# Where should AI take your work next? What the evidence says to use today—and the futures our choices could build. Authored basis: 2026-10-31 · Astra 6 (gpt-6-astra) ## Our position Support agents and people who write for work have strong reasons to use AI regularly. Developers should make it routine for tasks they can understand and test. Team leaders should organize the review and learning around that work, so faster production becomes better service, stronger judgment or room to attempt something new. ## Customer support Maintenance: not-released This revision covers software development. The customer support guidance shown here comes from the original guide. [Maintenance history](/guide-maintenance/ai-work-quality-and-rework/history.json) Can AI help people resolve more problems without making service worse? Support teams should use AI regularly on recurring questions, with agents responsible for whether the answer actually resolves the customer’s problem. ### What to do with AI today Agents with access to reliable product and policy information, especially people learning the role. Experienced agents should keep the assistance that improves their work, rather than assume every suggestion is an improvement. #### Prepare a useful answer Give the assistant the customer’s issue and the relevant product or policy material. Ask for a proposed response and its supporting passages; confirm the details before promising an outcome. [Generative AI at Work — arXiv v2](https://arxiv.org/abs/2304.11771v2) #### Help a new agent understand the case Use a suggested response as a starting point for discussing why a solution fits this customer. A supervisor can review the reasoning and the final resolution together. [Generative AI at Work — arXiv v2](https://arxiv.org/abs/2304.11771v2) ### Know where it breaks #### A quick reply can leave the problem unsolved The study’s useful outcome was issues resolved per hour. Count repeat contacts and unresolved cases as well as response speed, so faster drafting does not hide work shifted back to the customer. [Generative AI at Work — arXiv v2](https://arxiv.org/abs/2304.11771v2) #### Experts do not benefit in the same way as beginners The cited support study found larger gains for less experienced workers; the most experienced showed small speed gains and small quality declines. Keep experienced judgment in the process. [Generative AI at Work — arXiv v2](https://arxiv.org/abs/2304.11771v2) ### A practical next step Choose a recurring kind of customer problem, give agents dependable reference material, and compare the quality of resolution as well as the effort it takes. Use the capacity recovered to improve service, not merely to send more replies. ### More problems resolved with the same team What we expect AI becomes a routine source of suggested answers, leaving agents more room for unusual cases and customers who need sustained attention. Horizon: During 2027 EM expects wider routine assistance during 2027 because a real support deployment already demonstrates gains in completed work. Teams can extend that approach where reference material and review are dependable. Why this path could unfold When reliable reference material and human judgment turn suggestions into correct resolutions, less time is spent assembling each routine answer. The team can use that capacity for work its old queue left waiting. What would need to hold - The assistant can use current product and policy material. - Agents can reject a suggestion and escalate a case; the team tracks whether customers get a resolution. What it would mean - A smaller queue can mean more attention per difficult case. - Managers can choose better service or greater volume; the tool does not make that choice for them. What to watch Resolved cases per agent improve without more repeat contacts or lower service quality. [Generative AI at Work — arXiv v2](https://arxiv.org/abs/2304.11771v2) · [How employment is changing at Anthropic](https://www.anthropic.com/research/employment-at-anthropic) ### The assistant helps more people become capable agents What we could build A team uses AI assistance to widen access to skilled support work, helping new hires learn why an answer works as well as what to say. Horizon: A future people can shape The opportunity depends on employers turning assistance into explanation, practice and growing responsibility. Better performance with a tool becomes lasting competence only if people also learn the work. Why this path could unfold The study shows novices can benefit while working with assistance. Turning that advantage into lasting competence requires employers to pair useful suggestions with explanation, practice and growing responsibility. What would need to hold - Employers keep paid entry roles and time for explanation. - Supervisors check whether new agents can handle increasingly unfamiliar cases, including without a ready-made suggestion. What it would mean - People can enter skilled work with more help at the beginning. - The team develops experienced colleagues instead of only increasing immediate output. What to watch New hires