Directing Intelligence: DI Digest

One AI Agent for the Whole Team Beat One Agent Per Person

Todd McLees — Thu, 12 Nov 2026 15:52:19 GMT

Directing Intelligence Digest | November 12, 2026

Picture sixteen researchers sharing one computing budget that can’t cover everyone’s work, each with a personal AI agent told to get their jobs done. Most of those agents never did anything. In a new Anthropic study published yesterday, as teams grew from four agents to sixteen, the share of Claude Sonnet 5 agents that ever acted fell from 66% to 10%.

Today’s video moves from a crowded control room of idle assistants into a lecture hall where the exam rules are changing. Watch it asking: in your organization, who is deciding what the AI does, and who is checking what people really know?

The short version

One coordinator beat a crowd of personal agents: across four simulated workplaces, one agent serving everyone beat one agent per person, by about 10 to 45 percentage points for both Claude models tested. (Anthropic, Nov 11)
AI recommends, people decide, jobs follow: in a double-blind Swiss trial, refugee placements recommended by an algorithm, with officers keeping the final say, raised the employment rate about 11% at three years. (arXiv 2609.35448)
A regulator moves on assessment: England’s Office for Students will require assessments that reliably measure what students know, often AI-free, and courses that build AI skills, with fines available for institutions that fall short. (Office for Students, Nov 11)
Practice with AI can still help: in a four-week trial in English secondary schools, exam review with an AI science tutor beat students’ usual review by about 2.5 marks out of 35. Funded by the tutoring company. (arXiv 2609.14789)

Subscribe now

Here’s AI at work: a team of personal agents did worse than one coordinator

The research. Sahan Paliskara and Nattaput Namchittai (Anthropic Fellows Program), Jacqueline Tay (independent), and Avery Griffin and Andrew Lampinen (Anthropic) built four everyday settings where several people’s agents draw on one shared resource: a computing budget split among researchers, a clinic calendar shared by urgent patients, a group food order or trip booking, and a queue of code changes racing a release deadline. They compared one agent serving everybody with one agent per person, and scored the outcome for the group as a whole. Anthropic, “Worse Together”.

What it finds. The single coordinator won in all four settings for both Claude Opus 5 and Claude Sonnet 5. The main failure was doing too little, too late: as teams grew, more agents sat out entirely. In the group food order, a request for gluten-free sides from a user with celiac disease reached the agent placing the order in 94% to 99% of runs, but was still in that agent’s view at checkout in only 47% to 67%. What helped: on the budget, a designated team lead with stated priorities closed the gap. On the order, a rule blocking checkout until every message was read rescued 73% of failed Opus 5 runs.

How it fits. A field trial shows the structure that works. Kirk Bansak, Jens Hainmueller, Dominik Hangartner and colleagues (Stanford, ETH Zurich, UC Berkeley and others) ran a double-blind trial with Switzerland’s State Secretariat for Migration from 2020 to 2023, meaning neither officers nor refugees knew which kind of recommendation a case got. For about 2,000 refugee cases, officers received either a placement recommended by an algorithm or one drawn to mimic the usual process. Officers kept final authority and followed the recommendation about 97% of the time. At 36 months, the share of cases with an adult working was 5.2 points higher, about 11%, which the authors compare to the gains from hundreds of hours of intensive language training.

Limits. “Worse Together” is simulated users in simulated tasks, written by the company that makes the models it rates best. The Swiss trial is one country and one program, and its three-year average result only barely clears zero once statistical uncertainty is counted.

What to watch: whether workplace tools that give each employee a personal agent ship a coordinator role or shared priorities before they ship to whole teams.

Read it: Anthropic | Swiss refugee-matching trial | On X: The SI Upside on the Swiss trial

Who’s Directing? I run my own team of AI agents, and this paper describes the thing I’d plan for first: someone has to coordinate. If I managed an operations team about to give everyone a personal agent, I’d name one person or one agent as the lead and write the team’s priorities down before launch, because these agents didn’t sort that out among themselves.

Now, here’s what’s happening in learning: AI-free proof, AI-rich practice

The research. England’s higher education regulator, the Office for Students, and Advance HE, a sector body for teaching and leadership, published “Making AI work for students” on November 11. Through roundtables, online responses and interviews from June to September, they heard from 244 staff and 91 students and student representatives from at least 127 institutions. Office for Students.

