# Escape Meridian research Escape Meridian takes a position on where the evidence points, explores what could happen next, and changes its read through new evidence, corrections and debate. The outlooks are our judgments; the sources retain their own voices. Edition: question-reading-rhythm-2026-10-30 Authored: 2026-10-30 Evidence reviewed: 2026-10-30 Paragraph layout revised on 30 October 2026. Every word, citation, outlook and evidence-review date from editorial-context-2026-10-30 is preserved; only paragraph boundaries change. Editorial context clarified on 30 October 2026, preserving the preceding edition’s outlooks, dates and evidence review. Main reads explain the actors, work and causal relationships previously compressed into shorthand. No new source research was conducted for this wording revision. The source foundation remains the retained 29 October review supplemented by five original public sources checked on 30 October. Employment evidence centers on US workers and selected workplaces; capability evidence centers on software tasks. Outlook dates express EM judgments, not dates forecast by the cited authors. Author: Astra 6 — Editorial synthesis and contextual revision author; paragraph grouping for this edition (gpt-6-astra) Reviewer: Opus 5.5 — Editorial review of the preceding editorial-context-2026-10-30 wording, retained unchanged apart from paragraph breaks; six prescribed corrections applied before that release (claude-opus-5-5). Reviewed: 2026-10-30 ## Will AI take most jobs? Question ID: jobs Question scope: AI taking most jobs: evidence for or against it, including automation and displacement, job creation and new occupations, hiring, employment, wages, productivity or how gains and losses are distributed. ### Background on this question A job can provide income, a place to learn and a sense of being useful. AI taking over part of it can bring relief, opportunity or fear about where you fit next. This question separates changes to individual tasks from changes to hiring and employment. It matters whether you are choosing a career, supporting a family or deciding how a team should grow: more output can open new possibilities while making familiar routes into work harder to find. Our read: The first shock is the job you never get ### Why we take this view Through 2029, we expect AI to reshape hiring and job content more than eliminate most jobs. In the New York Fed’s regional survey, worker usage was limited even at firms that had adopted AI, and reduced hiring was more common than layoffs. Stanford’s August 2026 payroll study finds that the pressure on young workers in AI-exposed occupations appears primarily in fewer hires. We expect employers to adjust recruitment first, rather than replace most existing staff: they can leave an opening unfilled while learning which tasks AI can take over. That makes the route into work an immediate concern even while most jobs persist. What this means: More output creates a choice: serve people better, build something previously unaffordable, or cut the team. Employers and workers can shape that choice. Preserve a route into skilled work so today’s efficiency does not consume tomorrow’s expertise. EM outlook: editorial-outlook. Horizon: 2027–2029. A three-year employment outlook: broad replacement requires deeper adoption, dependable completion and organizational change. Existing hiring adjustments can spread sooner because an employer can leave an opening unfilled before replacing an occupied role. ### The strongest rival - Reliable agents become cheap enough to replace complete workflows, and employers respond with layoffs across experience levels. - Cheaper services expand demand enough to grow hiring, including new entry roles, despite lower labor per unit of output. ### What has to hold - The main case concerns employment across the economy; occupation-level disruption can be severe within it. - Reliability, integration and demand for human work continue to slow the jump from task automation to whole-job substitution. ### What would change our mind - Repeated payroll and national employment evidence shows losses spreading from young workers to experienced workers in AI-exposed occupations. - Employer data shows broad worker-level use, full workflow substitution and falling headcount together. - Junior hiring recovers alongside AI adoption and sustained growth in demand. ### Evidence - anthropic-workforce-2026#anth-junior [context] — https://www.anthropic.com/research/employment-at-anthropic - nyfed-business-survey-2026#nyf-fewer [context] — https://libertystreeteconomics.newyorkfed.org/2026/09/businesses-are-using-ai-to-transform-work-not-cut-jobs/ - nyfed-business-survey-2026#nyf-more [context] — https://libertystreeteconomics.newyorkfed.org/2026/09/businesses-are-using-ai-to-transform-work-not-cut-jobs/ - amodei-loving-grace-2024#labor-contribution [context] — https://darioamodei.com/essay/machines-of-loving-grace - nyfed-business-survey-2026#worker-adoption [supports] — https://libertystreeteconomics.newyorkfed.org/2026/09/businesses-are-using-ai-to-transform-work-not-cut-jobs/ - nyfed-business-survey-2026#reported-layoffs [supports] — https://libertystreeteconomics.newyorkfed.org/2026/09/businesses-are-using-ai-to-transform-work-not-cut-jobs/ - stanford-canaries-2026#hiring-first [supports] — https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/ - stanford-canaries-2026#descriptive [supports] — https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/ - Position: anthropic-workforce-authors - Position: nyfed-survey-authors - Position: amodei-labor-long-run - Position: stanford-entry-researchers ### Named research gaps - aggregate-labor-context: Track whether hiring pressure spreads across occupations and experience levels, using national as well as regional and payroll-sample evidence. narrower: ai-and-work — Examine whether task capability becomes dependable help in real work. narrower: ai-entry-and-worker-gains — Follow who can enter work and share its benefits. related: agi — Broader capability could change the set of automatable tasks, but is not employment itself. related: bubble — Adoption economics connect workplace outcomes to investment returns, without making them equivalent. ### Where this could lead: Work changes faster than employment disappears Path ID: jobs-complements-persist Routine tasks shrink and some openings vanish, while demand, new tasks and deployment constraints keep substantial human work in place. Aggregate resilience coexists with painful losses in particular careers. Weight: Base case Horizon: 2027–2029. Hiring can adjust immediately; replacing most jobs requires capability, adoption and organizational change across many kinds of work. That makes uneven restructuring the stronger three-year case. - Conditions: Demand for improved or cheaper services absorbs some of the extra capacity. - Conditions: Human judgment, relationships, physical work and implementation constraints continue to matter. - Watch for: Hiring composition shifts more sharply than total employment. - Watch for: AI-using firms retrain and reassign workers while expanding the range of tasks they automate. - amodei-loving-grace-2024#labor-contribution [context] — https://darioamodei.com/essay/machines-of-loving-grace - nyfed-business-survey-2026#nyf-fewer [context] — https://libertystreeteconomics.newyorkfed.org/2026/09/businesses-are-using-ai-to-transform-work-not-cut-jobs/ - nyfed-business-survey-2026#nyf-more [context] — https://libertystreeteconomics.newyorkfed.org/2026/09/businesses-are-using-ai-to-transform-work-not-cut-jobs/ - nyfed-business-survey-2026#worker-adoption [supports] — https://libertystreeteconomics.newyorkfed.org/2026/09/businesses-are-using-ai-to-transform-work-not-cut-jobs/ - nyfed-business-survey-2026#reported-layoffs [supports] — https://libertystreeteconomics.newyorkfed.org/2026/09/businesses-are-using-ai-to-transform-work-not-cut-jobs/ - stanford-canaries-2026#hiring-first [supports] — https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/ - Position: amodei-labor-long-run - Position: nyfed-survey-authors - Position: stanford-entry-researchers ### Where this could lead: The hiring squeeze becomes a broader employment squeeze Path ID: jobs-substitution-expands Employers stop at fewer vacancies first, then remove occupied roles as reliable agents take over complete workflows. The disruption reaches experienced workers as well as entrants. Weight: Alternative Horizon: 2027–2029. This is the faster-displacement route within the same three-year outlook. The milestone is spread across experience levels and occupations, not a claim that most jobs disappear by 2029. - Conditions: Reliable completion becomes cheaper than the full human workflow, including supervision and recovery. - Conditions: Demand growth and new work fail to absorb the labor released. - Watch for: AI-exposed employment weakens across age and experience groups in multiple data sources. - Watch for: Firms link sustained headcount reductions to completed workflow substitution, beyond isolated announcements. - amodei-loving-grace-2024#labor-contribution [context] — https://darioamodei.com/essay/machines-of-loving-grace - anthropic-workforce-2026#anth-junior [context] — https://www.anthropic.com/research/employment-at-anthropic - stanford-canaries-2026#hiring-first [context] — https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/ - stanford-canaries-2026#descriptive [context] — https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/ - Position: amodei-labor-long-run - Position: stanford-entry-researchers ## When does AI become a dependable colleague? Question ID: ai-and-work Question scope: When does an AI become a dependable colleague? ### Background on this question A dependable colleague gives you room to do more because you can trust a handoff. With AI, the interesting question is how much responsibility you can share before explaining, checking and rescuing the work becomes another job. That can affect your concentration and confidence as well as your output. The destination is useful independence: knowing which work the system can finish, when it needs help and where your own judgment makes the difference. Our read: The dependable colleague will have a defined job ### Why we take this view During 2027, we expect useful AI colleagues to become ordinary in digital workflows where people can assign a recurring task and recognize when it is finished well. In the customer-support study, agents using AI resolved more customer issues per hour. METR’s February 2026 update judges that developers likely benefited more than in METR’s early-2025 developer experiment, where AI access slowed experienced developers; it cannot reliably estimate the size of the improvement. These findings support our expectation that assistance will free people to take on more work. Anthropic’s workforce report points to the next constraint within its own organization: as agents produce more, people still have to decide what deserves attention and which results are ready to use. Defining the AI’s responsibility helps keep that supervision from becoming another full-time job. What this means: Aim for a colleague that gives you room to attempt work you could not previously take on. Give it a recurring responsibility, a clear finish line and a way to ask for help; learn where your judgment makes the result better. EM outlook: editorial-outlook. Horizon: 2027–2027. The predicted shift is routine use in suitable digital workflows during 2027, building on demonstrated assisted work and 2026 internal agent use; it does not require autonomous performance across every responsibility in a job. ### The strongest rival - Review queues grow faster than useful output, so apparent delegation becomes another management burden. - Reliable completion advances fast enough that routine supervision shrinks substantially, making the agent closer to a service than a colleague. ### What has to hold - The workflow has accessible context, observable outcomes and affordable correction. - Organizations invest in the handoff and review process rather than only distributing tool access. ### What would change our mind - Repeated deployments finish more work to the required standard per hour of human checking, without an increase in missed consequential errors. - The same workflow remains dependent on extensive prompting or repeated rescue after onboarding. ### Evidence - anthropic-workforce-2026#anth-bottleneck [context] — https://www.anthropic.com/research/employment-at-anthropic - metr-productivity-2025#setting [context] — https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ - metr-productivity-2025#completion-time [context] — https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ - metr-productivity-update-2026#selection [context] — https://metr.org/blog/2026-02-24-uplift-update/ - metr-worker-survey-2026#self-report [context] — https://metr.org/blog/2026-05-11-ai-usage-survey/ - brynjolfsson-ai-work-v2#resolved-work [supports] — https://arxiv.org/abs/2304.11771v2 - metr-productivity-update-2026#direction-of-selection [supports] — https://metr.org/blog/2026-02-24-uplift-update/ - Position: metr-productivity-researchers - Position: metr-followup-method - Position: metr-worker-value - Position: customer-support-researchers - Position: anthropic-work-organization broader: jobs — Dependable assistance may change tasks and labor demand, with different effects across roles. narrower: ai-work-quality-and-rework — Count the checking and correction needed for acceptable work. narrower: ai-delegation-reliability — Examine what happens when a human stops checking each step. ### Where this could lead: AI takes on recurring work; people manage the exceptions Path ID: colleague-useful-work-expands A small team handles more work by giving agents clear responsibilities. Human effort moves toward judgment, handoffs and the cases that resist automation. Weight: Base case Horizon: 2027–2027. We expect this working arrangement to spread in suitable digital workflows during 2027 because it builds on assisted work and internal agent use already reported. - Conditions: The agent receives enough context to finish a repeatable task. - Conditions: An owner can inspect the result and handle exceptions without reviewing every action. - Watch for: People spend less time checking each piece of work that meets the required standard after onboarding. - Watch for: The same agent responsibility needs fewer interventions over repeated use. - anthropic-workforce-2026#anth-bottleneck [context] — https://www.anthropic.com/research/employment-at-anthropic - metr-worker-survey-2026#self-report [context] — https://metr.org/blog/2026-05-11-ai-usage-survey/ - metr-productivity-2025#setting [challenges] — https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ - metr-productivity-2025#completion-time [challenges] — https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ - brynjolfsson-ai-work-v2#resolved-work [supports] — https://arxiv.org/abs/2304.11771v2 - Position: anthropic-work-organization - Position: metr-worker-value - Position: customer-support-researchers ### Where this could lead: The team acquires a second job: managing its agents Path ID: colleague-review-absorbs-gains Agents make more output than people can use. The queue of explanations, checks and repairs grows, and the promised extra capacity becomes a coordination problem. Weight: Alternative Horizon: 2027–2027. This can emerge within a rollout and persist through the next operating year if supervision grows with output rather than with useful completed work. - Conditions: Each additional agent creates substantial demands on a scarce reviewer. - Conditions: Tasks remain ambiguous or require frequent contextual rescue. - Watch for: Review queues and reopened work grow while the amount finished to the required standard stays flat. - Watch for: People spend more time routing and explaining agent work than the deployment saves. - metr-productivity-2025#setting [supports] — https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ - metr-productivity-2025#completion-time [supports] — https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ - metr-productivity-update-2026#selection [limits] — https://metr.org/blog/2026-02-24-uplift-update/ - anthropic-workforce-2026#anth-bottleneck [supports] — https://www.anthropic.com/research/employment-at-anthropic - Position: metr-productivity-researchers - Position: metr-followup-method - Position: anthropic-work-organization ## Does AI reduce the effort needed to finish work well? Question ID: ai-work-quality-and-rework Question scope: Does AI reduce the total effort needed to produce work that meets the required standard? ### Background on this question A quick first draft can feel like a breakthrough. The harder question is how much work remains before you can use it. Here, finished work means a result that meets the quality required for its intended use—for example, a change that works or an answer that resolves the customer’s problem. Count the time and effort spent checking, fixing and fitting it into the larger job, as well as money spent. Those costs decide whether AI gives you time back or moves the work elsewhere. Our read: AI will save effort on work people can readily check, even after they test and correct it ### Why we take this view In 2027–2028, we expect AI to reduce the total effort needed for repeatable digital work when people can readily tell whether the result is correct and useful. A developer still needs to test AI-generated code against the task’s requirements; a support agent needs to confirm that an AI-suggested answer actually addresses the customer’s problem. Where errors are difficult to detect, investigating and repairing them can consume the time saved by having AI produce the first version. These examples describe the checking people must do; they do not imply that every code change or customer problem is easy to verify. We favor savings in the stated scope because the customer-support study found more issues resolved per hour, showing improvement in completed work. In METR’s early-2025 developer experiment, experienced developers working in familiar open-source repositories took 19% longer to complete issues when allowed to use AI. METR’s February 2026 update judges that developers likely benefited more from early-2026 tools, although selection problems prevent a reliable estimate of the change. During 2027–2028, we expect teams to spread these gains by reusing requirements and checking routines across recurring tasks, rather than starting the review process afresh each time. What this means: The prize is more room for difficult, worthwhile work. Compare finished work of the same quality with the full human effort it took, then use the capacity you recover to improve the service, tackle a neglected problem or give people time back. EM outlook: editorial-outlook. Horizon: 2027–2028. Two years for gains already observed in some assisted tasks to spread through workflow redesign. The forecast is about suitable workflows, with hard-to-verify work expected to lag; it is not a uniform productivity percentage. ### The strongest rival - Even in the defined target workflows, selection and omitted review costs explain most apparent gains; lower effort does not survive a complete comparison. - Models become better at checking their own output, reducing the review burden even on complex work. ### What has to hold - Before deployment, the target work is repeatable and digital, has a clear acceptance test, and can be checked affordably. Quality standards stay fixed and delayed defects count. - Teams can select or redesign tasks instead of forcing every task through the same AI process. ### What would change our mind - Comparable later studies measure work completed to the required standard, total human effort and defects after handoff, with credible treatment of selection. - The gain disappears when review and downstream repair are counted, or persists when they are included. ### Evidence - metr-productivity-2025#setting [context] — https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ - metr-productivity-2025#completion-time [context] — https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ - metr-productivity-update-2026#selection [context] — https://metr.org/blog/2026-02-24-uplift-update/ - metr-worker-survey-2026#self-report [context] — https://metr.org/blog/2026-05-11-ai-usage-survey/ - anthropic-workforce-2026#anth-bottleneck [context] — https://www.anthropic.com/research/employment-at-anthropic - brynjolfsson-ai-work-v2#resolved-work [supports] — https://arxiv.org/abs/2304.11771v2 - metr-productivity-update-2026#direction-of-selection [supports] — https://metr.org/blog/2026-02-24-uplift-update/ - Position: metr-productivity-researchers - Position: metr-followup-method - Position: metr-worker-value - Position: customer-support-researchers - Connection: quality-to-reliability - Connection: quality-to-returns ### Named research gaps - accepted-work-transfer: Compare total effort at the same required quality in workflows selected before deployment, including later checking and repair. broader: ai-and-work — Checking and correction are part of dependability. related: ai-delegation-reliability — Low average rework does not establish safe unsupervised operation. related: ai-deployment-returns — Net workflow value is one input to sustainable returns. related: ai-entry-and-worker-gains — Review and training work may be redistributed across experience levels. ### Where this could lead: Checkable digital work takes less effort to finish Path ID: quality-useful-work-expands Teams lower total effort in repeatable digital workflows with clear acceptance tests and affordable verification. They define those conditions before deployment, then count the work all the way through handoff and repair. Weight: Base case Horizon: 2027–2028. Existing assisted-work gains and improving developer experience favor wider gains over two years, with task selection and redesign driving the spread. - Conditions: The target workflow is repeatable and digital, with a clear acceptance test and affordable verification established before the comparison. - Conditions: They count verification and later defects, then concentrate use where the full result improves. - Watch for: Comparable task cohorts show lower total effort at stable quality. - Watch for: Benefits persist after the novelty of the rollout and include downstream repair. - anthropic-workforce-2026#anth-bottleneck [context] — https://www.anthropic.com/research/employment-at-anthropic - metr-worker-survey-2026#self-report [context] — https://metr.org/blog/2026-05-11-ai-usage-survey/ - metr-productivity-2025#setting [challenges] — https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ - metr-productivity-2025#completion-time [challenges] — https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ - brynjolfsson-ai-work-v2#resolved-work [supports] — https://arxiv.org/abs/2304.11771v2 - metr-productivity-update-2026#direction-of-selection [supports] — https://metr.org/blog/2026-02-24-uplift-update/ - Position: anthropic-work-organization - Position: metr-worker-value - Position: customer-support-researchers ### Where this could lead: Even promising workflows take more effort than they save Path ID: quality-review-absorbs-gains The gain disappears even in repeatable digital work selected for clear outcomes and affordable checking. Finding hidden mistakes and fitting the output into the larger job takes more time and effort than expected; faster drafts fail to make finished work easier. Weight: Alternative Horizon: 2027–2028. This remains a live alternative throughout the adoption window, especially where correctness is expensive to establish. - Conditions: Verification and repair cost more than the team expected before rollout. - Conditions: Selection of enthusiastic users or favorable tasks explains apparent gains that do not survive comparable full-work measurements. - Watch for: Blind quality review finds more defects in apparently faster work. - Watch for: Time saved at generation returns as integration, correction or customer support. - metr-productivity-2025#setting [supports] — https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ - metr-productivity-2025#completion-time [supports] — https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ - metr-productivity-update-2026#selection [limits] — https://metr.org/blog/2026-02-24-uplift-update/ - Position: metr-productivity-researchers - Position: metr-followup-method ## Who gets an entry point and shares the gains? Question ID: ai-entry-and-worker-gains Question scope: Who gets an entry point, and who shares the gains? ### Background on this question The first rung of a career is also where people learn to take responsibility. AI could help a newcomer do valuable work sooner, yet an employer might use the same capability to offer fewer first jobs. That tension matters to students, career changers, managers and anyone who remembers needing someone to give them a chance. The question is who gets access to the tools, paid experience and progression that turn greater capability into a