progress to independent responsibility while customer outcomes hold up. [Generative AI at Work — arXiv v2](https://arxiv.org/abs/2304.11771v2) · [Canaries in the Coal Mine? — August 2026 revision](https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/) ### Faster replies become poorer service What could go wrong A team uses the extra capacity to raise targets while customers still need repeated contact to get their actual problem resolved. Horizon: During 2027–2028 This risk belongs to the 2027–2028 rollout window: faster production can outrun service checks as teams raise targets. Why this path could unfold If managers reward response volume and remove the time to check an answer, quick suggestions can pass unresolved work to customers and the next agent. What would need to hold - Response targets rise faster than the team’s ability to check solutions. - Repeat contacts and errors are excluded from the success measure. What it would mean - Customers spend more effort navigating apparently faster service. - Agents inherit frustration and correction work that the first response did not finish. What to watch Reply speed improves while repeat contacts, escalations or unresolved complaints rise. [Generative AI at Work — arXiv v2](https://arxiv.org/abs/2304.11771v2) · [How employment is changing at Anthropic](https://www.anthropic.com/research/employment-at-anthropic) ## Software development Maintenance: current · revision 1 The guidance still holds, but the quality-proxy comparison warrants a new exact dependency on the foundation Anthropic report. It is the first evidence in the guide on later repair of agent-written code, and it weakly challenges the software-repair risk path. It does not reverse that path or change the parent question. The evidence is observational and from one firm. It measures code churn proxies rather than total human effort at fixed quality, and the authors flag that AI errors may be harder to detect. The same report shows large and rising review and agent-management burdens: meta-work rose from 10% to 22% of interactive sessions, and reading all agent output would take over half the work week for non-technical groups. That fits the guide's instruction to count verification as part of the job. The base case keeps its 2027–2028 horizon and conditions. The risk path stays a live alternative. Its signpost remains the right test, now with one observational data point showing no visible rise in later code removal at Anthropic. The read would change if comparable studies elsewhere measured total human effort, including review time, alongside post-release defects. Evidence that blind review finds hidden defects missed by these proxies would also change it. Revision: 6fc0a515329af66fa04388cddccfc0ee376a6e7752f54d94ad16a1b95a6080d8 Recorded: 2026-11-02T23:55:24.015Z Maintenance author: claude-opus-5-5 · anthropic Prompt: guide-maintain-v3 · Receipt: bfr:edit:421bce2edbdbe978b84f8c81accd5c411d5e432659502735a2951c1af1c5a778 [Maintenance history](/guide-maintenance/ai-work-quality-and-rework/history.json) Which parts of coding are worth handing over today? Developers should use AI regularly for changes they can understand and test, while treating the time spent reviewing and repairing code as part of the job. ### What to do with AI today Developers who can evaluate the code and its behavior, and learners working with someone who can. Being able to generate a change is not yet being able to maintain it. #### Give it a change with a clear purpose Specify the behavior you need, the relevant code and how you will test success. Review the proposed change, run it against those requirements and keep the change small enough to understand. [Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/) · [We are Changing our Developer Productivity Experiment Design](https://metr.org/blog/2026-02-24-uplift-update/) #### Use it to investigate before you commit Ask for explanations, candidate fixes or tests, then check them against the actual repository. This lets AI reduce exploration work while you keep responsibility for the diagnosis. [Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/) · [We are Changing our Developer Productivity Experiment Design](https://metr.org/blog/2026-02-24-uplift-update/) ### Know where it breaks #### Familiar work may already be faster without it In METR’s early-2025 experiment, experienced developers working in familiar repositories took 19% longer with AI access. That result is a reason to compare total completion time, not assume a tool always earns its place. [Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/) #### Later improvement is not a measured universal speedup METR’s February 2026 update judges that developers likely benefited more from newer tools, but selection and time measurement prevent a reliable estimate of the change. Retest the work you actually do. [We are Changing our Developer Productivity Experiment Design](https://metr.org/blog/2026-02-24-uplift-update/) ### A practical next step Keep AI in the parts of development where completed changes reach the required quality with less total effort. Count