What it finds. Students described peers finishing assessments with AI “sometimes in just a few minutes” and getting a similar grade. Detection tools don’t reliably identify AI use. In response, the regulator will require valid, reliable assessment, which in many cases means secure, AI-free formats such as closed-book exams, presentations and practicals. It will also require courses to build relevant skills, and it can impose conditions, fines or funding suspensions.

How it fits. AI-free assessment doesn’t mean AI-free learning. In a trial led by What Worked Education and Durham University, 929 students in English secondary schools were randomly assigned to four weeks of GCSE science review (GCSEs are the national exams English students take around age 16) with the Medly AI tutor or their usual review. Among the 644 who finished, the tutor group scored 13.95 out of 35 against 11.39. Medly funded the evaluation, attrition was 30.7%, and the tests were curriculum-aligned rather than standardized.

Limits. The regulator’s report is qualitative and England-only, with no estimate of how common AI-completed assessment is. The tutoring trial is a vendor-funded preprint with short follow-up.

What to watch: U.S. accreditors and state systems, and whether any follow England in tying AI-free assessment to compliance.

Read it: Office for Students report (PDF) | AI tutoring trial | On X: Ashley Booker on the tutoring trial

Who’s Directing? If I were a community college dean, I’d sort every course’s major assessments into two piles this term: the ones that prove what a student can do alone, and the ones built for practice. Then I’d make sure the first pile is secure and the second pile actually uses AI well, because employers read the credential as proof.

Quick hits

Few companies define AI fluency by role. In an FT Longitude survey of 300 U.K. and U.S. business leaders commissioned by Andela, 60% said they have an organization-wide definition of AI fluency, but only 42% set specific expectations by role or function. Vendor-commissioned survey. Andela, Nov 9

Triangulation: signal and noise

Signal. A lab simulation, a field trial in public administration and a national regulator’s review point the same way: AI does its best work inside a structure a person sets. Someone coordinates the agents, someone keeps the final say, and someone checks what a student can do without the tool.

Noise. The Anthropic study is simulated and self-authored; the tutoring trial is vendor-funded; the regulator’s evidence is qualitative.

What I’d do. Before rolling out personal agents, name the coordinator and write down the priorities. Before redesigning a course for AI, decide which assessment proves the learning and protect it.

Until tomorrow, that’s the intelligence. How it’s directed is up to you.

Todd

Sources

Subscribe now

AI Built Into the Store Lifted Tiny Sellers’ Sales 61%

Todd McLees — Wed, 11 Nov 2026 21:42:28 GMT

Directing Intelligence Digest | November 11, 2026

For the smallest sellers on a leading Chinese online marketplace, paying market rates for product photos, videos and listing copy would have cost the median firm in the bottom size quarter 22.7% of a year’s sales. In the top quarter, 0.4%. Then the platform built generative AI into the software every seller already uses, and the smallest shops grew fastest. That’s from a new NBER working paper from the University of Wisconsin-Madison and two Zhejiang universities, passed around on X this week by Raffaella Sadun and Kevin Bryan.

Today’s video walks through a night market of tiny stalls, then into a classroom where the lights dim. Watch it asking: in your organization, is AI something each person has to figure out alone, or is it built into the tools your newest people already use?

The short version

Built in, the smallest gain most: small sellers who adopted the platform’s built-in AI saw monthly sales rise 51.6%; the bottom quarter by size grew 61.5%, the top quarter 10.3%. (NBER w36003)
Left to individuals, seniors pull ahead: in a PA Consulting survey of 2,000 U.K. employees, 70% of senior leaders say AI helps them do things they couldn’t before, versus 38% of entry-level and junior staff. (PA Consulting, Nov 9)
A teacher chatbot backfired: in a semester-long trial with 193 teachers in Türkiye, a GPT-4o teaching assistant lowered students’ motivation and confidence, and scores fell for students of lower-performing teachers. (Stanford SCALE review, Nov 10)
The guidance gap: 79% of surveyed students use AI for homework or assignments; 47% say they were shown how. (OpenAI, citing ThinkYoung, Nov 10)

Here’s AI at work: the smallest shops gained the most

The research. Panle Jia Barwick (University of Wisconsin-Madison), Lu Fang (Zhejiang University of Finance and Economics), and Jinhan Liu, Huilan Xu and Zhe Yuan (Zhejiang University) studied a leading Chinese e-commerce platform (anonymized) that in 2024 embedded generative AI in the software sellers use to run their stores: tools that generate product images, videos and titles, plus ready-made analysis of demand and competitors. With complete records of each firm’s AI use and monthly sales, they compared adopters with matched non-adopters among firms selling RMB 5 million a year or less. NBER Working Paper 36003.

