How it works

A field experiment with about 1,000 high-school students found that unrestricted GPT-4 access raised practice scores 48% and lowered unassisted exam scores 17%, while a tutor version that gave hints instead of answers kept the practice gains and removed the exam loss (Bastani et al. 2025, PNAS). In a randomized study of 52 developers learning a new library, the AI-assisted group scored 17 points lower on a follow-up quiz, with the largest gap on debugging; what predicted learning was interaction style, not AI use as such (Anthropic 2026). Explanation without a recommendation produced both good decisions and learning; a recommendation with an explanation produced good decisions and no learning (Gajos and Mamykina 2022). Built in: the answer-versus-hint split, and the scaffold mode.

Experienced open-source developers were 19% slower with AI tools and believed they had been 20% faster (METR 2025, 16 developers, 246 tasks). Confidence in the AI predicted less critical thinking across 936 real tasks from 319 knowledge workers (Lee et al. 2025, CHI). Built in: the ledger is behavioral. Felt effort and self-rated competence are never used as signals.

Endoscopists’ adenoma detection rate fell from 28.4% to 22.4% in colonoscopies performed without AI after a few months of AI-assisted practice (Budzyń et al. 2025, Lancet Gastroenterology and Hepatology, observational, 1,443 procedures). Students’ prompting profiles diverged within four assignments of a single semester (Misiejuk et al. 2026). Built in: drift detection on the delegated share over four trailing weeks.

Across 53 studies, skill loss grew with the retention interval and was larger for abstract, feedback-free cognitive procedures than for natural or physical tasks (Arthur et al. 1998). Spacing beat massing in 259 of 271 comparisons (Cepeda et al. 2006). Generating an answer beats reading it, d = 0.40 across 86 studies (Bertsch et al. 2007); trying a problem before instruction helps conceptual knowledge and not rote procedure (Sinha and Kapur 2021). Built in: each domain is a card under the open-source FSRS power law, with priors by decay class and attempt-first aimed at conceptual work. Labelled a guess: the priors themselves, which were fitted on flashcards, not skills.

Lifetime GPS use correlated with worse hippocampus-dependent spatial memory, and heavier use over three years predicted steeper decline in a small longitudinal sample (Dahmani and Bohbot 2020). Navigation deficits precede memory deficits in preclinical Alzheimer’s disease, and in Sea Hero Quest genetic-risk carriers separated on wayfinding but not on memory tests (Coughlan et al. 2019, PNAS). Calculators, by contrast, came out neutral or positive in two meta-analyses, and the Google effect failed replication in 2018. Built in: a navigation domain from day one. Left out: any claim about search engines or handwriting.

The three-process model predicts alertness from time awake and clock time with published constants and was validated on 136 airline crew and 8,040 sleepiness ratings (Åkerstedt and Folkard 1997; Ingre et al. 2014). About six hours of demanding work raised lateral-prefrontal glutamate and shifted choices toward low effort in one study (Wiehler et al. 2022). The daily willpower tank did not survive replication: d = 0.04 across 23 labs and d = 0.06 across 36 (Hagger et al. 2016; Vohs et al. 2021), and dopamine does not drain when you avoid stimulation. Built in: the three-process curve with a chronotype shift. Labelled a hypothesis: the fatigue term. Left out: any dopamine variable.

Cognitive forcing functions cut overreliance on AI more than explanations did, and “people assigned the least favorable subjective ratings to the designs that reduced the overreliance the most” (Buçinca, Malaya, Gajos 2021). Explanations reduce overreliance only when they make checking cheaper than deferring (Vasconcelos et al. 2023). Built in: the verdict before the send, a verification checklist in co-pilot mode, user-set intensity, and a quiet switch.

A breathing pause before an app opened led people to close it in 36% of attempts and cut openings 57% over six weeks (Grüning et al. 2023, PNAS). Reviews of digital self-control tools find short-term reductions and no lasting habit change from blocking (Roffarello and De Russis 2023). Switching between videos to escape boredom makes people more bored (Tam and Inzlicht 2024). Office workers took about 25 minutes to return to an interrupted task (Mark et al. 2005). The media-multitasking literature is small and contested. Built in: the pause, a self-set budget shown rather than enforced, switches per hour reported as behavior. Left out: any claim that screens damage attention.

The 2024 Lancet Commission names 14 modifiable factors accounting for up to 45% of dementia, as a theoretical ceiling (Livingston et al. 2024). The best multidomain trials show differences of 0.02 to 0.03 standard deviations a year (FINGER; US POINTER 2025). Young-onset dementia runs about 1 per 100,000 at ages 30 to 34, so a detector for young adults has no predictive value (Hendriks et al. 2021). Brain-training games improve the trained task and little else across 11,430 people (Owen et al. 2010). Built in: the vocabulary of trends and habits. Left out: risk scores, detection, any disease claim, and games.

Knowing a task activates a region says little about what exercising the region does, because most regions serve many processes (Poldrack 2006). The lateral prefrontal and parietal “multiple-demand” network lights up for nearly any hard task (Duncan 2010). Open resources exist for the mapping itself: the Cognitive Atlas, Neurosynth, and O*NET’s abilities taxonomy. Built in: a one-line group-average note under each domain. Left out: any meter that claims to measure a region.