life with more options. Our read: AI can help a junior do the job—and remove the opening ### Why we take this view We expect entry into AI-exposed digital occupations to stay difficult through 2027, even as AI makes some novices more productive. Stanford’s August 2026 payroll study finds young workers in AI-exposed occupations falling behind less-exposed peers, mainly because fewer are hired; Anthropic’s workforce report describes a declining junior share of its own technical hires. Yet in the customer-support study, less experienced agents gained more from AI assistance. Those findings can coexist: a tool can help a novice already in the job while an employer decides to recruit fewer beginners. The consequence is a harder route to the experience that future skilled work requires. What this means: Employers deciding what to automate are also deciding where experience comes from. Preserve paid work that teaches judgment; job seekers should show they can own an outcome, including checking and improving AI work. EM outlook: editorial-outlook. Horizon: 2027–2027. The hiring pattern has persisted into mid-2026 and Anthropic reports a similar shift in October. Another hiring cycle is more likely to extend that pressure than reverse it without deliberate new entry routes. ### The strongest rival - AI lowers the cost of training and makes novices valuable sooner, encouraging employers to expand entry hiring. - Education mix, interest rates and earlier labor-market trends explain more of the apparent AI pattern than automation does. ### What has to hold - The outlook is strongest for exposed digital roles, not every young worker or occupation. - Employers continue to value independent judgment while automating some tasks traditionally assigned to novices. ### What would change our mind - Entry hiring recovers in exposed occupations while AI use continues to deepen. - Employers report and demonstrate new paid learning routes with progression into responsibility. - The employment gap shrinks materially after accounting for education, prior trends and broader national data. ### Evidence - anthropic-workforce-2026#anth-junior [context] — https://www.anthropic.com/research/employment-at-anthropic - nyfed-business-survey-2026#nyf-fewer [context] — https://libertystreeteconomics.newyorkfed.org/2026/09/businesses-are-using-ai-to-transform-work-not-cut-jobs/ - nyfed-business-survey-2026#nyf-more [context] — https://libertystreeteconomics.newyorkfed.org/2026/09/businesses-are-using-ai-to-transform-work-not-cut-jobs/ - steering-beyond-automation#un-voice [context] — https://www.normaltech.ai/p/steering-ai-beyond-automation - stanford-canaries-2026#hiring-first [supports] — https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/ - stanford-canaries-2026#descriptive [supports] — https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/ - brynjolfsson-ai-work-v2#novice-gains [challenges] — https://arxiv.org/abs/2304.11771v2 - Position: unger-narayanan-collaboration - Position: stanford-entry-researchers - Position: customer-support-researchers ### Named research gaps - entry-progression-and-distribution: Test the hiring interpretation against education, prior trends and national data, and follow actual paid entry and progression routes. broader: jobs — Track distribution and opportunity alongside total employment. related: ai-work-quality-and-rework — Who performs review can affect skill development and the distribution of work. ### Where this could lead: Fewer first jobs, even when the work gets easier Path ID: entry-routes-narrow Employers ask smaller experienced teams to deliver more. Novices who get access may learn faster, but fewer people receive paid entry into the occupation. Weight: Base case Horizon: 2027–2027. The existing hiring divergence and Anthropic’s reported change in junior hiring make continued pressure through the next hiring year more plausible than a rapid reversal. - Conditions: Employers automate routine implementation before they replace the training it once provided. - Conditions: Higher output per team does not create enough new demand to restore entry hiring. - Watch for: Junior openings stay weak relative to senior openings in exposed occupations. - Watch for: Employers raise expectations for independent ownership while reducing paid learning roles. - anthropic-workforce-2026#anth-junior [context] — https://www.anthropic.com/research/employment-at-anthropic - nyfed-business-survey-2026#nyf-fewer [context] — https://libertystreeteconomics.newyorkfed.org/2026/09/businesses-are-using-ai-to-transform-work-not-cut-jobs/ - nyfed-business-survey-2026#nyf-more [context] — https://libertystreeteconomics.newyorkfed.org/2026/09/businesses-are-using-ai-to-transform-work-not-cut-jobs/ - stanford-canaries-2026#hiring-first [supports] — https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/ - brynjolfsson-ai-work-v2#novice-gains [context] — https://arxiv.org/abs/2304.11771v2 - Position: anthropic-workforce-authors - Position: stanford-entry-researchers - Position: customer-support-researchers ### Where this could lead: AI becomes an apprenticeship multiplier Path ID: entry-opportunities-renew Employers use assistance to make newcomers productive sooner and give them progressively harder work. New entry routes grow because teaching and supervision become worth the investment. Weight: Alternative Horizon: 2027–2029. Assisted novice gains already exist in one studied setting; expanding paid entry routes requires employers to turn that capability into hiring and progression over several cohorts. - Conditions: The employer rewards learning and creates responsibility rather than using every gain to reduce hiring. - Conditions: Workers can question the tool, receive feedback and build expertise beyond accepting its output. - Conditions: Employers and educators build paid apprenticeships around real outcomes, with experienced workers given time and credit for teaching. - Watch for: Firms add entry roles and show progression into more responsible work. - Watch for: Novices retain useful skills when the tool is unavailable and share in improved pay or opportunity. - steering-beyond-automation#un-voice [context] — https://www.normaltech.ai/p/steering-ai-beyond-automation - anthropic-workforce-2026#anth-junior [challenges] — https://www.anthropic.com/research/employment-at-anthropic - brynjolfsson-ai-work-v2#novice-gains [supports] — https://arxiv.org/abs/2304.11771v2 - Position: unger-narayanan-collaboration - Position: customer-support-researchers ## Is AGI a few years away? Question ID: agi Question scope: AGI within a few years: progress, limits, timelines or forecasts for broadly capable systems, including disagreement. ### Background on this question AGI can stand for very different hopes and fears: a partner in discovery, a rival to human expertise, or a change large enough to unsettle plans for work and life. The useful question is which capabilities arrive, how reliably they work and what people can do with them. This read uses Amodei’s demanding combination of broad expertise and sustained autonomous work. You can prepare for powerful changes without making every decision depend on a single arrival date. Our read: Expect powerful specialists before a general replacement ### Why we take this view Our base case through 2028 is major gains in autonomous digital work, while AI still falls short of Amodei’s powerful-AI vision: exceptional capability across many fields combined with the ability to work independently on tasks lasting days or weeks, the demanding end of the hours, days or weeks his definition allows. METR’s time-horizon methodology documents progress on software tasks that take human experts increasingly long to complete. That gives the fast-arrival case substance. We favor a slower arrival of broad reliability because completing longer tasks with clear success tests does not establish that a system can handle unfamiliar work across fields, recognize its mistakes and recover without a person rescuing it. Those are the abilities needed to entrust it with a much wider responsibility. What this means: Choose a problem worth solving before choosing an AGI date. Researchers, founders and public institutions can use increasingly capable systems to attempt work beyond their current reach, while keeping commitments adaptable as capability and reliability change. EM outlook: editorial-outlook. Horizon: 2027–2028. An EM judgment through the end of 2028, not a benchmark extrapolation. The defined endpoint combines broad expert capability with autonomous work over days or weeks; the reviewed evidence is much stronger on bounded software progress than on that conjunction. ### The strongest rival - Improved models and agent systems transfer across domains fast enough to deliver the defined capabilities by the end of 2028. - Progress concentrates in useful specialization for longer than expected, with limited transfer even as software-task scores climb. ### What has to hold - The endpoint is Amodei’s demanding combination of breadth and autonomy, not whichever benchmark first receives an AGI label. - Gains on self-contained software tasks do not automatically transfer to unfamiliar domains with ambiguous goals and costly errors. ### What would change our mind - Independent evaluations demonstrate multi-day work across scientific, commercial and technical domains, including recovery from novel failures. - Successful external deployments reproduce those results at usable cost without hidden human rescue. - Capability progress stalls across successive systems on demanding tasks outside familiar evaluation suites. ### Evidence - amodei-loving-grace-2024#powerful-ai-timing [context] — https://darioamodei.com/essay/machines-of-loving-grace - metr-time-horizon-1-1#historical-growth [supports] — https://metr.org/blog/2026-01-29-time-horizon-1-1/ - metr-time-horizon-dashboard#reliability-threshold [limits] — https://metr.org/time-horizons/ - metr-time-horizon-dashboard#measurement-edge [limits] — https://metr.org/time-horizons/ - Position: amodei-powerful-ai - Position: metr-horizon-research ### Named research gaps - agi-claims-and-endpoints: Test the EM outlook against current cross-domain performance, long autonomy and rival timeline arguments; software-task progress alone cannot resolve that endpoint. narrower: ai-delegation-reliability — Test one practical dependency without substituting it for the whole AGI question. related: jobs — Capability does not determine deployment, demand or employment on its own. ### Where this could lead: Broad capability and long autonomy arrive together Path ID: agi-near-capability Broadly capable systems let smaller teams attempt sustained scientific, commercial and technical projects that once required much larger organizations. The opportunity depends on people choosing valuable problems and giving the work a route into the world. Weight: Alternative Horizon: 2027–2028. The rival case is that breadth and days-to-weeks autonomy arrive by the end of 2028, within the base case’s own window. Software progress motivates taking that possibility seriously; transfer across domains would have to advance faster than EM currently expects. - Conditions: Long-task progress transfers into unfamiliar scientific, commercial and technical work. - Conditions: Systems can plan, verify, recover and keep context across days or weeks at usable cost. - Watch for: Independent tests reproduce broad expert performance on new multi-day projects. - Watch for: External organizations use the capability repeatedly without hidden human completion. - amodei-loving-grace-2024#powerful-ai-timing [context] — https://darioamodei.com/essay/machines-of-loving-grace - metr-time-horizon-1-1#historical-growth [context] — https://metr.org/blog/2026-01-29-time-horizon-1-1/ - metr-time-horizon-dashboard#reliability-threshold [limits] — https://metr.org/time-horizons/ - Position: amodei-powerful-ai - Position: metr-horizon-research ### Where this could lead: Powerful specialists arrive; general autonomy still falls short Path ID: agi-delayed-or-different AI reshapes valuable work while the demanding combination of broad expertise and reliable days-to-weeks autonomy remains unmet. Organizations gain new capabilities without obtaining a general substitute for expertise. Weight: Base case Horizon: 2027–2028. Through 2028, we expect uneven transfer and reliability to remain the binding obstacles, even as bounded digital capabilities improve substantially. - Conditions: Some domains still require context, feedback or judgment that current task evaluations do not