What it finds. Adoption raised monthly sales 51.6%, mostly through 47% more orders, with prices barely moving. The gains tilt hard toward the smallest: 61.5% for the bottom quarter by size, 10.3% for the top. The smallest sellers’ negative reviews fell 42.5%, while human customer-service time per order fell 5.6% across adopters. Categories where the tools mattered most drew more new sellers, which the authors read as market expansion. Their line: “Scale economies in producing AI need not become scale advantages in using it.”

How it fits. Self-serve adoption looks different. PA Consulting’s survey of 2,000 U.K. employees, published November 9, found senior leaders using AI more than three times as often as entry-level and junior staff, and 70% of them saying it helps them do what they couldn’t before, versus 38% of juniors. A third of the workforce, 32%, hasn’t seen meaningful benefits. PA’s Kate Spencer: “The employees who already possess greater expertise and judgement are currently deriving the greatest benefits.”

Limits. A working paper, not peer reviewed. One platform in China, physical-goods sellers only. Sellers chose whether to adopt; the authors confirm the results with instrumental-variable estimates and other designs, but it isn’t a randomized trial. PA’s is a self-reported survey from a consulting firm.

What to watch: whether U.S. platforms that embed AI for small businesses publish seller-level results alongside adoption counts.

Read it: NBER | PA Consulting | On X: Raffaella Sadun, Kevin Bryan

Who’s Directing? I’ve built 35 of my own AI agents, and this paper names the thing I’d watch most closely: who carries the setup cost. If I ran a company or a campus, I’d stop counting licenses and list the three systems our newest people open every day, then ask what AI built into those would take off their plate.

Now, here’s what’s happening in learning: the teacher chatbot that lowered motivation

The research. Stanford’s SCALE Initiative (Seth Walker, Lily Fesler, JP Martinez Claeys, Chris Agnew and Susanna Loeb) updated its review of the strongest causal evidence on generative AI in education on November 10. Of 868 papers added to its repository, 201 were trials or quasi-experiments; 32 met its bar, bringing the evidence base to 52. Most new student studies are from colleges, not K-12. Stanford SCALE.

What it finds. The headline: AI helps more when students use it to understand material than to produce work, though neither reliably improves learning. The study that should stop school leaders is a randomized trial in Türkiye by Sungu, Lira and Duckworth (an SSRN preprint, as summarized in the review): 193 teachers and 2,816 students in grades 5 to 12. Teachers got a GPT-4o chatbot for lesson planning, assessment, feedback and admin for a semester. They mostly used it to generate materials and accepted them with minimal revision. Students’ motivation and confidence fell, even though they weren’t told their teachers were using AI. Average scores didn’t change; for students of below-median teachers, they fell. On the student side, a Brazilian trial found ChatGPT cut presentation prep from 5.8 to 3.2 hours, and students who prepared the traditional way did better on a surprise conceptual test 45 days later.

How it fits. OpenAI, citing new ThinkYoung research published November 10, reports that 79% of surveyed students use AI for homework or assignments, but only 47% were shown how. It adds that as AI improves, “the role of teachers becomes more important, not less.” The Stanford review says the same thing in evidence: effects depend on how AI is designed and used.

Limits. Many studies are preprints with small samples, and the review finds no evidence yet on whether gains last. The Türkiye result is one trial. ThinkYoung surveyed teens across Europe, the Middle East and Africa, and OpenAI is reporting it.

What to watch: district and campus AI contracts that require a test of what students can do without the tool.

Read it: Stanford SCALE | OpenAI | On X: David Ross

Who’s Directing? If I were a provost or superintendent, I’d add one line to every AI pilot: measure what students can do after the tool is gone. And I’d build teacher training around editing AI materials, because in the Türkiye trial teachers mostly accepted the drafts as written, and their students’ motivation fell.

Quick hits

Freelance demand is shifting. OpenAI and Upwork found that in 43 of 63 work categories, activities common in ChatGPT lost share of Upwork job posts. Company research on its own platforms. OpenAI, Nov 10
Saved time, not reinvested. Payscale’s AI Workforce Impact Report: 43% of employees say AI saves them three or more hours a week; 19% say the time goes into more strategic work. Vendor survey. Payscale, Nov 10

Triangulation: signal and noise

Signal. An economics paper on a marketplace, a U.K. workforce survey and a review of 52 education trials point the same way: gains go to people who get AI built into their work and keep their judgment in it.

