One in Twenty UK Adults Pays for AI. The Jobs Map Is Uneven.
Warwick and LSE tracked fourteen million accounts. IT at about twenty percent; social care near two. Oxford: curricula help people unlearn a sticky rule.
Directing Intelligence Digest | November 6, 2026
Today’s video covers who personally pays for AI in the UK, and why simple-to-complex curricula help people escape a sticky rule. Watch it asking: who already buys the tool, and who still gets sequenced practice?
The short version
Who pays: about 5.1% of UK adults paid for an AI assistant in June 2026, up from 2.3% a year earlier (~14M accounts). (Who Pays for AI)
By job: IT 20.3%; lecturers 16.0%; finance 15.8%; school teachers 7.5%; nurses 2.5%; social care 1.7%. (Who Pays for AI)
Wage prices: schooling return fell 0.55 log points in AI-exposed US occupations after ChatGPT; early-career experience returns rose. (arXiv)
Dull-work claim: after adjustment, the LLM link from AI exposure to “dull/dirty/dangerous” scores largely fades. (arXiv)
Curriculum unlearning: simple-to-complex lifted rule-plus-exception mastery to 65.3% vs 44.4% (N=207). (arXiv)
Scaffolded essays: AI use did not independently predict ethics-essay quality once prior writing was controlled (Auckland, N=1,894 surveys). (arXiv)
WORK: Who writes the check for AI
The research. Nikhil Datta, Peter John Lambert and Patrick Schneider (Warwick, LSE, Imperial) match UK open-banking and card records covering about 14.2 million accounts to more than 200 AI products, impute app-store AI subscriptions, and re-weight to UK adults. Occupations come from professional-body and employer payment signals. whopaysfor.ai, November 5, 2026.
What it finds. About 5.1% of adults paid in June 2026 (about 2.8 million), up from 2.3% a year earlier. Claude’s share of direct spend rose from 7.4% to 29.4% in seven months. Ages 20–29 are 5.9× as likely as 60+ to have ever paid. By job: IT 20.3%, lecturers 16.0%, finance 15.8%, teachers 7.5%, nurses 2.5%, social care 1.7%.
How it fits. The wage paper below finds exposed US occupations saw schooling returns fall and early experience returns rise. Who Pays shows consumer demand already concentrates among younger knowledge workers.
Limits. Consumer payments only; employer seats invisible; app-store share estimated; UK only.
What to watch: whether seat data ever closes the gap between who pays personally and who is assigned a tool.
Read it: site | PDF
On X: Peter John Lambert
Who’s Directing? I’d put two columns on the AI dashboard: who already pays out of pocket, and who gets protected practice time. If those don’t overlap, the strategy is theater.
WORK: Schooling got cheaper. Early practice got dearer.
The research. Levy Yeyati, Judzik and Barbiero (UTDT / IDB) compare US CPS earnings, 2018 through June 2026, in top vs bottom quartiles of pre-ChatGPT AIOE. arXiv 2610.36008.
What it finds. After ChatGPT, the return to a year of schooling fell 0.55 log points in exposed occupations. Experience returns rose 0.26 log points at three years, were flat at fifteen, and fell 0.12 at twenty-five (crossover 18.1 years under their quadratic). Codified knowledge got cheaper; practice-based judgment did not move the same way.
Limits. Potential experience is a proxy; exposure is occupational, not observed use.
What to watch: whether junior roles still include the practice the wage premium now prices.
Read it: arXiv
Who’s Directing? If I were redesigning junior roles, I’d protect the first three years of practice the market is starting to pay for.
WORK: “Dull, dirty, dangerous” is a weak AI map
The research. Siyi Yu (Purdue) links 21 O*NET items across 754 US occupations to LLM exposure and AIOE. arXiv 2611.06888.
What it finds. After adjusting for physical demands, preparation and wages, the fully adjusted LLM interval includes zero. Within Dull, restricted discretion tracks exposure more than repetition. About 31.6% of employment sits in high-condition, low-exposure jobs.
Limits. Observational; no worker preference data.
What to watch: business cases that sell “AI takes the dull work” without naming the condition.
Read it: arXiv
Who’s Directing? I’d ban the DDD slogan until someone names the condition and asks the workers.
LEARNING: Curricula that help you unlearn
The research. Rubruck, Zhang, Saxe and Summerfield (Oxford; Saxe also UCL) test why simple-to-complex curricula work on a rule-plus-exception task, in humans and yoked nets. arXiv 2611.04565.
What it finds. Exp 1 (N=207): people latch onto a one-feature rule and underfit exceptions. At test, rule-plus-exception mastery hit 65.3% under simple-to-complex vs 44.4% with no curriculum (OR 2.35). Complex-to-simple hurt. Exp 2 (N=1,027): curricula help most after compressed one-dimensional priors. Their phrase: curriculum unlearning.
How it fits. Yesterday’s practice-dose pattern and today’s wage paper both point at practice. Sequence decides whether people escape a sticky rule.
Limits. Lab task; short training; not a classroom trial.
What to watch: tutoring and onboarding that graduate exceptions after the rule, then retest the sticky answer.
Read it: arXiv
On X: Summerfield Lab
Who’s Directing? If I were building an AI literacy path, I’d teach the rule, stage the exceptions, then test whether people can still drop the sticky answer.
LEARNING: Scaffolded AI use didn’t buy better ethics essays
The research. Shakil, Denny and colleagues (University of Auckland et al.) track GenAI on ethical essays across three years of a first-year computer systems course (1,894 survey responses). arXiv 2611.08431.
What it finds. Voluntary AI use held near two-thirds; a 2026 draft-before-AI scaffold pushed adoption above ninety percent. AI use did not independently predict essay quality once prior writing was controlled. Markers coded critical engagement in about 77–81% of scaffolded submissions.
Limits. One course; self-report; no unscaffolded 2026 control.
What to watch: draft-first checklists with marked reflection, not detectors.
Read it: arXiv
Who’s Directing? I’d copy the draft-first rule before I copy any AI detector.
Quick hits from X
METR on Anthropic R&D acceleration: AECI growth about 1.9× prior [1.5–2.8], heavy redactions. METR
Alex Imas posted Organizing Intelligence (Koh, Legg): competence reshapes wages; coordination can break firm-size limits. Imas · arXiv
Worth a listen
Nothing cleared the bar in the last thirty hours.
Triangulation: signal and noise
Signal. Who Pays, the wage paper and Oxford’s curriculum study agree: codified knowledge got cheaper; practice and sequenced judgment did not.
Noise. Employer seats are invisible; the curriculum paper is a lab task; wages price occupations, not users.
My read. Measure who already pays, protect early practice, and sequence exceptions after the rule.
Until tomorrow, that’s the intelligence. How it’s directed is up to you.
Todd