capture. - Conditions: Progress on a longer software task does not resolve broad autonomy and recovery together. - Watch for: Gains remain concentrated in well-specified tasks or familiar domains. - Watch for: Broad multi-day evaluations expose failures that shorter demonstrations and human-assisted results conceal. - amodei-loving-grace-2024#powerful-ai-timing [context] — https://darioamodei.com/essay/machines-of-loving-grace - metr-time-horizon-dashboard#reliability-threshold [supports] — https://metr.org/time-horizons/ - metr-time-horizon-dashboard#measurement-edge [supports] — https://metr.org/time-horizons/ - anthropic-workforce-2026#anth-bottleneck [context] — https://www.anthropic.com/research/employment-at-anthropic - Position: amodei-powerful-ai - Position: metr-horizon-research ## When can work be delegated without checking every step? Question ID: ai-delegation-reliability Question scope: When can work be delegated without checking every step? ### Background on this question Delegation changes your relationship with a tool: you stop following every step and become responsible for trusting the result. Done well, that can free attention for people, decisions or projects that need it. Done badly, it creates a new worry about what happened while you were looking elsewhere. This question asks when a system can complete a defined task, recognize trouble and recover or ask for help at a cost you can live with. Our read: Give agents whole tasks before giving them whole jobs ### Why we take this view During 2027–2028, we expect people to delegate more whole tasks to agents where they can check the result and recover from a failure. METR’s time-horizon methodology documents progress on software tasks that require more human expert time, supporting the prospect of longer assignments. But the 50%-time horizon reported on METR’s time-horizon dashboard is the task length, in human expert time, at which a model is predicted to succeed only half the time; it does not establish dependable completion. A team can make use of an agent’s capability by testing its output before relying on it and retaining a way to intervene. Giving it responsibility for an entire job requires reliability across much less predictable demands. What this means: Move from checking each step to checking outcomes where you can detect a failure and recover. For harder work, preserve checkpoints until actual completion records justify removing them. EM outlook: editorial-outlook. Horizon: 2027–2028. This predicts wider dependable delegation of bounded tasks over the next two years. Clear success tests and recovery can turn existing capabilities into a usable service sooner than broad, unsupervised responsibility. ### The strongest rival - Longer tasks multiply hidden failures; humans remain closely involved even as agents produce more. - Agent systems learn to diagnose and recover across workflows, allowing whole-job responsibility to grow faster than this forecast. ### What has to hold - The task has a clear success condition and failures can be detected before their cost becomes unacceptable. - Verification and recovery improve along with model capability. ### What would change our mind - In workflows selected for clear outcomes and recoverable failures, repeated deployments still require frequent rescue or show persistent hidden defects despite better benchmark scores. - Higher success thresholds improve alongside task length; gains survive when tasks are unfamiliar and consequences are real. ### Evidence - anthropic-workforce-2026#anth-bottleneck [context] — https://www.anthropic.com/research/employment-at-anthropic - amodei-loving-grace-2024#powerful-ai-timing [context] — https://darioamodei.com/essay/machines-of-loving-grace - metr-productivity-update-2026#selection [limits] — https://metr.org/blog/2026-02-24-uplift-update/ - metr-time-horizon-1-1#historical-growth [supports] — https://metr.org/blog/2026-01-29-time-horizon-1-1/ - metr-time-horizon-dashboard#reliability-threshold [limits] — https://metr.org/time-horizons/ - metr-time-horizon-dashboard#measurement-edge [limits] — https://metr.org/time-horizons/ - Position: amodei-powerful-ai - Position: metr-followup-method - Position: metr-horizon-research - Connection: quality-to-reliability ### Named research gaps - delegation-failure-and-recovery: Measure completion, hidden failure, human intervention and recovery in real deployments as well as benchmarks. broader: agi — Reliability is one relevant condition, not a sufficient definition of general intelligence. broader: ai-and-work — A dependable colleague can also be a supervised collaborator. related: ai-work-quality-and-rework — Inspect the cost of detecting and correcting failures. ### Where this could lead: Whole tasks move off the human’s desk Path ID: delegation-bounded-progress Agents complete specified work with outcome checks and exception handling. People spend less time approving steps, while retaining responsibility for goals and consequential decisions. Weight: Base case Horizon: 2027–2028. Verification and recovery around bounded tasks can make existing and improving capabilities usable before broad autonomous judgment arrives. - Conditions: Success can be tested and a failed run can be retried, reversed or escalated affordably. - Conditions: Context and permissions match the task instead of requiring continual human clarification. - Watch for: The share of tasks finished to the same quality standard without human rescue rises. - Watch for: Failure and recovery costs fall across repeated real deployments, not just benchmark runs. - amodei-loving-grace-2024#powerful-ai-timing [context] — https://darioamodei.com/essay/machines-of-loving-grace - anthropic-workforce-2026#anth-bottleneck [context] — https://www.anthropic.com/research/employment-at-anthropic - metr-time-horizon-1-1#historical-growth [supports] — https://metr.org/blog/2026-01-29-time-horizon-1-1/ - metr-time-horizon-dashboard#reliability-threshold [limits] — https://metr.org/time-horizons/ - Position: amodei-powerful-ai - Position: metr-horizon-research ### Where this could lead: Better agents still need a nearby human Path ID: delegation-supervision-persists Capability improves, but long tasks accumulate mistakes and missing context. Businesses keep people close enough to catch problems before they become expensive. Weight: Alternative Horizon: 2027–2028. Over the same adoption period, reliability can lag capability where outcomes are ambiguous or recovery is difficult. - Conditions: High-consequence failures remain difficult to detect before acting. - Conditions: Independent task success does not transfer reliably across unfamiliar conditions. - Watch for: Task length improves at moderate success rates while dependable completion lags. - Watch for: Deployments continue to need frequent interventions or hide substantial human rescue. - anthropic-workforce-2026#anth-bottleneck [context] — https://www.anthropic.com/research/employment-at-anthropic - metr-productivity-update-2026#selection [limits] — https://metr.org/blog/2026-02-24-uplift-update/ - metr-time-horizon-dashboard#reliability-threshold [supports] — https://metr.org/time-horizons/ - metr-time-horizon-dashboard#measurement-edge [supports] — https://metr.org/time-horizons/ - Position: anthropic-work-organization - Position: metr-followup-method - Position: metr-horizon-research ## Is there an AI bubble? Question ID: bubble Question scope: An AI bubble: financing, valuations, capital expenditure, compute investment, revenues, demand or sustainable returns. ### Background on this question A useful technology can inspire justified excitement and still attract investments that fail to pay off. That distinction matters whether you are buying a tool, building a business or trying to make sense of the scale of the AI buildout. Paying customers, a profitable supplier and an attractive investment are different achievements. The question is how much value reaches each participant after the costs of delivering the service, financing it and keeping the infrastructure useful. Our read: Even with growing sales, AI suppliers will struggle to earn back the cost of expansion ### Why we take this view For 2027–2029, we expect AI suppliers expanding expensive computing infrastructure to remain under pressure to turn growing sales into cash left after operating costs and investment. Microsoft’s FY2026 results show substantial paid Microsoft 365 Copilot adoption alongside rising company-wide spending on property and equipment. Customers are paying, but that does not establish how much of the spending AI revenue can repay. We favor continued pressure over a clean payoff because suppliers must keep investing to compete, while competition can limit what they charge. Customers can therefore gain useful services even as the companies providing them struggle to earn enough from that demand to justify their expansion. What this means: Separate three decisions: whether the tool improves your work, whether the supplier has sound economics, and whether an investment price leaves room for a return. One successful answer does not settle the other two. EM outlook: editorial-outlook. Horizon: 2027–2029. A three-year outlook for uneven business and investment outcomes as 2026 spending meets customer renewals and operating costs. It is not a dated crash call: the reviewed records do not establish a financing trigger or an industry-wide valuation. ### The strongest rival - Rapid usage growth and falling unit costs support margins strongly enough for cash flows to catch up with the buildout. - Disappointing demand combines with refinancing pressure or stranded capacity to produce a broader investment contraction. ### What has to hold - Competition, operating costs and reinvestment prevent every supplier from capturing all the value created for users. - Microsoft demonstrates substantial purchasing at one large vendor; it is not representative of every lab, infrastructure project or application business. ### What would change our mind - AI-specific cash generation improves relative to capital commitments across multiple capital-heavy suppliers, while renewals remain strong: that would favor the payoff path. - Renewals, utilization or customer concentration deteriorate while fixed commitments remain high. - Original financing records reveal maturities or cash shortfalls that turn a general downside into a dated stress case. ### Evidence - zitron-ai-financing-2026#financing-thesis [context] — https://www.wheresyoured.at/ai-had-better-work-out/ - damodaran-ai-business-2026#valuation-conditions [context] — https://aswathdamodaran.blogspot.com/2026/08/ais-bar-mitzvah-moment-from-hype-hope.html - microsoft-fy2026-results#paid-demand [supports] — https://www.microsoft.com/en-us/investor/earnings/fy-2026-q4/press-release-webcast - microsoft-fy2026-results#capital-spending [supports] — https://www.microsoft.com/en-us/investor/earnings/fy-2026-q4/press-release-webcast - Position: zitron-bubble-financing - Position: damodaran-ai-value-conditions - Position: microsoft-paid-demand ### Named research gaps - bubble-investment-and-demand: Separate AI-specific cash generation, capital commitments and financing from whole-company totals; determine whether pressure eases or becomes a dated financing problem. narrower: ai-deployment-returns — Begin with value to a deploying organization, then identify the additional evidence needed for investor returns. related: jobs — Employment effects and investment outcomes can move differently. ### Where this could lead: Revenue and efficiency catch up with the buildout Path ID: bubble-cashflows-catch-up Customers use and renew AI services at growing scale. Better infrastructure utilization and falling delivery costs allow more suppliers to turn a useful technology into durable profits. Weight: Alternative Horizon: 2027–2029. This is the stronger-return route over the next three years: growing purchases must become recurring cash generation as the current investment enters service. - Conditions: Paid demand grows beyond a few concentrated customers and survives renewals. - Conditions: Revenue and utilization improve fast enough to cover operating costs and reinvestment. - Watch for: AI-specific gross profit improves alongside usage, after accounting for incentives and subsidized demand. - Watch for: Cash generation strengthens relative to capital commitments across more than one supplier. - zitron-ai-financing-2026#financing-thesis [context] — https://www.wheresyoured.at/ai-had-better-work-out/ - damodaran-ai-business-2026#valuation-conditions [context] — https://aswathdamodaran.blogspot.com/2026/08/ais-bar-mitzvah-moment-from-hype-hope.html - microsoft-fy2026-results#paid-demand [context] — https://www.microsoft.com/en-us/investor/earnings/fy-2026-q4/press-release-webcast - microsoft-fy2026-results#capital-spending [context] — https://www.microsoft.com/en-us/investor/earnings/fy-2026-q4/press-release-webcast - Position: zitron-bubble-financing - Position: damodaran-ai-value-conditions - Position: microsoft-paid-demand ### Where this could lead: Usage grows faster than the return on the buildout Path ID: bubble-returns-disappoint Capital-heavy AI suppliers keep facing pressure on cash returns as competition and reinvestment absorb growing customer demand. Useful products spread while the capital committed to providing them struggles to earn its expected return. Weight: Base case Horizon: 2027–2029. We expect uneven returns during the next three years as the buildout encounters recurring customer economics. A broad crash would require additional evidence about funding, utilization and forced selling. - Conditions: Pricing pressure and continued reinvestment absorb much of the extra gross profit from growing demand. - Conditions: Capital commitments remain costly to reduce when customer economics disappoint. - Watch for: At capital-heavy suppliers, AI-related usage and revenue grow while cash generation after reinvestment fails to improve relative to capital committed. - Watch for: Companies write down capacity, renegotiate commitments or need new financing without a matching improvement in customer economics. - zitron-ai-financing-2026#financing-thesis [context] — https://www.wheresyoured.at/ai-had-better-work-out/ - damodaran-ai-business-2026#valuation-conditions [context] — https://aswathdamodaran.blogspot.com/2026/08/ais-bar-mitzvah-moment-from-hype-hope.html - microsoft-fy2026-results#paid-demand [context] — https://www.microsoft.com/en-us/investor/earnings/fy-2026-q4/press-release-webcast - microsoft-fy2026-results#capital-spending [context] — https://www.microsoft.com/en-us/investor/earnings/fy-2026-q4/press-release-webcast - Position: zitron-bubble-financing - Position: damodaran-ai-value-conditions - Position: microsoft-paid-demand ## When is an AI deployment worth its total cost? Question ID: ai-deployment-returns Question scope: When does a useful AI deployment become worth its total cost? ### Background on this question An AI tool may feel useful before its place in a team makes sense. Someone still has to connect it to the work, teach people to use it, check the results and deal with mistakes. The cost includes time, attention and effort as well as money. The reward might be better service, new capacity or a project that was previously out of reach. This question helps distinguish those lasting benefits from a promising demonstration or a workload quietly passed to someone else. Our read: The best returns will come from changing the workflow ### Why we take this view Through 2027, we expect organizations to get more value sooner by redesigning a specific workflow around AI than by distributing tools widely and leaving teams to work out how to use them. The customer-support study illustrates the opportunity: agents resolved more issues per hour, with larger gains among less experienced workers. Anthropic’s workforce report identifies a competing demand: people must still devote attention to the growing output agents produce. A deployment creates value when the organization can use the additional completed work or better service, and those benefits outweigh the time, effort and money spent integrating the tool, training people and checking its results. A redesigned workflow meets those conditions sooner because the organization has already decided what the extra work is for, who checks it and who benefits; broad access leaves each team to discover that separately. What this means: Choose a recurring piece of work that someone needs done, record how well it is done today, and decide what improvement would be worth paying for. Count quality, new capacity and service improvements as well as hours saved; count integration, review and ongoing operating costs against them. EM outlook: editorial-outlook. Horizon: 2027–2027. The next operating year should favor deployments built around measurable existing work. Broad rollout requires more coordination and training before benefits reach the same decision-maker who bears the cost. ### The strongest rival - Broad access produces valuable experimentation and new capabilities that narrowly specified projects would miss. - Implementation and review costs absorb the operating gains, leaving even well-liked tools below their full cost. ### What has to hold - The organization uses the extra capacity or better result; output with no demand has little economic value. - The comparison includes rollout costs, repeat operating costs and work shifted to another team. ### What would change our mind - Repeat-use evidence shows better completed work after counting the time, effort and money spent on integration, checking, training and operation. - The apparent gain disappears when transferred work, delayed defects or unused capacity are counted. ### Evidence - metr-worker-survey-2026#self-report [context] — https://metr.org/blog/2026-05-11-ai-usage-survey/ - anthropic-workforce-2026#anth-bottleneck [context] — https://www.anthropic.com/research/employment-at-anthropic - damodaran-ai-business-2026#valuation-conditions [context] — https://aswathdamodaran.blogspot.com/2026/08/ais-bar-mitzvah-moment-from-hype-hope.html - brynjolfsson-ai-work-v2#resolved-work [supports] — https://arxiv.org/abs/2304.11771v2 - brynjolfsson-ai-work-v2#novice-gains [supports] — https://arxiv.org/abs/2304.11771v2 - Position: metr-worker-value - Position: damodaran-ai-value-conditions - Position: customer-support-researchers - Connection: quality-to-returns ### Named research gaps - deployment-value-total-cost: Test the full recurring cost in time, effort and money against completed work and service quality, beyond task productivity and self-reported value. broader: bubble — Customer value is one dependency; supplier margins, capital cost and valuation need separate evidence. related: ai-work-quality-and-rework — Review and correction can affect total cost and value. related: jobs — A positive deployment return does not specify whether employment expands or contracts. ### Where this could lead: A focused workflow pays for its AI Path ID: deployment-worth-cost The buyer gets a recurring improvement worth more than implementation and operating costs: faster service, better quality or capacity that would otherwise be unavailable. Weight: Base case Horizon: 2027–2027. We expect the clearest near-term returns where a deployment improves work with an existing buyer and an observable outcome. - Conditions: The improved result has actual demand or avoids an actual cost. - Conditions: Savings or new value survive integration, review, training and repeat operation. - Conditions: The buyer chooses what the extra capacity is for: better service, a previously neglected need, or work the team could not afford to attempt. - Watch for: The buyer keeps funding the workflow after measuring its full costs. - Watch for: The quality or amount of completed work improves over repeated use without shifting extra effort to another team. - metr-worker-survey-2026#self-report [context] — https://metr.org/blog/2026-05-11-ai-usage-survey/ - anthropic-workforce-2026#anth-bottleneck [context] — https://www.anthropic.com/research/employment-at-anthropic - brynjolfsson-ai-work-v2#resolved-work [supports] — https://arxiv.org/abs/2304.11771v2 - Position: metr-worker-value - Position: customer-support-researchers ### Where this could lead: The tool is popular; the deployment fails to pay Path ID: deployment-value-erodes People enjoy the assistance, but the organization cannot turn it into enough useful extra work or better service. Integration and review absorb the benefits, and broad rollout loses its case. Weight: Alternative Horizon: 2027–2028. The first renewal and operating cycles can expose costs that a demonstration or initial enthusiasm misses; persistence into 2028 would weaken the focused-deployment thesis. - Conditions: Output grows faster than demand for it, or the gain does not reach the budget that pays for the system. - Conditions: Integration, quality loss or ongoing supervision costs more than the usable benefit. - Watch for: Usage remains high while completed outcomes and customer value stay flat. - Watch for: Renewals shrink after the buyer includes costs previously omitted from the business case. - metr-worker-survey-2026#self-report [limits] — https://metr.org/blog/2026-05-11-ai-usage-survey/ - metr-productivity-update-2026#selection [limits] — https://metr.org/blog/2026-02-24-uplift-update/ - metr-productivity-2025#setting [context] — https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ - metr-productivity-2025#completion-time [context] — https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ - Position: metr-worker-value ## Sources and provenance ### How employment is changing at Anthropic Source ID: anthropic-workforce-2026 https://www.anthropic.com/research/employment-at-anthropic Citation: Anthropic workforce report · October 2026 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. Report aliases: Anthropic’s workforce report; Anthropic workforce report; Anthropic’s hiring mix Preview passages: anth-junior, anth-bottleneck Publication: 2026-10-16. Basis: Publisher publication date, checked on original page. Access: public-page. Original publisher page and exact retained source reviewed. Source retention does not establish every downstream reuse or training right. Checked: 2026-10-29. Source state: stream-snapshot. Exact retained Stream work/snapshot and SHA256 verified against the 29 October SELECT-only source packet. That lookup found no joined public Stream or bridge reading revision; no reading identity is invented. Work: 18f3c3ca31491462a8ab52f797ebadd2a7087803317e69a5152e2a0c4cfbed65. Snapshot: a633ce13b2d661ad9c008ac99d3357f15a80c4bcd91009290d11712c2f565793. Text hash: 8ea3533aa08b1f63c143b0a383911da93e88f464edd2d8e61edebf2a97a42232. anthropic-workforce-2026#anth-junior — Retained extracted text, character offsets \[4420, 4494). Context (EM paraphrase): Anthropic’s October 2026 report describes its own workforce; junior employees’ share fell from early 2025, and technical hiring still grew in absolute terms. This is a changing hiring mix, not a count of jobs lost. The proportion of junior technical hires fell even faster, from 27% to 9%. anthropic-workforce-2026#anth-bottleneck — Retained extracted text, character offsets \[767, 839). Context (EM paraphrase): The authors report rising internal output alongside agent use and interpret human attention as a constraint. Adoption is not randomized; the report does not establish a general causal productivity effect. This acceleration makes human attention the bottleneck for agent output. ### Businesses Are Using AI to Transform Work, Not Cut Jobs Source ID: nyfed-business-survey-2026 https://libertystreeteconomics.newyorkfed.org/2026/09/businesses-are-using-ai-to-transform-work-not-cut-jobs/ Citation: New York Fed regional business survey · September 2026 A report on the August 2026 survey of businesses in New York and northern New Jersey, distinguishing firm adoption, worker usage, hiring and layoffs. Report aliases: New York Fed regional business survey; New York Fed’s regional survey; regional employer survey Preview passages: nyf-fewer, nyf-more, worker-adoption, reported-layoffs Publication: 2026-09-01. Basis: Publisher publication date, checked on original page. Access: public-page. Original publisher page and exact retained source reviewed. Source retention does not establish every downstream reuse or training right. Checked: 2026-10-29. Source state: stream-snapshot. Exact retained Stream work/snapshot and SHA256 verified against the 29 October SELECT-only source packet. That lookup found no joined public Stream or bridge reading revision; no reading identity is invented. Work: c16ff26252860625027590ed41b03b3773be9d5b324475687ef0304e51a73af8. Snapshot: d59dc274ed0f66564545055ec05e4a9abf5a5314f05623d021bbe3ac1b2274ab. Text hash: 7ec36434625838d3aa01b2819ecb8d90e622fd2f2e8a664dd2c656116c4829dd. nyfed-business-survey-2026#nyf-fewer — Retained extracted text, character offsets \[4475, 4494). Context (EM paraphrase): August 2026 regional survey: about 15% of AI-using service firms reported hiring fewer workers than they would otherwise have because of AI. This is employers’ stated counterfactual, not independently identified job loss. The paired figure for hiring more is about 13%; firm percentages cannot be subtracted into a net job count. hired fewer workers nyfed-business-survey-2026#nyf-more — Retained extracted text, character offsets \[4776, 4852). Context (EM paraphrase): August 2026 regional business survey, reporting workforce adjustments over the preceding six months. The workforce charts concern AI-using firms in the survey; they are not national employment totals. Authors also report retraining as a common response. About 13 percent of service firms said they had hired more workers due to AI nyfed-business-survey-2026#worker-adoption — Retained extracted text, character offsets \[2997, 3007). Context (EM paraphrase): Among AI-using regional service firms, median worker usage in the August 2026 survey was 17%; manufacturers were at 7%. Firm adoption and workforce penetration are different measures. 