Noise. The marketplace paper is one platform in China; PA, Payscale and OpenAI are reporting their own surveys; the Türkiye trial is a single preprint.

What I’d do. Build AI into the tools your least-resourced people already use, and keep a person editing what it produces. Then ask two questions: who on your team is figuring AI out alone, and what can your students do without it?

Until tomorrow, that’s the intelligence. How it’s directed is up to you.

Todd

Sources

AI Coding Tools More Than Tripled Commits. Releases Rose 30 Percent.

Todd McLees — Wed, 11 Nov 2026 08:45:21 GMT

Directing Intelligence Digest | November 10, 2026

A working paper by economists at MIT and Wharton, first posted in May, revised in September, has been circulating on X. The authors followed more than 500,000 developers on GitHub and looked at four of the biggest software stores: Apple’s App Store, Google Play, the Chrome Web Store and SourceForge. Since early 2025, monthly new apps on those stores have roughly doubled, on average. Total engagement with those new apps after launch has not risen in any of them.

The short version

More code, fewer finished things than you’d think: across 500,000+ GitHub developers, AI coding tools, agents included, lifted weekly commits about 240% on a cumulative basis. Projects rose 80% and releases about 30%. (NBER w35275)
Twice the apps, same audience: monthly new apps roughly doubled, on average, across four big stores since early 2025. Total engagement with new apps after launch didn’t rise in any. (NBER w35275)
One company’s answer: Sourcegraph now has a bot route each change to Ship, Show or Ask; anything touching authentication or sensitive data still goes to a person. (Sourcegraph, Nov 10)
How the rule is worded: across AI policies at 101 U.S. universities and 40 interviews, rules framed as dishonesty led some students to hide their use; rules about owning what you submit drew openness. (University of Iowa, arXiv)

WORK: The code got written. The releases didn’t keep pace.

The research. Mert Demirer and Liyuan Yang (MIT) and Leon Musolff (Wharton) linked public GitHub activity for more than 500,000 developers to Microsoft’s internal data on when each began using AI coding tools: autocomplete, sync agents that work alongside the developer (for example Claude Code or Copilot’s agent mode) and async agents that work on their own (for example Codex). Each adopter is matched to a similarly active developer from the same week a year earlier. NBER Working Paper 35275.

What it finds. Weekly commits rose about 30% with autocomplete. Adding sync agents brought the cumulative effect to about 180%, and adding async agents brought it to about 240%. Then the gains shrink at every step toward the customer: that 240% becomes 80% more projects and 30% more releases. After adopting sync agents, developers added 7.0 times as many lines of code and deleted 12.2 times as many, and created 79% more branches. Across GitHub, the share of new repositories with no activity after their first month rose from 60% in 2021 to 69%. The authors read the pattern as consistent with a weak link: review, quality control and release capacity limit how much early work reaches the finish line.

How it fits. Sourcegraph published its own response today. Its engineers built a bot that inspects each change and routes it to Ship (no more review), Show (review after merge) or Ask (review first). Anything touching authentication or how sensitive data is exposed always goes to a person. The company reports fewer requests for a quick stamp and calls it an internal experiment. Atlassian’s Q3 State of AI Impact in Engineering Report, cited in a November 4 product launch, says teams that rely heavily on AI to write code wait 71 minutes for a human review, up from 32.

Limits. A working paper, not peer reviewed. Open-source software and four app stores only, with little enterprise coverage; the authors cite an enterprise study with the same pattern and smaller effects. It relies partly on Microsoft data, and at least one co-author discloses outside relationships.

What to watch: engineering dashboards that report releases and customer use next to commits.

Read it: NBER | Sourcegraph | Atlassian (vendor) | On X: Diego Arias pointer

Who’s Directing? I’ve built 35 of my own AI agents, so I read this one closely. If I ran an engineering or IT team, I’d stop reporting AI output in commits and start reporting what shipped and who signed off on it, then write down which changes always need a person, the way Sourcegraph did.

LEARNING: How the policy is worded can shape whether students tell you

The research. Sabiha Mehreen, Xiangzhe Yuan, Rishab Nithyanand and Naimul Hoque (University of Iowa) collected 694 AI policy and guidance documents from 101 U.S. universities, found nine rules that keep recurring, then interviewed 20 students and 20 instructors about them. arXiv 2611.11072, posted November 8.

