Attention's raw inputs — air, trees, movement — and a 3.2-million-problem warning about offloading to AI
This week the strongest evidence is about what feeds attention rather than what attention does: tree canopy tracks better directed attention in nearly 12,000 ABCD children with fine particulates as a mediator, a Belgian birth cohort ties last-month air pollution to slower saccades, and a 104-child randomized crossover puts a measurable sympathetic mediator between a bout of exercise and steadier inhibitory control. In parallel, single neurons in the human medial temporal lobe show that error monitoring switches off during mind-wandering, and a macaque locus coeruleus study complicates the popular 'LC burst = attention' shorthand. The loudest signal is a preprint: across ten years of ALEKS data, time-on-task for AI-solvable math problems has fallen ~27% since ChatGPT and proctored retention odds are down 25% — large, quasi-experimental, and not yet peer-reviewed.
cumulative decline in the odds of a correct answer on randomly assigned, proctored retention items for AI-solvable math topics after ChatGPT's release (ALEKS, 3.2M learning records; arXiv preprint, not peer-reviewed)
Where the human evidence is
Count by statusEvidence-stage map
Stage × promise (1–10)Attention domain impact
Strength 0–10How a bout of exercise steadies attention
MechanismMedicine & Science in Sports & Exercise, September 2026. In a randomized crossover of 104 children, the pathway from acute exercise to more consistent inhibitory control ran through a measurable drop in sympathetic tone — the same arousal system the locus coeruleus story is about. It converges with this week's environmental findings: green space and clean air were associated with better directed attention, suggesting attention's raw material is partly physiological state.
Environment as an input to attention
Green space tracks better directed attention in 11,868 children — and fine particulates carry part of the effect
Fowler, Luby, Barch and colleagues tested Attention Restoration Theory in the ABCD study (N=11,868, mean age 9.96). Census-tract tree-canopy and park cover was positively associated with performance-based directed attention (β=0.76) and crystallized intelligence (β=0.83), adjusting for area deprivation, ADHD diagnosis, stimulant use, sleep, screen use, perceived safety, prenatal tobacco, physical activity and PM2.5. PM2.5 partially mediated both associations (indirect β=0.06 each). The design is cross-sectional and the standardized effects are small; the contribution is the size and the explicit test of an air-quality pathway rather than a purely 'restorative' one.
Read the paperRecent air-pollution exposure is associated with slower, fewer saccades — an oculomotor marker of cognitive workload — in 9-11-year-olds
In the ENVIRONAGE birth cohort, 196 children completed eye-tracking paradigms indexing cognitive workload and emotion-related attentional bias. After multiple-testing correction, an IQR increase in last-month black carbon or NO2 was associated with roughly 11°/s lower saccade velocity, and last-month NO2 and PM2.5 with fewer saccades; first-trimester and first-1000-day windows also related to saccade velocity. Pupil diameter, fixation count and the attention-bias outcomes did not survive correction. Objective outcomes and modelled residential exposure are strengths; the sample is modest and the oculomotor-to-attention inference is indirect.
Read the paperBody to brain — arousal, movement and the attentional filter
A bout of exercise steadies children's inhibitory control via lower sympathetic tone — with a measured mediator
Hillman's group randomized 104 children through acute exercise, the Trier Social Stress Test for Children (active control) and seated rest, measuring salivary alpha-amylase (sAA; a sympathetic/LC-NE proxy), cortisol, P3 amplitude/latency, attentional inhibition and academic achievement. Exercise lowered sAA relative to both controls and lowered cortisol relative to the stressor; the post-exercise fall in sAA statistically mediated reduced trial-to-trial reaction-time variability on incongruent flanker trials, i.e. more consistent inhibitory control. Exercise effects on P3 amplitude depended on baseline cortisol. sAA is an indirect LC-NE marker, and mediation in a crossover is still correlational at the mediator step.
Read the paperSpatial attention in a moving head: alpha lateralization and micro head-turns are coupled, and the turns improve accuracy
Wöstmann and Obleser recorded EEG alpha and head motion while 33 listeners attended to lateralized speech. Alpha power fell contralateral to targets and rose contralateral to distractors, the familiar filter signature; gyroscopes showed consistent head turns toward targets and more variable movement to distractors, with miniature turns persisting even when told not to move. Head rotation improved accuracy by reducing target uncertainty, and trial-by-trial alpha modulation preceded lateral eye and head movements. Single lab, modest n, but a useful reminder that 'covert' attention in the lab is partly overt sampling in life.