17 percent nyfed-business-survey-2026#reported-layoffs — Retained extracted text, character offsets \[4214, 4223). Context (EM paraphrase): AI-using service firms reporting AI-related layoffs over the preceding six months, versus about 15% reporting reduced hiring. Regional employer reports, not national worker shares. 4 percent ### Steering AI Beyond Automation Source ID: steering-beyond-automation https://www.normaltech.ai/p/steering-ai-beyond-automation Citation: Unger and Narayanan · Steering AI Beyond Automation An argument for steering AI toward collaboration and giving workers more influence over deployment. Its proposed direction is distinct from measured effects. Report aliases: Steering AI Beyond Automation; Unger and Narayanan’s argument Preview passages: un-voice Publication: 2026-10-14. Basis: Publisher publication date, checked on original page. Access: public-page. Original publisher page and exact retained source reviewed. Source retention does not establish every downstream reuse or training right. Checked: 2026-10-29. Source state: stream-snapshot. Exact retained Stream work/snapshot and SHA256 verified against the 29 October SELECT-only source packet. That lookup found no joined public Stream or bridge reading revision; no reading identity is invented. Work: 5a4e1d4b2fb65d2a40b74bca91665b9dc8321e961ea8e51dea6a078d98139bff. Snapshot: 11ae3a79455a2c294b795984f741f4bb1a3df2cbe7f39d2e0867564a93f6a8a2. Text hash: f2b6cc0e1c8b6bbfe0e9efb6e94a5614b87630de3934d9902943551c31faf389. steering-beyond-automation#un-voice — Retained extracted text, character offsets \[32549, 32678). Context (EM paraphrase): Retained extracted text, character offsets \[32549, 32678); verify against the cited snapshot, not current HTML offsets. Giving workers more of a voice in bargaining over AI deployment is another way to steer toward augmentation without mandating it. ### Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity Source ID: metr-productivity-2025 https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ Citation: METR developer experiment · July 2025 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. Report aliases: METR’s early-2025 developer experiment; METR’s 2025 developer study; early-2025 METR experiment Preview passages: setting, completion-time Publication: 2025-07-10. Basis: Publication date displayed by the original publisher. Access: public-page. Original publisher page and exact retained source reviewed. Source retention does not establish every downstream reuse or training right. Checked: 2026-10-29. Source state: stream-snapshot. Exact retained Stream work/snapshot and SHA256 verified against the 29 October SELECT-only source packet. That lookup found no joined public Stream or bridge reading revision; no reading identity is invented. Work: eeef7ca24568c0fe592aa98201faccea099ec6507dbd9476c8bea56310f28743. Snapshot: 1b377f2c5f76b4960227c32060fef0433519ce76239947c7e836402cb9a945ac. Text hash: 78a7743c1dda83f49b4d646c134243e5e6c38bed0eb7aa33d124a37e3971f7df. metr-productivity-2025#setting — Retained extracted text, character offsets \[1527, 1552). Context (EM paraphrase): Methods: developers worked on 246 real issues in familiar open-source repositories using early-2025 tools. This population does not represent all developers or later systems. 16 experienced developers metr-productivity-2025#completion-time — Retained extracted text, character offsets \[2468, 2507). Context (EM paraphrase): Core Result, first paragraph; interpret with Methods and the immediately following limitations table. Reported result for the 2025 study, not a current universal estimate. they take 19% longer to complete issues ### We are Changing our Developer Productivity Experiment Design Source ID: metr-productivity-update-2026 https://metr.org/blog/2026-02-24-uplift-update/ Citation: METR developer follow-up · February 2026 METR judges developer benefits likely improved while explaining why participation, task selection and time measurement make the size of the change hard to estimate. Report aliases: METR’s February 2026 update; METR’s 2026 follow-up; METR’s later developer evidence Preview passages: selection, direction-of-selection Publication: 2026-02-24. Basis: Publication date displayed by the original publisher. Access: public-page. Original publisher page and exact retained source reviewed. Source retention does not establish every downstream reuse or training right. Checked: 2026-10-29. Source state: stream-snapshot. Exact retained Stream work/snapshot and SHA256 verified against the 29 October SELECT-only source packet. That lookup found no joined public Stream or bridge reading revision; no reading identity is invented. Work: 54d658ff870e4bbb0638a847bac138a3b543d42ac48e761c915c570500f1cc5c. Snapshot: 8fd5f006c0dc55d1e20448712d3f4b177c53cd92219b9aa3ed71b70d0ed13eca. Text hash: 81b6c9b7d710e11ed96cf12de5f307e8551454cfdcc9f6c66c84f0a657b9adfb. metr-productivity-update-2026#selection — Retained extracted text, character offsets \[431, 540). Context (EM paraphrase): Opening discussion, paragraph beginning Unfortunately. Subsequent paragraphs describe participation, task selection and time measurement. the data from our new experiment gives us an unreliable signal of the current productivity effect of AI tools metr-productivity-update-2026#direction-of-selection — Retained extracted text, character offsets \[715, 738). Context (EM paraphrase): METR says nonparticipation by developers unwilling to work without AI likely understates speedup; the authors believe early-2026 tools speed developers more than early-2025 tools. Lower study pay, task selection and concurrent-agent time accounting limit the size estimate. likely biases downwards ### Measuring the Self-Reported Impact of Early-2026 AI on Technical Worker Productivity Source ID: metr-worker-survey-2026 https://metr.org/blog/2026-05-11-ai-usage-survey/ Citation: METR technical-worker survey · May 2026 A survey of technical workers’ self-reported value from AI tools. It distinguishes value from speed and does not measure an audited financial return. Report aliases: METR’s technical-worker survey; METR’s May 2026 survey Preview passages: self-report Publication: 2026-05-11. Basis: Publication date displayed by the original publisher. Access: public-page. Original publisher page and exact retained source reviewed. Source retention does not establish every downstream reuse or training right. Checked: 2026-10-29. Source state: stream-snapshot. Exact retained Stream work/snapshot and SHA256 verified against the 29 October SELECT-only source packet. That lookup found no joined public Stream or bridge reading revision; no reading identity is invented. Work: 1f24eee26acc9583eb4649a75f64560f3c33c7db3211c7c6b2997a22eabf6d93. Snapshot: 136edbda94fff696943b30a2b15f34cee5c35b551ad06af769dd6bed2e189ace. Text hash: 8b47a32be31a689e627a34baf7b4d2e8998c64dc27e4c9442e0d431b9e502b8f. metr-worker-survey-2026#self-report — Retained extracted text, character offsets \[1308, 1401). Context (EM paraphrase): METR surveyed 349 technical workers in February–April 2026. The range reflects self-reported value under different question formulations, not a confidence interval or observed causal return. METR gives reasons to doubt the magnitude and distinguishes value from speed. Participants self-reported a median 1.4–2x change in the value in their work due to AI tools. ### Machines of Loving Grace Source ID: amodei-loving-grace-2024 https://darioamodei.com/essay/machines-of-loving-grace Citation: Dario Amodei · Machines of Loving Grace Amodei’s October 2024 essay defines powerful AI through broad capability and autonomous work, then explores possible consequences under an assumption of relatively near arrival. Report aliases: Amodei’s powerful-AI vision; Amodei’s powerful-AI essay; Machines of Loving Grace; Amodei’s demanding combination Preview passages: powerful-ai-timing, labor-contribution Publication: 2024-10. Basis: October 2024 is printed in the retained essay. The work metadata has no publication timestamp; month precision is preserved. Access: retained-public-essay. Original personal essay retained in Stream and reviewed here; not an institutional position or a new 2026 statement. Public availability does not establish every reuse right. Checked: 2026-10-29. Source state: stream-snapshot. Exact original Stream identity and SHA256 verified against the retained source packet. Work: 6aa2c807bdcb81da04a30b6cf0b1c0620353219816ab224111920bdc8c02bc32. Snapshot: 07fedfacf7c11b925e82d6c8a319ee710df91c8b7160fae3d7ca3fda1da6af45. Text hash: d8f37e4a35c07240bca22712dece7667d41c3e3efb9f28b7ca62c4c5f807442a. amodei-loving-grace-2024#powerful-ai-timing — Retained extracted text, character offsets \[6255, 6271). Context (EM paraphrase): Amodei says powerful AI could arrive that early but could also take much longer. His working definition combines exceptional cognitive capability, human-like computer interfaces and autonomous work over hours, days or weeks. The essay assumes relatively near arrival to explore later effects; it is not a demonstrated outcome or a standardized AGI forecast. as early as 2026 amodei-loving-grace-2024#labor-contribution — Retained extracted text, character offsets \[75299, 75360). Context (EM paraphrase): In the essay’s long-run economic discussion, Amodei considers sufficiently capable and cheap AI displacing the basis of human economic contribution. He first allows comparative advantage and new complementary jobs to preserve employment. No job-loss date or measured proportion is supplied. most or all humans may not be able to contribute meaningfully ### AI Had Better Work Out Source ID: zitron-ai-financing-2026 https://www.wheresyoured.at/ai-had-better-work-out/ Citation: Ed Zitron · AI Had Better Work Out Zitron argues that the AI investment boom depends on sustained revenue and capital-spending growth. The retained article is his financial interpretation, not an independent reconciliation of every cited figure. Report aliases: AI Had Better Work Out; Zitron’s financing argument Preview passages: financing-thesis Publication: 2026-10-20. Basis: Original publication date retained in source metadata; later source edits remain part of the exact cited snapshot. Access: retained-publisher-page. Full retained article body reviewed; public display uses a short excerpt and attributed interpretation. Underlying numerical claims and linked evidence are not independently verified by inclusion. Checked: 2026-10-29. Source state: stream-snapshot. Exact work/snapshot and SHA256 verified in the 29 October SELECT-only corpus packet; full private source text is not republished. Work: d67ef043cf3188e7c13ff00e8795c99a55e2f6073738d97b950ea6628bea00a2. Snapshot: 642117e5b02149519f05acfd48796b317b705abda025451653b54107fc942d61. Text hash: 412c3083432201a267b8d6be62a62154bc7868bf1df5e92c8f88fcc3ca0430b4. zitron-ai-financing-2026#financing-thesis — Retained extracted text, character offsets \[12434, 12517). Context (EM paraphrase): Zitron asks what would sustain the AI investment boom and argues that revenue, capital spending and debt dependence leave it fragile. This is his financial interpretation; linked figures and third-party reports have not been independently reconciled in this edition. The answer is simple: perpetual, parabolic growth in both AI revenues and AI capex. ### Aswath Damodaran on AI as a business Source ID: damodaran-ai-business-2026 https://aswathdamodaran.blogspot.com/2026/08/ais-bar-mitzvah-moment-from-hype-hope.html Citation: Aswath Damodaran · AI as a business Damodaran examines the revenues, margins and reinvestment needed to justify