What it finds. In those interviews, rules that set a similar expectation landed differently depending on their words. Where passing off AI work was framed as academic dishonesty, some students concealed their use. “I pretend not to use it, I pretend that I generate everything by myself,” one said. Where the rule made students accountable for whatever they submit, they described owning the result: “it falls back on you, because you’re the one submitting it at the end of the day.” The most common rule, check with your instructor first, appeared at all 101 universities. One student won’t ask: “I don’t want to find out that I’m not supposed to be using it, and everyone else is using it.” One instructor called enforcement “just sort of a trust situation.”

On campus. The University of Illinois Chicago now requires every syllabus to state acceptable and prohibited student AI uses and any disclosure expectations, and requires faculty to disclose their own planned AI use in course design, materials, grading support and feedback (UIC policy). In Scotland, the University of Dundee’s policy, updated October 7, says it outright: “Disclosure should not be treated as admission of misconduct,” and students should be ready to “discuss, explain, justify, or defend” their work (Dundee).

Limits. A preprint. Forty interviewees, recruited through outreach, volunteer sign-ups and snowball sampling, so this shows how wording is experienced, not how often students hide use. The policy coding was AI-assisted with a manual check.

What to watch: universities rewriting “academic dishonesty” AI clauses into accountability and disclosure language this spring.

Read it: arXiv

Who’s Directing? If I were a president or provost, I’d ask three students to read our AI policy aloud and tell me whether it makes them want to disclose or hide. Then I’d keep the accountability line, put Dundee’s sentence about disclosure not being a confession right next to it, and make the faculty side as transparent as UIC did.

Quick hits

Surrendering judgment. University of Victoria researchers followed 43 senior software engineering students through a semester team project. Some leaned on AI until they were stuck in code they’d built but couldn’t understand. Participants’ accounts credit relaxed deadlines and a lecture and lab on working with AI with helping some take judgment back. Exploratory, one course. arXiv 2611.12646
Feels easier, works harder. In a Colorado State lab study of 20 novice programmers, AI help raised measured mental workload (reaction times on a side task up 11.2%) while students’ own sense of effort didn’t change. Small sample; the arXiv comments page lists it as accepted to ICSE 2027. arXiv 2611.12019

Triangulation: signal and noise

Signal. Half a million developers, one company’s review redesign and 101 universities’ policies point at the human step at the end, approving, owning, disclosing, as where AI’s gains get kept or lost.

Noise. The developer study is a working paper on open-source code using Microsoft data; Sourcegraph and Atlassian report on themselves; the Iowa findings come from 40 interviews.

My read. Name the person or rule at the end of the process, and make that step easy to use honestly. Who stamps the release in your organization, and would a student tell your policy the truth?

Until tomorrow, that’s the intelligence. How it’s directed is up to you.

Todd

Sources

One in Twenty UK Adults Pays for AI. The Jobs Map Is Uneven.

Todd McLees — Fri, 06 Nov 2026 11:53:46 GMT

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

Sources

AI Keeps the Answers Right While Skills Slip, Two New Papers Warn

Todd McLees — Thu, 05 Nov 2026 15:26:59 GMT

Directing Intelligence Digest | November 5, 2026

Today’s video covers two stories: a KIT and IBM Research model of skill eroding while output looks fine, and a Google Research and NFTE trial of AI built to make students practice. Watch it asking one question: in your organization, who’s still doing the reps?

The short version

Hidden deskilling: a new KIT and IBM Research model shows answers can stay right while skill slips, and standard reliance scores can call that appropriate. (arXiv)
Lock-in: an Oxford model finds a long stretch of cheap AI can leave a population dependent after the advantage fades. (arXiv)
Gartner: 23% of 1,303 functional leaders cut headcount due to AI; 17% added. (Gartner)
MIT: AI lifted worker accuracy 21 points, yet workers gave it about a quarter of its rational weight. (arXiv)
Google and NFTE: push-back AI teammates, +7.1 points on business plans; students who skipped practice saw no gain. (arXiv)
UChicago: tech-free Core classrooms, Claude for every student, and a no-AI writing course. (Chicago Maroon)

WORK: The skill loss you can’t see in the output

The research. Karlsruhe Institute of Technology researchers, with Wei Sun of IBM Research, built a theoretical model: on each task a worker accepts the AI’s answer or solves it alone, and skill grows or decays with practice.

What it finds. Novices get trapped and experts deskill while output accuracy holds at the AI’s level. “Appropriate reliance” scores can rate those acceptances as fine. Paying for effort restores reliance but not learning; paying for correct unaided answers does, and pays novices most.