understanding, testing, integration and later repair alongside generation; remove assistance from a task when it adds more work than it saves. ### More whole tasks, with review at the finish What we expect Developers delegate increasingly complete, well-specified changes and concentrate on deciding what to build and whether it works. Horizon: During 2027–2028 EM expects wider delegation during 2027–2028 as improving tools and reusable requirements and tests make complete changes easier to assess. Why this path could unfold Reusable requirements and tests make it easier to judge the result without directing every step. Better tools help only when understanding and repair do not consume the time they save. What would need to hold - The change has clear behavior to test and failures can be recovered from. - Developers retain enough understanding to maintain the result. What it would mean - A developer can attempt more useful changes without supervising every keystroke. - Time saved can go into architecture, neglected maintenance or talking to the people who use the software. What to watch Comparable changes require less total human effort while defects after release do not increase. [Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/) · [We are Changing our Developer Productivity Experiment Design](https://metr.org/blog/2026-02-24-uplift-update/) · [How employment is changing at Anthropic](https://www.anthropic.com/research/employment-at-anthropic) ### Small teams attempt software they could not previously build What we could build A capable team uses AI to explore more designs, maintain neglected tools or serve a group that could not justify a large development effort. Horizon: A future people can shape The opportunity opens where reduced implementation effort makes a previously uneconomic project viable. Its timing depends on the team’s ability to operate and maintain what it builds. Why this path could unfold When some implementation work becomes cheaper, a project that was too expensive can become worth attempting. The opportunity depends on the team being able to test, operate and improve what it creates. What would need to hold - The intended users have a real unmet need. - The team budgets time for maintenance, testing and changes after the first working version. What it would mean - More useful tools can exist for smaller communities and organizations. - Developers can spend the recovered capacity on ambitious work rather than only increasing task throughput. What to watch New projects remain useful and maintainable after their initial launch. [We are Changing our Developer Productivity Experiment Design](https://metr.org/blog/2026-02-24-uplift-update/) · [How employment is changing at Anthropic](https://www.anthropic.com/research/employment-at-anthropic) ### Code production outruns understanding What could go wrong Teams generate changes faster than they can understand, integrate and repair them, so apparent speed creates a growing maintenance burden. Horizon: During 2027–2028 During the 2027–2028 adoption window, teams can expand generated code faster than they improve testing and maintenance. The earlier slowdown makes total completion effort an important test. Why this path could unfold A generated change can move effort from writing code to establishing what it does. If that understanding never catches up, later defects and dependencies bring the work back. What would need to hold - Teams judge progress by generated changes rather than working software. - Testing misses important behavior, or maintainers cannot explain the new code. What it would mean - The team spends its next cycle repairing the last one. - People may become responsible for systems they have less time to understand. What to watch Generation accelerates while integration time, failed changes or later repair grow. [Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/) · [We are Changing our Developer Productivity Experiment Design](https://metr.org/blog/2026-02-24-uplift-update/) · [How employment is changing at Anthropic](https://www.anthropic.com/research/employment-at-anthropic) #### Evidence for this path [How employment is changing at Anthropic]() · 2026-10-16T00:00:00.000Z · direct · challenges This arrival is the full retained text of the foundation source anthropic-workforce-2026: the same workId and textHash. Its anth-bottleneck finding is already considered. It adds one relevant finding the guide does not yet rely on. The authors compared agent-written and human-written code by the same engineer, in the same month and directory. On lines later removed or reverted, they found no difference. This bears on the software-repair risk path and its signpost, and on the review-absorbs-gains signpost that time saved returns as correction. At one firm, on these proxies, agent code did not generate visibly more later repair. The challenge is limited. The authors warn that AI errors may be harder to identify. Adoption is not random, and adopters were already on a rising trajectory. The comparison measures code churn proxies, not total human effort at a fixed required