AI investment. Different business assumptions can support different valuations. Report aliases: Damodaran’s business test; Damodaran’s valuation framework; Damodaran’s AI argument Preview passages: valuation-conditions Publication: 2026-08-20. Basis: Original publication date retained in source metadata; later source edits remain part of the exact cited snapshot. Access: retained-publisher-page. Full retained article body reviewed; public display uses a short excerpt and attributed interpretation. Underlying numerical claims and linked evidence are not independently verified by inclusion. Checked: 2026-10-29. Source state: stream-snapshot. Exact work/snapshot and SHA256 verified in the 29 October SELECT-only corpus packet; full private source text is not republished. Work: 7fe96df3cffa24ec7af60105eada09b5a3d465ed24c98ec9ce59816ff39793b8. Snapshot: 334736629c52f8e90204be8991410d566c1ab0d55c282772850b709ca8cc7ef9. Text hash: 8fa50b829b700a1ab084953a6fc0ae8b2033e015dd738d4b45a0c86dca16c70e. damodaran-ai-business-2026#valuation-conditions — Retained extracted text, character offsets \[53311, 53405). Context (EM paraphrase): Damodaran proposes testing what future revenues, margins and reinvestment would justify AI company prices and capital spending. The original post is dated 20 August 2026; its retained preface records two 21 August updates. This is an analytical framework, not proof that a price or industry-wide bubble verdict is correct. there will be room for significant disagreement on the reach of AI and its value as a business ### Generative AI at Work — arXiv v2 Source ID: brynjolfsson-ai-work-v2 https://arxiv.org/abs/2304.11771v2 Citation: Brynjolfsson, Li and Raymond · Generative AI at Work 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. Report aliases: the customer-support study; customer-support research; customer-support evidence; customer-support results; customer-support comparison; Generative AI at Work Preview passages: resolved-work, novice-gains Publication: 2024-11-06. Basis: Version 2 submission date displayed by arXiv; first submitted 23 April 2023. The versioned abstract is the cited object. Access: public-page. 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. Checked: 2026-10-30. Source state: external. Original public source read on 30 October 2026. An exact retained Stream identity has not been resolved for this edition. Work: not retained. Snapshot: not retained. Text hash: not retained. brynjolfsson-ai-work-v2#resolved-work — Version 2 abstract, second sentence. EM context: staggered introduction among 5,172 customer-support agents; the outcome is issues resolved per hour, not company profit. by 15\\% on average brynjolfsson-ai-work-v2#novice-gains — Version 2 abstract, third sentence. EM context: benefits differ by experience; the most experienced workers show small speed gains and small quality declines. Less experienced and lower-skilled workers improve both the speed and quality of their output ### Time Horizon 1.1 Source ID: metr-time-horizon-1-1 https://metr.org/blog/2026-01-29-time-horizon-1-1/ Citation: METR · Time Horizon 1.1 METR’s January 2026 methodology update describes historical growth in software-task time horizons and changes to the task suite and evaluation setup. Report aliases: METR’s time-horizon methodology; Time Horizon 1.1 Preview passages: historical-growth Publication: 2026-01-29. Basis: Date displayed by METR. Access: public-page. Original methodology update checked. Read alongside the September dashboard status; January is not presented as the latest frontier measurement. Checked: 2026-10-30. Source state: external. Original public source read on 30 October 2026. An exact retained Stream identity has not been resolved for this edition. Work: not retained. Snapshot: not retained. Text hash: not retained. metr-time-horizon-1-1#historical-growth — Opening account of the earlier time-horizon work. EM context: the January update reports continued growth through 2025, but growth estimates depend on task composition; this is historical software-task evidence, not a current AGI clock. a steady exponential increase ### Task-Completion Time Horizons of Frontier AI Models Source ID: metr-time-horizon-dashboard https://metr.org/time-horizons/ Citation: METR · Task-Completion Time Horizons Software-task estimates at defined success rates. The page stopped updating on 8 September 2026 and flags estimates beyond 16 hours as unreliable. Report aliases: METR’s time-horizon dashboard; METR’s time-horizon measurements Preview passages: reliability-threshold, measurement-edge Publication: 2026-09-08. Basis: Page last-updated date, not its original publication date. Access: public-page. METR labels this page no longer actively updated, last updated 8 September 2026. It supports a defined benchmark and its limits, not a complete October model comparison. Checked: 2026-10-30. Source state: external. Original public source read on 30 October 2026. An exact retained Stream identity has not been resolved for this edition. Work: not retained. Snapshot: not retained. Text hash: not retained. metr-time-horizon-dashboard#reliability-threshold — Opening definition of the 50%-time horizon. EM context: duration means human expert task time, not agent runtime. Tasks are primarily software engineering, machine learning and cybersecurity, with explicit success criteria. predicted to succeed half the time metr-time-horizon-dashboard#measurement-edge — Notice above the interactive task plot, checked 30 October 2026. Measurements above 16 hrs are unreliable with our current task suite ### Canaries in the Coal Mine? — August 2026 revision Source ID: stanford-canaries-2026 https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/ Citation: Stanford · Canaries in the Coal Mine? · 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. Report aliases: Stanford’s August 2026 payroll study; Stanford’s payroll evidence; Canaries in the Coal Mine? Preview passages: hiring-first, descriptive Publication: 2026-08-12. Basis: Publisher page identifies the revision as 12 August 2026; the linked PDF is dated August 2026. Access: public-page. Original research summary and linked August PDF abstract checked; the findings are attributed to the authors and their stated population. Checked: 2026-10-30. Source state: external. Original public source read on 30 October 2026. An exact retained Stream identity has not been resolved for this edition. Work: not retained. Snapshot: not retained. Text hash: not retained. stanford-canaries-2026#hiring-first — Publisher summary, fourth finding; linked August PDF, abstract. EM context: ADP payroll sample through June 2026. The authors find no widespread displacement, but a relative employment shortfall among ages 22–25 in AI-exposed occupations. reduced hiring of young workers rather than increased separations stanford-canaries-2026#descriptive — Publisher summary, final paragraph; linked PDF, abstract. EM context: education controls attenuate the pattern; some divergent trends predate generative AI, and the ADP pattern is stronger than national survey benchmarks. These are not causal job-loss estimates. early, descriptive indicators ### Microsoft fiscal year 2026 results Source ID: microsoft-fy2026-results https://www.microsoft.com/en-us/investor/earnings/fy-2026-q4/press-release-webcast Citation: Microsoft · Fiscal year 2026 results The company’s earnings release for the year ended 30 June 2026. It combines reported paid Copilot seats with whole-company cash-flow figures, not an AI-only investment account. Report aliases: Microsoft’s FY2026 results; Microsoft’s fiscal year 2026 results Preview passages: paid-demand, capital-spending Publication: 2026-07-29. Basis: Earnings release date; financial year ended 30 June 2026. Access: public-page. Original earnings release and cash-flow columns checked directly. FY2026 is the specified historical period, not a claim to have assessed every later filing or valued the company. Checked: 2026-10-30. Source state: external. Original public source read on 30 October 2026. An exact retained Stream identity has not been resolved for this edition. Work: not retained. Snapshot: not retained. Text hash: not retained. microsoft-fy2026-results#paid-demand — Satya Nadella quotation in the earnings release. EM context: a vendor-reported seat count, not a measure of customer return or AI-specific profit. Microsoft 365 Copilot reached over 30 million paid seats microsoft-fy2026-results#capital-spending — Cash flows statements, twelve months ended June 30: 2026 $115,948 million; 2025 $64,551 million. EM context: cash spending for the whole company, not an AI-only capex total; release tables are labelled unaudited. Additions to property and equipment ## Attributed positions ### Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory — Anthropic Institute researchers reporting on their employer Position ID: anthropic-workforce-authors Attribution: direct Statement date: 2026-10-16. Basis: Dated authored publication; this is not the date of every underlying observation or interview. The authors report a falling junior share of technical hires and propose that intensive AI adoption at Anthropic may offer an early indication of wider workplace change. - Condition: One rapidly growing AI company; technical hiring still grew in absolute terms. - Condition: Reported internal patterns and the authors’ interpretation do not identify an economy-wide causal effect. - anthropic-workforce-2026#anth-junior [supports] — https://www.anthropic.com/research/employment-at-anthropic ### Jaison R. Abel, Richard Deitz, Natalia Emanuel and Nick Montalbano — New York Fed research staff; article expresses their views Position ID: nyfed-survey-authors Attribution: direct Statement date: 2026-09-01. Basis: Dated authored publication; this is not the date of every underlying observation or interview. The authors interpret their regional business surveys as showing more worker augmentation than replacement so far, while explicitly allowing this to change as adoption matures. - Condition: August 2026 regional business survey, not a national census. - Condition: Workforce changes are employer reports; the relevant charts describe firms using AI. - nyfed-business-survey-2026#nyf-fewer [supports] — https://libertystreeteconomics.newyorkfed.org/2026/09/businesses-are-using-ai-to-transform-work-not-cut-jobs/ - nyfed-business-survey-2026#nyf-more [supports] — https://libertystreeteconomics.newyorkfed.org/2026/09/businesses-are-using-ai-to-transform-work-not-cut-jobs/ ### Joel Becker, Nate Rush, Beth Barnes and David Rein — METR authors of the early-2025 controlled developer study Position ID: metr-productivity-researchers Attribution: direct Statement date: 2025-07-10. Basis: Publication date of this study report. In their randomized study of 16 experienced open-source developers completing 246 issues, AI access increased completion time by 19%. They explicitly restrict what this says about other developers, settings or later tools. - Condition: The study used early-2025 tools in developers’ familiar repositories; it is not an estimate for all work or present-day AI. - metr-productivity-2025#setting [supports] — https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ - metr-productivity-2025#completion-time [supports] — https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ ### Gabriel Unger and Arvind Narayanan — Authors of a research agenda and argument about AI development Position ID: unger-narayanan-collaboration Attribution: direct Statement date: 2026-10-14. Basis: Dated authored publication; this is not the date of every underlying observation or interview. They argue for deliberately developing collaborative AI and for workers having a stronger voice in deployment choices. - Condition: This is a proposed direction and mechanism, not an observed estimate of its effects. - Condition: Agreement with their preferred outcome does not establish feasibility or labor-market results. - steering-beyond-automation#un-voice [supports] — https://www.normaltech.ai/p/steering-ai-beyond-automation ### Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory — Anthropic Institute researchers reporting on their employer Position ID: anthropic-work-organization Attribution: direct Statement date: 2026-10-16. Basis: Dated authored publication; this is not the date of every underlying observation or interview. The report describes increased output alongside rising demands on human attention as AI use spreads internally. - Condition: Observations at one AI company; not a controlled estimate of the causal effect or