How it fits. Oxford’s Renaud Lambiotte, posted a day earlier, models lock-in: same average pay can hide users at skill 0.12 next to non-users at 0.92. Clinically, endoscopists who’d used AI detected fewer adenomas once it was withdrawn.

Limits. Illustrative parameters, no field data.

What to watch: employers reporting unaided accuracy.

Read it: KIT and IBM paper | Oxford paper

Who’s Directing? I’d add one number to every AI dashboard I sign off on: what the team gets right with the tool switched off. Then I’d put practice on the calendar like any other work.

WORK: Gartner says freed time leaks away

The research. A Gartner survey of 1,303 senior functional leaders, January to April 2026 (press release; full report for clients).

What it finds. “Twenty-three percent of functional leaders reported a reduction in headcount due to AI, while 17% reported an increase,” says analyst Grace Gstalder. Her line: “AI-generated capacity is a resource, not a return.” Gartner’s steps: name priority outcomes for freed time, build redeployment into performance management, tie gains to financial results.

Limits. Advisory-firm survey; only the press summary is public.

What to watch: whether “redeployment” includes the practice time the KIT model says skills need.

Read it: Gartner release

Who’s Directing? If your AI business case counts freed hours, I’d ask where those hours go. Some of them should buy back practice.

WORK: Workers underuse good AI advice

The research. MIT researchers including Asuman Ozdaglar and Daniel Huttenlocher model how a worker combines her own evidence with an AI’s, then test it on ChatBench data.

What it finds. AI raised accuracy 21 points, yet workers applied about a quarter of the Bayes-rational weight. If a worker ignores how much her evidence overlaps the AI’s, a stronger AI can make her less accurate.

Limits. Lab benchmark; I’m working from the abstract.

What to watch: trust calibration in onboarding.

Read it: arXiv paper

Who’s Directing? Teach people to ask what the AI knows that they don’t. That’s the judgment worth training.

LEARNING: AI built to argue made students practice

The research. Google Research and the Network for Teaching Entrepreneurship randomized 52 middle and high school students at four summer BizCamps. AI teammates pushed back, pitched flawed ideas and stated false claims to test whether students checked. Comparison students did the same tasks on worksheets and could use ordinary AI tools. Google tested its own platform.

What it finds. Final business plans scored 67.6% vs 60.5% (+7.1 points), graded by teachers and judges on NFTE rubrics written before the study. Ten students with access completed zero tasks and averaged 59.3%.

How it fits. It mirrors the KIT model: gains followed practice. UChicago, below, protects practice by design.

Limits. Small, short, vendor-run; 25.5% didn’t enjoy it at all; the dose pattern isn’t randomized.

What to watch: a larger, independent replication.

Read it: arXiv paper

Who’s Directing? If I were piloting AI in a course, I’d judge it by one log: did students actually practice? Usage minutes wouldn’t count.

LEARNING: UChicago gives everyone Claude and protects the reps

The story. The Chicago Maroon reports that the University of Chicago has banned technology from humanities and social sciences Core classrooms, even as every student now has access to Anthropic’s Claude suite. A summer memo called the social sciences Core “best understood as a pedagogical setting without AI.” The new ICA writing course, taken by about 70% of first-years and required from 2027-28, bars generative AI from producing, revising or evaluating coursework.

What changed for people. Instructors moved to in-class exams, paper notes and discussion before writing. “The hum Core and sosc Core are training the muscle of your mind,” says senior instructional professor Samantha Fenno. One student called all-in-class essays “going a bit too far.”

Limits. Student-newspaper reporting; no outcome data yet.

What to watch: whether ICA reports writing gains.

Read it: Chicago Maroon, Nov 2

Who’s Directing? The part I’d copy is the pairing: give everyone the tool, then protect the rooms where the skill gets built.

Quick hits from X

Addy Osmani (essay, not research) maps a slide from offloading to “cognitive surrender” and suggests: “Once a week, write a bit of code without an agent.” It lands on the same prescription as the KIT paper, from a practitioner. On X: Osmani’s article

Worth a listen

Nothing cleared the bar today. This week’s podcasts I checked were product talk, not research or deployment evidence.

Triangulation: signal and noise

Signal. Two independent models, a related October paper and a clinical study point the same way: reliance can hide skill loss. Google’s trial adds a dose pattern, and UChicago is already designing for it.

Noise. None of the models are calibrated against workers yet, and the strongest learning result comes from the vendor.

My read. Measure unaided skill and budget practice now; it’s cheap insurance.

Until tomorrow, that’s the intelligence. How it’s directed is up to you.