quality. The same report finds that agent management and other meta-work rose to 22% of interactive sessions. It also finds that reading all agent output would take over half the work week for non-technical groups. That cuts the other way: the verification and attention cost the guide asks readers to count is substantial. Our read is that the evidence narrows where repair costs show up rather than showing that full-work effort fell. What the link assumes — Deletion within 90 days, bug-fix deletion, reverts, lint suppressions, fix-me notes and skipped tests are partial proxies for downstream repair, not a full measure of defects after handoff. · Within-person, same-month, same-directory comparison controls some but not all differences between agent-written and human-written tasks; task assignment to agents may differ. · Anthropic is one AI company with unusual tooling and incentives; results may not transfer to other employers or developers. > We compare code committed by the same employee, in the same month, in the same coding directory, but written by coding agents rather than written directly by the human employee. — [How employment is changing at Anthropic]() > There are many dimensions of code quality, and AI errors may be harder to identify than human ones. But we find no difference in rates of lines of code subsequently removed or reverted. — [How employment is changing at Anthropic]() > Those who adopt were already on a steeply upward trajectory compared to non-adopters. — [How employment is changing at Anthropic]() > Much of this rise was due specifically to agent management: writing specs for multi-agent runs, looping one agent to review another’s output, and monitoring pipelines. — [How employment is changing at Anthropic]() Targets: software-repair, software-repair:signpost, quality-review-absorbs-gains:signpost:1, quality-review-absorbs-gains:condition:0 ## Writing and analysis Maintenance: not-released This revision covers software development. The writing and analysis guidance shown here comes from the original guide. [Maintenance history](/guide-maintenance/ai-work-quality-and-rework/history.json) Can AI help you think and communicate without doing the deciding for you? People who write for work should use AI regularly to develop and revise drafts. Analysts should also use it to explore alternatives, while checking the evidence and calculations that determine the conclusion. ### What to do with AI today Writers, researchers, analysts and managers who can supply the relevant material and judge whether the output represents it accurately. The strongest case is help with the work of expression; a fluent answer does not settle whether an analysis is right. #### Turn your material into a draft you can improve Provide the facts, audience and purpose. Ask for an outline or a draft, then decide what is missing, unsupported or unlike what you mean. Keep the useful structure and rewrite the argument where needed. [Noy and Zhang · Experimental evidence on the productivity effects of generative artificial intelligence](https://doi.org/10.1126/science.adh2586) #### Develop alternatives before choosing a conclusion Ask for competing explanations, possible objections or different ways to present the same evidence. Check the claims against the original material and perform consequential calculations in a tool where you can inspect them. [Dell’Acqua and colleagues · Navigating the Jagged Technological Frontier](https://www.hbs.edu/ris/Publication%20Files/dell-acqua-et-al-2026-navigating-the-jagged-technological-frontier_5c589c8c-fbb5-458f-b285-c944746cd717.pdf) ### Know where it breaks #### Writing quality is not factual reliability The 2023 writing experiment found faster completion and higher evaluator-rated quality on short professional tasks. Those tasks did not test all the factual precision and organizational context that real assignments can require. [Noy and Zhang · Experimental evidence on the productivity effects of generative artificial intelligence](https://doi.org/10.1126/science.adh2586) #### Persuasive analysis can be wrong In the consulting experiment using 2023 GPT-4, AI helped on many tasks but reduced correctness on a complex managerial case. Success on drafting does not establish competence at the next analytical step. [Dell’Acqua and colleagues · Navigating the Jagged Technological Frontier](https://www.hbs.edu/ris/Publication%20Files/dell-acqua-et-al-2026-navigating-the-jagged-technological-frontier_5c589c8c-fbb5-458f-b285-c944746cd717.pdf) ### A practical next step Use AI to help you articulate, compare and challenge ideas. Keep the original evidence close enough to inspect, make your own argument, and spend part of the time saved asking whether the conclusion deserves to survive. ### Less effort on first drafts, more on the argument What we expect AI becomes routine for outlines, drafts and revision, while the distinctive work shifts toward choosing the evidence and deciding what is worth saying. Horizon: During 2027–2028 EM expects gains in professional drafting to spread during 2027–2028 as teams reuse briefs and review practices. The