of 0.87 at two years and 0.76 at five.
Ball et al., 2002 (ACTIVE randomised trial)10
sessions were enough to see it
ACTIVE delivered its gains in ten sessions spread over five to six weeks. Cognitive training is measured in weeks of short sessions, not years.
Ball et al., 2002 · Rebok et al., 2014 (ten-year follow-up)Fifteen minutes today, and you have the number to compare everything against.
Free · no signup needed · open methodology