Todd

Sources

AI Skill-Matching’s Leader Fell to 16th Under a Stricter Test

Todd McLees — Thu, 05 Nov 2026 03:20:25 GMT

Directing Intelligence Digest | November 4, 2026

Proton Will Re-Interview Staff on AI. Bring Proof.

Todd McLees — Fri, 30 Oct 2026 11:40:18 GMT

Directing Intelligence Digest | October 29, 2026

Today’s video covers two stories. A privacy company is re-interviewing staff about how they actually use AI, and learning scientists mapped where AI helps students and where it gets in the way. Watch it asking what you could show for your own AI use this year.

The short version

Proton will re-interview staff on how they use AI, and the results feed a review that could change roles.
Statistics Canada finds youth hiring fell across AI-exposed and less-exposed jobs alike, and the slide started before ChatGPT.
Stanford’s TaskNet had workers describe their own jobs, and it found 79 tasks the government’s job database was missing.
A couple running a cave-house rental in Spain gave an AI agent their bookings, their plant care and, eventually, their press outreach.
A University of Minnesota panel of 30 experts says students should learn first, practice with AI, then think for themselves.
Soongsil University in South Korea set three levels of AI use, and the University of Iowa added Claude for Education to its toolkit.

WORK

Proton will re-interview staff on how they use AI

An internal email sent Monday, October 26, and leaked to 404 Media, tells Proton staff that re-interviews on AI use will run November 4 to 11. The email calls them “consequential,” and the findings feed a wider review that could change roles or whether people stay. Interviewers come from outside each person’s team, picked from leaders who “demonstrate strong AI capability.” Staff should bring what they changed, the results, and what didn’t work, since “a list of tools alone will not demonstrate this.” On Reddit, Proton said it is “not doing layoffs and using AI as an excuse,” that it is hiring, and that it will offer support.

[IMAGE: Top WORK story]

Who’s Directing? I like the question Proton is asking, and I’m uneasy about the stakes attached to it. If I ran this, I’d ask everyone to bring a before-and-after first and hold the job decisions until people had a season to learn.

Read it: Tom’s Guide on the Proton re-interviews

Canada’s young workers were losing hiring ground before ChatGPT

Statistics Canada researchers Tahsin Mehdi and Marc Frenette tracked hiring for workers aged 15 to 29 through August 2026. Between the six months to November 2022 and the six months to August 2026, youth hiring fell in jobs where AI is likely to help (2.9% to 2.5%), in jobs more open to substitution (3.2% to 3.0%), and in less-exposed jobs too (4.0% to 3.3%). Coding-heavy jobs fell the most, from 2.5% to 1.9%. The decline began before ChatGPT, and separations held steady. The agency finds no broad evidence yet that AI is cutting hiring in Canada.

What to watch: whether coding-heavy jobs keep pulling away from the rest.

Who’s Directing? This is the calm number I want leaders to read before they blame AI for a hiring freeze. Young people are struggling to get a first job, and that problem deserves its own fix whatever the cause turns out to be.

Read it: Statistics Canada study · The Daily summary

Stanford asked workers what they actually do all day

TaskNet, from Stanford’s Digital Economy Lab working with Mercor and ONET analysts, had 132 workers file 2,726 reports on their own work. ONET analysts accepted all 202 resulting task statements, against about 65% of usual worker write-ins. The new map adds 79 tasks. It covered 89% of what held-out workers reported, against 79% for O*NET, and it raises measured AI exposure for software developers and lawyers.

Who’s Directing? The people doing the job saw the change before the database did. I’d ask my own team the same question before I bought anyone’s exposure score.

Read it: TaskNet paper on SSRN · Stanford project page On X: Molly Kinder on TaskNet

A cave-house rental in Spain hired an AI agent without meaning to

VJ Roworth and Adrian Beadell moved from England to run Leonard’s Lookout, a four-bedroom holiday rental carved into a cave in Spain. They pay about 20 euros a month for ChatGPT, and their agent, which they call Scout, now replies to Airbnb bookings, cross-checks calendars across platforms and keeps a plant care schedule. It even found a Forbes contributor and drafted the pitch. “The only time it won’t automatically send anything is if it has an impact or a change on how we run things currently,” Roworth said. The details are the owners’ own account.

Who’s Directing? The rule I’d copy is the one about what the agent can’t send on its own. They decided where its judgment stops, and that’s what keeps it their business.