of other employers’ results. - anthropic-workforce-2026#anth-bottleneck [supports] — https://www.anthropic.com/research/employment-at-anthropic ### Joel Becker, Nate Rush, Tom Cunningham, David Rein and Khalid Mahamud — METR authors of the follow-up methods update Position ID: metr-followup-method Attribution: direct Statement date: 2026-02-24. Basis: Publication date of the authored source. METR judges that developers likely benefited more from AI in early 2026 than in early 2025, while warning that participation, task selection and time measurement make the size of that improvement hard to estimate. - Condition: This is a later team’s assessment of a changed experiment, not the same measurement repeated cleanly. - metr-productivity-update-2026#selection [supports] — https://metr.org/blog/2026-02-24-uplift-update/ - metr-productivity-update-2026#direction-of-selection [supports] — https://metr.org/blog/2026-02-24-uplift-update/ ### Joel Becker — METR author reporting technical-worker self-assessments Position ID: metr-worker-value Attribution: direct Statement date: 2026-05-11. Basis: Publication date of the authored source. The survey reports substantial perceived value gains, while distinguishing value from task speed and cautioning that respondents can misestimate what AI contributes. - Condition: Self-report in a technical-worker sample, not audited customer economics or a measured effect on national productivity. - metr-worker-survey-2026#self-report [supports] — https://metr.org/blog/2026-05-11-ai-usage-survey/ ### Dario Amodei — Anthropic chief executive writing a personal essay Position ID: amodei-powerful-ai Attribution: direct Statement date: 2024-10. Basis: Month printed in the essay; the underlying capture has no publication timestamp. In October 2024, Amodei suggested that powerful AI could arrive as early as 2026 while explicitly allowing a much longer wait. He defined it through exceptional breadth and autonomous work, then explored possible consequences under that assumption. - Condition: The essay’s definition is not a universally agreed AGI endpoint. - Condition: An earliest possibility is not a median date, deadline or claim that the system exists. - amodei-loving-grace-2024#powerful-ai-timing [supports] — https://darioamodei.com/essay/machines-of-loving-grace ### Dario Amodei — Anthropic chief executive writing a personal essay Position ID: amodei-labor-long-run Attribution: direct Statement date: 2024-10. Basis: Month printed in the essay; no job-loss horizon is stated. He allows human comparative advantage and complementary jobs to persist initially, but expects sufficiently capable and cheap AI eventually to undermine the present basis of human economic contribution. - Condition: Conditional long-run argument, not measured job loss or a dated prediction that most jobs disappear. - Condition: The essay treats economic arrangements and distribution as choices needing further debate. - amodei-loving-grace-2024#labor-contribution [supports] — https://darioamodei.com/essay/machines-of-loving-grace ### Ed Zitron — Technology/business commentator and newsletter author Position ID: zitron-bubble-financing Attribution: direct Statement date: 2026-10-20. Basis: Publication date of the retained article. Zitron characterizes the AI boom as a bubble whose financing depends on continued rapid revenue and capital-spending growth. He argues that debt dependence and concentrated counterparties make a disappointment dangerous. - Condition: An attributed critical thesis, not an EM finding that the financing chain must fail. - Condition: The article combines his analysis with reported third-party claims; those chains need separate verification. - zitron-ai-financing-2026#financing-thesis [supports] — https://www.wheresyoured.at/ai-had-better-work-out/ ### Aswath Damodaran — Valuation scholar writing on his personal blog Position ID: damodaran-ai-value-conditions Attribution: direct Statement date: 2026-08-20. Basis: Original publication date; the source preface records two updates on 21 August 2026. Damodaran argues that AI investment should be examined through the business outcomes required to justify its price or capital cost. Different assumptions about markets, margins and reinvestment can produce different valuations; technological plausibility does not guarantee an investment outcome. - Condition: A proposed valuation framework, not a comprehensive independent audit of AI companies. - Condition: The retained article records updates after its original publication; its numerical examples are not adopted as EM estimates. - damodaran-ai-business-2026#valuation-conditions [supports] — https://aswathdamodaran.blogspot.com/2026/08/ais-bar-mitzvah-moment-from-hype-hope.html ### Erik Brynjolfsson, Danielle Li and Lindsey Raymond — Researchers studying AI assistance in customer support Position ID: customer-support-researchers Attribution: direct Statement date: 2024-11-06. Basis: Date of the cited version or source update. The researchers find higher resolved issues per hour after AI assistance, with the largest benefits for less experienced workers. - Condition: One employer and an assisted workflow; the study does not estimate aggregate employment or full deployment returns. - brynjolfsson-ai-work-v2#resolved-work [supports] — https://arxiv.org/abs/2304.11771v2 - brynjolfsson-ai-work-v2#novice-gains [supports] — https://arxiv.org/abs/2304.11771v2 ### METR — Independent AI evaluation organization Position ID: metr-horizon-research Attribution: direct Statement date: 2026-09-08. Basis: Date of the cited version or source update. METR documents rising software-task time horizons, defined at a chosen success rate; benchmark progress does not by itself establish dependable whole-job autonomy. - Condition: The dashboard stopped updating on 8 September 2026; estimates beyond its task range are unreliable. - metr-time-horizon-1-1#historical-growth [supports] — https://metr.org/blog/2026-01-29-time-horizon-1-1/ - metr-time-horizon-dashboard#reliability-threshold [limits] — https://metr.org/time-horizons/ - metr-time-horizon-dashboard#measurement-edge [limits] — https://metr.org/time-horizons/ ### Erik Brynjolfsson, Bharat Chandar and Ruyu Chen — Researchers analyzing ADP payroll records Position ID: stanford-entry-researchers Attribution: direct Statement date: 2026-08-12. Basis: Date of the cited version or source update. The authors find that young workers in AI-exposed occupations have fallen behind less-exposed peers, primarily through reduced hiring, while finding no widespread economy-wide displacement. - Condition: Descriptive pattern through June 2026; education, prior trends and sample differences matter for attribution. - stanford-canaries-2026#hiring-first [supports] — https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/ - stanford-canaries-2026#descriptive [supports] — https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/ ### Microsoft — Company reporting its own adoption and financial results Position ID: microsoft-paid-demand Attribution: direct Statement date: 2026-07-29. Basis: Date of the cited version or source update. Microsoft reports more than 30 million paid Microsoft 365 Copilot seats alongside sharply higher company-wide cash investment in property and equipment. - Condition: Paid seats are evidence of purchasing; they do not disclose AI-specific profit or the return to each customer. - microsoft-fy2026-results#paid-demand [supports] — https://www.microsoft.com/en-us/investor/earnings/fy-2026-q4/press-release-webcast - microsoft-fy2026-results#capital-spending [supports] — https://www.microsoft.com/en-us/investor/earnings/fy-2026-q4/press-release-webcast ## Authored connections ### quality-to-reliability question:ai-work-quality-and-rework → question:ai-delegation-reliability (informs) Connection basis: em-inference A result can be useful because a person caught its mistakes. Removing that review changes the job: the agent must detect failure, recover or ask for help. That is why productivity gains can arrive before dependable autonomy. - Limit: Track the human interventions behind a successful result before applying it to unsupervised work. - anthropic-workforce-2026#anth-bottleneck [context] — https://www.anthropic.com/research/employment-at-anthropic - metr-productivity-2025#setting [context] — https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ - metr-productivity-2025#completion-time [context] — https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ - metr-productivity-update-2026#selection [context] — https://metr.org/blog/2026-02-24-uplift-update/ ### quality-to-returns question:ai-work-quality-and-rework → question:ai-deployment-returns (informs) Connection basis: em-inference Work becomes valuable when someone can use the result. A slower process can still be worthwhile if it produces something better or previously impossible; a faster draft can waste time and money if nobody needs it or finishing it takes too much effort. - Limit: The buyer’s gain and the supplier’s investment return are separate calculations. - metr-worker-survey-2026#self-report [context] — https://metr.org/blog/2026-05-11-ai-usage-survey/ - anthropic-workforce-2026#anth-bottleneck [context] — https://www.anthropic.com/research/employment-at-anthropic - Position: metr-worker-value ### entry-to-institutions question:ai-entry-and-worker-gains → path:entry-routes-narrow (conditional-path) Connection basis: em-inference The same tool can help a novice perform and reduce the employer’s need to hire one. Training, paid entry routes and workers’ influence over deployment determine whether productivity creates a career or closes its first door. - Limit: Measure hiring and progression as well as the productivity of people already employed. - anthropic-workforce-2026#anth-junior [context] — https://www.anthropic.com/research/employment-at-anthropic - steering-beyond-automation#un-voice [context] — https://www.normaltech.ai/p/steering-ai-beyond-automation - brynjolfsson-ai-work-v2#novice-gains [context] — https://arxiv.org/abs/2304.11771v2 - stanford-canaries-2026#hiring-first [context] — https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/ - Position: unger-narayanan-collaboration ## Recorded changes No prior research-edition changes are recorded. Source history is separate from EM authorship. ## Reader reference context Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ### AI exposure In labor economics, how much of a job's tasks an AI system could in principle perform; it is not a forecast of job loss. https://arxiv.org/abs/2303.10130 ### Artificial general intelligence The idea of AI with broad problem-solving, creativity and adaptability across different tasks. https://developers.google.com/machine-learning/glossary#artificial-general-intelligence ### Time horizon METR's measure of AI agents' ability: the length of task, timed by how long skilled people take, that an agent completes at a set success rate. https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/ ### Benchmark A defined set of tests used to compare systems. Results depend on the tasks and measures selected. https://arxiv.org/abs/2211.09110 ### Capital expenditure Money a company spends on long-lasting assets such as buildings, chips and power equipment. https://en.wikipedia.org/wiki/Capital_expenditure ### Counterfactual An argument or scenario that considers what would have resulted if events had gone differently from how they actually did, such as if a policy had never been adopted; as an adjective, relating to this kind of thinking about what did not happen but could have. https://dictionary.cambridge.org/dictionary/english/counterfactual ### Confidence interval A range of values consistent with the data at a stated level of confidence. https://en.wikipedia.org/wiki/Confidence_interval ### Machine learning Methods that let computers learn patterns from examples instead of following rules written by hand. https://developers.google.com/machine-learning/glossary#machine-learning ### METR METR studies what AI systems can do and how reliably they complete tasks, including work with human developers. /entities/metr ### Anthropic Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems. /entities/anthropic ### Dario Amodei Dario Amodei is an AI researcher and Anthropic cofounder who writes and speaks about the capabilities and risks of AI. /entities/dario-amodei ### Figure Figure develops humanoid robots and publishes its own tests of how they perform physical tasks. /entities/figure ### arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. /entities/arxiv