Read the paperClinical — attentional bias and the validity of attention ratings
Tablet-based attention-bias modification shifts a neural marker in major depression, but the clinical signal is unclear
Kratochwil, Frodl and colleagues randomized MDD outpatients and healthy controls to ten tablet-based dot-probe sessions over 14 days of positivity training versus control training. The primary outcome was EEG Early Posterior Negativity to positive stimuli. In MDD there was a time × training interaction (η²p=0.049; between-treatment difference 1.65 μV, 95% CI 0.19-3.11). Depression symptoms fell in both arms without a reported arm difference, so the trial shows a small neural target engagement, not efficacy. The authors themselves call for higher stimulus contrast and longer EEG follow-up.
Read the paperMultilingual children are rated as more inattentive by parents yet are not diagnosed more — a validity problem in a 70,102-child cohort
Köder and colleagues linked parental CBCL/disruptive-behaviour ratings of attention difficulties and multilingual language use for 70,102 MoBa children to registry ADHD diagnoses. Multilingual children received higher attention-difficulty ratings than monolinguals, but ADHD prevalence did not differ by multilingualism or parental language background; there were indications of under-diagnosis among children whose mothers came from North Africa, the Middle East, and South/East Asia, and daughters of non-Norwegian-speaking mothers were diagnosed significantly later. The implication is measurement: parent questionnaires may not travel across linguistic and cultural contexts, and diagnostic pathways are uneven. Observational, registry-based, Norway-specific.
Read the paperAttention in the AI era — offloading at population scale
After ChatGPT, students spend ~27% less time on AI-solvable math problems and retain less of them — 3.2 million ALEKS records
Rismanchian and McGraw Hill colleagues exploit within-curriculum variation between AI-susceptible (text word problems) and AI-resistant (interactive graph) topics across a ten-year ALEKS panel. Post-ChatGPT, college students' learning time on susceptible problems fell 2.8% per quarter, cumulating to 26.9% over eleven quarters (high school 31.3%, middle school 9.0%, Grade 5 no change); the divergence vanished under proctoring, arguing against genuine efficiency gains. On randomly assigned proctored retention items, odds of a correct answer fell 25% cumulatively, while unproctored items moved the opposite way. This is the largest behavioural evidence yet for cognitive offloading eroding durable learning, but it is a preprint with a difference-in-differences design, platform-employed authors and no direct measure of AI use.
Read the paperAmong 5,748 students, a 'problem-dependent' generative-AI profile — escapism, low AI literacy, high dependency — carries the most depression and anxiety
Shi and colleagues identified four GenAI-use profiles: Rational-Tool, Moderate-Recreational, Problem-Dependent (15%) and Light-Exploratory. The Problem-Dependent profile showed the highest BDI-II and BAI scores; the Rational-Tool profile (high AI literacy, instrumental motivation) the lowest. In an exploratory random-forest model, depression, smartphone addiction and executive-function difficulties were the strongest correlates of problem-dependent use, and conscientiousness of rational use. Cross-sectional, self-report, single country: it cannot separate cause from consequence, but it converges with earlier findings that how AI is used matters more than how much.
Read the paperTreat air quality as an attention variable, not only a respiratory one: in ABCD (N=11,868) part of the green-space benefit for directed attention ran through PM2.5, and in ENVIRONAGE (n=196) last-month black carbon and NO2 tracked slower saccades. For families and schools, filtration and siting decisions have a cognitive rationale.
A single exercise bout before demanding work is now backed by a mediator, not just an outcome: the sympathetic-tone drop (sAA) after exercise explained steadier inhibitory control in a 104-child randomized crossover. Expect consistency gains more than speed gains, and note the effect on neural markers varied with baseline cortisol.
Mind-wandering costs more than the missed target — human MTL neurons stop signalling errors during it (Hippocampus 2026). In high-stakes checking work, an attentional lapse also removes the safety net, so build external verification rather than relying on self-correction.
Where retention matters, do not let the model do the first pass: across 3.2M ALEKS interactions, students shortened time-on-task on AI-solvable problems by ~27% and proctored retention odds fell 25%. This is a preprint from platform-affiliated authors, so cite it as a large warning signal, not settled evidence.
Do not read 'phasic LC activity' as a synonym for attention or decision in a paper or a pitch: in two macaques most LC neurons fired with the saccade and were independent of accuracy. Pupil-based 'focus' claims that lean on that shorthand deserve a second look.