Read it: Forbes on Leonard’s Lookout and Scout

LEARNING

Thirty learning scientists agree: learn it first, then practice with AI

The University of Minnesota asked 30 learning scientists and education researchers to rate how generative AI could help or hurt 65 learning processes, and they reached consensus on 42. Their advice is to learn without AI first, use it to practice through quizzes, feedback and tutoring, and then do your own thinking when you apply what you know. They rated critical thinking, metacognition, writing and belonging as vulnerable, with little upside, and saw elementary students as most at risk. They rated the evidence on most processes as thin. “If I were a teacher, I would definitely not provide access to my students if my goal is for them to develop certain skills,” lead author Panayiota Kendeou told MPR News.

[IMAGE: Top LEARNING story]

Who’s Directing? This matches what I’ve seen in my own learning. The struggle at the start is where it sticks, and I’d rather my kids meet the AI as a practice partner after they’ve done that work themselves.

Read it: University of Minnesota consensus report · MPR News coverage

Soongsil University sets three levels of AI use

Soongsil University in Seoul has new AI guidelines with three levels: Level 0 bans AI on an assignment, Level 1 allows limited use, and Level 2 permits it. Instructors state the level in the syllabus, and misconduct can’t be decided by AI detection results alone. “These guidelines are not rules intended to prevent AI use, but common standards to help students develop their ability to ask questions and make judgments for themselves,” said Lim Jong-in, who chairs the AI committee.

Who’s Directing? A level printed on the syllabus tells students what the work is for before they start it. I’d want that clarity in every course my kids take.

Read it: Financial News on Soongsil’s guidelines

Iowa adds Claude for Education and asks what students will do with it

At the University of Iowa’s Faculty Senate on October 27, senior vice provost Barrett Thomas said students, faculty and staff now have an AI Gateway and Claude for Education, with university data staying inside the university. Copilot and ChatGPT Edu are also offered, and a HIPAA-ready Claude is planned for 2027. How AI is used in a course stays with faculty. “Nobody knows, not us, not OpenAI, not Anthropic, what these use cases really are,” Thomas said, so the university wants to gather the data. Some faculty raised concerns about critical thinking.

Who’s Directing? I respect a provost who admits nobody knows the use cases yet. I’d add one thing to the data plan, which is asking students what they stopped doing themselves.

Read it: The Daily Iowan on Iowa’s AI tools

Quick hits from X

Sridhar Vembu on rebalancing. The Zoho founder says one smart person with AI can now do the work of ten, and he expects a painful rebalancing as industries that couldn’t afford engineers or marketers start hiring them. Read the post
Kevin Bryan on doing new things. Bryan argues AI pays off most when work gets redesigned, using a design tool that makes hundreds of variations free as his example. Read the thread
Keith Humphreys on a no-tech class. Humphreys says his Stanford guest lecture took 50% longer than last year because a no-tech rule brought far more student questions. Read the post

Who’s Directing? Humphreys lost time and gained questions, and I’d take that trade in any classroom. Saved time only matters if we spend it on something people can’t get from the tool.

Worth a listen

The a16z Show: Mark Cuban on Why AI Could Turn Solo Founders Into Trillionaires (October 28). About halfway through, Cuban says the biggest struggle for most organizations is not knowing how to put AI to work. “There will be the trillionaire in the basement,” he says. Watch the episode

That’s the Digest for Thursday. Write down one thing AI changed in your work this month, and what didn’t work, before anyone asks.

Todd

Sources

Tom’s Guide, Aleksandar Stevanović, “Proton to re-interview staff on AI use, and jobs may be on the line,” Oct 27, 2026 (reporting a leaked email first published by 404 Media): article
Statistics Canada, Tahsin Mehdi and Marc Frenette, “Hiring and job separation rates in the era of generative artificial intelligence: Early evidence,” Oct 28, 2026: study · The Daily
Wang, Haupt, Robinson, Lewis, Brynjolfsson and Yang, “TaskNet,” SSRN, Oct 26, 2026: paper · Stanford Digital Economy Lab
Forbes, Christine Haughney Dare-Bryan (contributor), “What Happens When Founders Embrace A Full Time AI Publicist,” Oct 26, 2026: article
Kendeou, Greene, Nixon and colleagues, “Ways generative AI can support and threaten learning in K-20 US education,” University of Minnesota CEHD, Oct 2026: report
MPR News, Elizabeth Shockman, Oct 28, 2026: article
Financial News (Korea), Soongsil University AI guidelines, Oct 27, 2026: article
The Daily Iowan, Oct 27, 2026: article
The a16z Show, Mark Cuban episode, Oct 28, 2026: episode