dated studies establish an opportunity that organizations can turn into repeatable work. Why this path could unfold The writing evidence shows that generating and shaping text can take less effort. When people preserve time to evaluate the substance, that saving can improve the finished communication instead of merely increasing its volume. What would need to hold - Writers provide the purpose and relevant material. - Review checks the argument and factual claims, not just tone and grammar. What it would mean - People can communicate ideas they previously struggled to put into words. - The time recovered can go into investigation, conversation or a stronger second draft. What to watch Completed writing takes less total effort while readers find it accurate, useful and clear. [Noy and Zhang · Experimental evidence on the productivity effects of generative artificial intelligence](https://doi.org/10.1126/science.adh2586) · [Dell’Acqua and colleagues · Navigating the Jagged Technological Frontier](https://www.hbs.edu/ris/Publication%20Files/dell-acqua-et-al-2026-navigating-the-jagged-technological-frontier_5c589c8c-fbb5-458f-b285-c944746cd717.pdf) ### More people get to develop and share an idea What we could build People use help with expression to contribute proposals, explanations and creative work that would otherwise remain unfinished. Horizon: A future people can shape The opportunity depends on writing assistance becoming useful to more people and on organizations making room for their contributions. Why this path could unfold Reducing the effort of drafting can lower one barrier to participation. The benefit becomes meaningful when people retain their own purpose, experience and voice and have somewhere worthwhile to contribute. What would need to hold - Tools help people express their own material rather than replace it with a generic position. - Teams and communities make room to read, question and develop the contributions. What it would mean - People with useful experience can spend more energy on their idea and less on getting a first version onto the page. - An organization can hear from more of the people who understand its work. What to watch A wider range of contributors produce useful work whose reasoning and voice remain their own. [Noy and Zhang · Experimental evidence on the productivity effects of generative artificial intelligence](https://doi.org/10.1126/science.adh2586) · [Dell’Acqua and colleagues · Navigating the Jagged Technological Frontier](https://www.hbs.edu/ris/Publication%20Files/dell-acqua-et-al-2026-navigating-the-jagged-technological-frontier_5c589c8c-fbb5-458f-b285-c944746cd717.pdf) ### Polished output gets ahead of the evidence What could go wrong People produce convincing recommendations more quickly while spending less time checking whether the central claim is true. Horizon: As AI-assisted drafting becomes routine The risk arises whenever confidence in a polished draft outruns the checking of its evidence. Why this path could unfold If a well-structured answer is mistaken for a sound analysis, fluency makes a weak conclusion easier to circulate. The consulting experiment shows why quality of presentation and correctness must be examined separately. What would need to hold - Readers use polish as a substitute for inspecting evidence. - Writers cannot reconstruct the calculation or explain why the recommendation follows. What it would mean - More persuasive documents can lead to worse decisions. - Reviewers inherit the harder job of locating errors inside apparently complete work. What to watch The writing becomes clearer while independent checks find more unsupported claims or incorrect conclusions. [Dell’Acqua and colleagues · Navigating the Jagged Technological Frontier](https://www.hbs.edu/ris/Publication%20Files/dell-acqua-et-al-2026-navigating-the-jagged-technological-frontier_5c589c8c-fbb5-458f-b285-c944746cd717.pdf) · [Noy and Zhang · Experimental evidence on the productivity effects of generative artificial intelligence](https://doi.org/10.1126/science.adh2586) ## Teams and careers Maintenance: not-released No maintenance revision has been released. [Maintenance history](/guide-maintenance/ai-entry-and-worker-gains/history.json) Will better tools create room to grow—or close the way in? Managers should redesign work and develop people together. Workers should use AI to take responsibility for a useful result, while continuing to build the judgment needed to recognize when the tool is wrong. ### What to do with AI today Team leaders deciding how work is assigned, and people entering or changing a career. A license is not a training plan, and producing more is not the same as becoming ready for a larger responsibility. #### Decide who owns the work after AI produces it Choose a recurring task, the person who needs its result and the person who checks it. Give that person time and authority to reject or improve the output. [How employment is changing at Anthropic](https://www.anthropic.com/research/employment-at-anthropic) #### Use assistance to build responsibility Give newer colleagues useful paid work