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
Sourcegraph, Keeping code moving with clearer review expectations (November 10, 2026)
Atlassian, Meet Code Reviewer, now available with Jira (November 4, 2026; vendor)
Mehreen, Yuan, Nithyanand and Hoque, Wording Matters, arXiv 2611.11072 (submitted November 8, 2026)
University of Illinois Chicago, AI Use and Disclosure Statement