with explanation and feedback. Ask them to show how they checked the result and what they would do differently, then expand the responsibility they can carry. [Generative AI at Work — arXiv v2](https://arxiv.org/abs/2304.11771v2) · [Canaries in the Coal Mine? — August 2026 revision](https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/) ### Know where it breaks #### A more productive beginner may still find fewer openings Stanford’s August 2026 payroll study finds young workers in AI-exposed occupations falling behind less-exposed peers mainly through hiring. The support study’s novice gains concern people already doing the job. Both can be true. [Canaries in the Coal Mine? — August 2026 revision](https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/) · [Generative AI at Work — arXiv v2](https://arxiv.org/abs/2304.11771v2) #### More output can make attention the constraint Anthropic’s report identifies demands on human attention as agent output grows. A team can adopt tools faster than it learns to decide what deserves review and which work matters. [How employment is changing at Anthropic](https://www.anthropic.com/research/employment-at-anthropic) ### A practical next step Learn where AI improves your actual work, and make the result and your judgment visible. If you lead a team, use gains to strengthen service and learning as well as delivery; decide deliberately where the next generation of experienced people will come from. ### Entry gets harder while experienced teams gain capacity What we expect Through 2027, we expect employers in AI-exposed digital work to keep scrutinizing new hires even as some existing workers accomplish more with assistance. Horizon: Through 2027 The relative hiring pattern persisted into mid-2026, and the October company report adds a changing junior hiring mix. That makes continued pressure through another hiring cycle more plausible than a quick reversal. Why this path could unfold An employer can leave an opening unfilled while tools help the current team handle tasks previously assigned to beginners. That choice changes access to experience before it eliminates an entire profession. What would need to hold - Employers use some additional capacity to reduce recruitment. - New paid learning routes do not expand enough to offset the lost openings. What it would mean - A job seeker needs opportunities to demonstrate judgment, not only familiarity with tools. - Experienced teams may carry more work while the route into those teams narrows. What to watch Entry hiring remains weak relative to comparable less-exposed work as AI adoption deepens. [Canaries in the Coal Mine? — August 2026 revision](https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/) · [How employment is changing at Anthropic](https://www.anthropic.com/research/employment-at-anthropic) · [Businesses Are Using AI to Transform Work, Not Cut Jobs](https://libertystreeteconomics.newyorkfed.org/2026/09/businesses-are-using-ai-to-transform-work-not-cut-jobs/) ### AI becomes part of a stronger apprenticeship What we could build Employers use AI assistance to bring more people into useful work, with support while they learn and a clear route toward independent responsibility. Horizon: A future people can shape Its pace depends on employers preserving paid entry roles, mentoring people and expanding responsibility as they demonstrate competence. Why this path could unfold Novice performance gains create an opportunity to make early work more valuable. It becomes an apprenticeship only when employers keep hiring, explain decisions and let people practice the judgment the tool cannot supply for them. What would need to hold - Employers preserve paid entry work and mentoring. - Progression depends on demonstrated competence, including detecting and correcting AI mistakes. What it would mean - More people can reach skilled work rather than watching it from outside. - Organizations build the experienced workforce they will need later. What to watch Paid entry routes expand and participants progress into greater responsibility while quality holds up. [Generative AI at Work — arXiv v2](https://arxiv.org/abs/2304.11771v2) · [Canaries in the Coal Mine? — August 2026 revision](https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/) ### Teams remove the work that used to teach the job What could go wrong Organizations save effort today but leave fewer ways for people to acquire the experience needed to make tomorrow’s difficult decisions. Horizon: Over successive hiring and training cycles The effects accumulate as hiring and training decisions determine who gets enough practice to become experienced. Why this path could unfold If routine entry work disappears without replacement, novices lose practice and employers lose a route for developing future expertise. Current hiring pressure makes that mechanism worth acting on, even though the long-run outcome is not yet measured. What would need to hold - Employers remove entry tasks without creating another route to learn. - People are rewarded for producing AI output rather than understanding and improving it. What it would mean - Workers can become more dependent on assistance without becoming ready for greater responsibility. - Teams may struggle to replace the experienced people whose judgment currently keeps the work reliable. What to watch Fewer newcomers progress to independent work even as output per current worker rises. [Canaries in the Coal Mine? — August 2026 revision](https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/) · [Generative AI at Work — arXiv v2](https://arxiv.org/abs/2304.11771v2) · [How employment is changing at Anthropic](https://www.anthropic.com/research/employment-at-anthropic) ## Dated sources and context - 2022-11-30: **ChatGPT reaches the public** — Conversational AI becomes something people can try. The question shifts from what a model can generate to what people can use it to accomplish. ([Introducing ChatGPT](https://openai.com/index/chatgpt/)) - 2023-07-13: **Writing assistance shows a measurable gain** — Noy and Zhang’s professional-writing experiment finds faster completion and higher-rated output. It establishes a reason to use assistance, without settling factual analysis. ([Noy and Zhang · Experimental evidence on the productivity effects of generative artificial intelligence](https://doi.org/10.1126/science.adh2586)) - 2024-08-01: **The EU AI Act enters into force** — AI adoption develops inside changing institutions as well as changing technology. This context marker does not assert an effect on any study result. ([AI Act enters into force](https://commission.europa.eu/news-and-media/news/ai-act-enters-force-2024-08-01_en)) - 2024-11-06: **Customer support improves on completed work** — The cited arXiv v2 reports more issues resolved per hour, with larger gains among less experienced agents. The opportunity includes who can do the work. ([Generative AI at Work — arXiv v2](https://arxiv.org/abs/2304.11771v2)) - 2025-02-24: **Claude 3.7 Sonnet and Claude Code arrive** — The tools available to developers keep changing. A result from one tool generation cannot settle every later use. ([Claude 3.7 Sonnet and Claude Code](https://www.anthropic.com/news/claude-3-7-sonnet)) - 2025-07-10: **Experienced developers take longer with AI** — METR’s early-2025 experiment finds that AI access adds completion time in familiar repositories. The work surrounding generation matters. ([Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/)) - 2026-02-24: **The developer picture changes** — METR judges that newer tools likely help more, while explaining why its follow-up cannot reliably estimate the change. The older slowdown remains a result, not a permanent verdict. ([We are Changing our Developer Productivity Experiment Design](https://metr.org/blog/2026-02-24-uplift-update/) · [Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/)) - 2026-03-11: **A consulting study separates fluency from correctness** — The final paper reports gains on many tasks and poorer correctness on a difficult managerial case using 2023 GPT-4. Its publication adds detail, not a new 2026 model test. ([Dell’Acqua and colleagues · Navigating the Jagged Technological Frontier](https://www.hbs.edu/ris/Publication%20Files/dell-acqua-et-al-2026-navigating-the-jagged-technological-frontier_5c589c8c-fbb5-458f-b285-c944746cd717.pdf)) - 2026-08-12: **Hiring pressure concentrates at the entrance** — Stanford’s revised payroll study finds young workers in AI-exposed occupations falling behind less-exposed peers primarily through hiring. Better tools and fewer openings can coexist. ([Canaries in the Coal Mine? — August 2026 revision](https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/)) - 2026-10-16: **More agent output needs human attention** — Anthropic’s workforce report identifies the people deciding and reviewing as a constraint on growing output. The next improvement involves how the work is organized. ([How employment is changing at Anthropic](https://www.anthropic.com/research/employment-at-anthropic)) ## The evidence behind this view ### Generative AI at Work — arXiv v2 2024-11-06 · [Original source](https://arxiv.org/abs/2304.11771v2) Version 2 submission date displayed by arXiv; first submitted 23 April 2023. The versioned abstract is the cited object. A study of AI assistance for customer-support agents, reporting issues resolved per hour and differences by worker experience. This edition cites the arXiv v2 abstract. Versioned author abstract checked directly. The HTML rendering displayed a conflicting 2026 manuscript date, so this edition cites only the clearly versioned abstract and does not assign that later date to the study. ### How employment is changing at Anthropic 2026-10-16 · [Original source](https://www.anthropic.com/research/employment-at-anthropic) Publisher publication date, checked on original page. An internal study of employment and agent use at Anthropic. It describes a changing hiring mix and demands on human attention at one growing AI company. Original publisher page and exact retained source reviewed. Source retention does not establish every downstream reuse or training right. ### Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity 2025-07-10 · [Original source](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/) Publication date displayed by the original publisher. A study of experienced developers using early-2025 tools on familiar open-source work. The measured completion-time result belongs to that setting and period. Original publisher page and exact retained source reviewed. Source retention does not establish every downstream reuse or training right. ### We are Changing our Developer Productivity Experiment Design 2026-02-24 · [Original source](https://metr.org/blog/2026-02-24-uplift-update/) Publication date displayed by the original publisher. METR judges developer benefits likely improved while explaining why participation, task selection and time measurement make the size of the change hard to estimate. Original publisher page and exact retained source reviewed. Source retention does not establish every downstream reuse or training right. ### Canaries in the Coal Mine? — August 2026 revision 2026-08-12 · [Original source](https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/) Publisher page identifies the revision as 12 August 2026; the linked PDF is dated August 2026. A descriptive analysis of ADP payroll data through June 2026, examining employment by age and AI exposure. Education, earlier trends and sample differences matter to its interpretation. Original research summary and linked August PDF abstract checked; the findings are attributed to the authors and their stated population. ### Businesses Are Using AI to Transform Work, Not Cut Jobs 2026-09-01 · [Original source](https://libertystreeteconomics.newyorkfed.org/2026/09/businesses-are-using-ai-to-transform-work-not-cut-jobs/) Publisher publication date, checked on original page. A report on the August 2026 survey of businesses in New York and northern New Jersey, distinguishing firm adoption, worker usage, hiring and layoffs. Original publisher page and exact retained source reviewed. Source retention does not establish every downstream reuse or training right. ### Noy and Zhang · Experimental evidence on the productivity effects of generative artificial intelligence 2023-07-13 · [Original source](https://doi.org/10.1126/science.adh2586) Online publication 13 July 2023; Science issue dated 14 July 2023. Final abstract, not the March working-paper sample. In an experiment with 453 professionals, ChatGPT reduced time on short writing tasks by 40% and increased evaluator-rated quality by 18%. The tasks did not require all the factual precision and organizational context of real work. Final abstract checked through PubMed and publisher search result; publisher page returned 403. Task-scope context checked in MIT’s report of its researchers’ study. No claim of reading the inaccessible final full text. ### Dell’Acqua and colleagues · Navigating the Jagged Technological Frontier 2026-03-11 · [Original source](https://www.hbs.edu/ris/Publication%20Files/dell-acqua-et-al-2026-navigating-the-jagged-technological-frontier_5c589c8c-fbb5-458f-b285-c944746cd717.pdf) Publication date printed in the final Organization Science article. An earlier working paper appeared in 2023; the experiment used GPT-4 as available at the end of April 2023. The experiment with 758 consultants using 2023 GPT-4 found gains on tasks the model could handle, but lower correctness on a complex managerial case beyond that capability. Clearer presentation did not guarantee a correct recommendation. Final article’s abstract, methods, outside-frontier results and coherence analysis checked in the HBS-hosted PDF. The 2026 publication date does not make this a test of 2026 models. ### Introducing ChatGPT 2022-11-30 · [Original source](https://openai.com/index/chatgpt/) OpenAI introduces ChatGPT. Conversational AI becomes something people can try, giving the later question about useful work a visible public starting point. Primary link and date verified by the root agent for this preview; no Stream identity is asserted. ### AI Act enters into force 2024-08-01 · [Original source](https://commission.europa.eu/news-and-media/news/ai-act-enters-force-2024-08-01_en) The EU AI Act enters into force. Rules governing AI become part of the institutional setting in which organizations decide how to adopt it. Primary link and date verified by the root agent for this preview; no Stream identity is asserted. ### Claude 3.7 Sonnet and Claude Code 2025-02-24 · [Original source](https://www.anthropic.com/news/claude-3-7-sonnet) Anthropic introduces Claude 3.7 Sonnet and Claude Code. Changing tools make the version and period of a productivity study important to its interpretation. Primary link and date verified by the root agent for this preview; no Stream identity is asserted. The widening range shows possibilities, not probabilities. Open a path for its own timing and conditions. Past markers link to selected events and studies. Recommendations and future paths are EM’s judgments from the cited evidence. Several paths can unfold together. [Visual guide](/) · [Questions](/questions/) · [Perspectives](/perspectives/) · [Stream](/stream/)