The Return to Judgment

A firm that buys better output has not necessarily bought a more capable organization. It may have acquired a useful tool, a temporary production advantage, a different division of labor, or some combination of the three. The economic return depends on which of these changes persists when the task changes, the model fails, the vendor raises its price, or a worker has to make a decision independently.

The patent-lawyer experiment by Autor and coauthors is the central evidence this month. Randomized access improves assisted drafting, while gains on an unassisted follow-up assessment concentrate among senior lawyers. The junior estimate is imprecise and close to zero. It does not establish general deskilling. It does establish why a production metric and a learning metric should be recorded separately. S01S01 · MIT Stone Center / NBERDoes AI Assistance Enhance or Erode Expertise? Evidence from a Three-Month Field Experiment in Patent DraftingFull text reviewed; working paper

That distinction travels further than the experiment itself, provided the change of scale remains explicit. An AI credential can improve a candidate’s interview prospects without proving that the candidate has acquired a productive skill. A stock-price response can reveal an expectation of future productivity without recording an increase in national output. A data center can benefit a county while reducing the value of the homes nearest its boundary. The central analytical task is to identify the relevant outcome before deciding whether the evidence is favorable.

September also offers a useful corrective to a simple displacement narrative. Fairlie and Wu do not find a significant summer-2026 increase in relative unemployment among recent college graduates using their CPS comparisons. That result deserves a place beside evidence of hiring pressure in particular exposed occupations. It neither invalidates every narrow displacement finding nor licenses a claim that the labor market is unaffected. S02S02 · IZA@LISERThe Early Impacts of AI on Employment among Recent College GraduatesFull text reviewed; working paper

For an executive, the immediate implication is to commission an evaluation that follows a task through delivery and tests what the organization retains. For a policymaker, it is to distinguish total surplus from the people and places that receive it. For an economist, it is to treat the links between task performance, skill formation and aggregate productivity as separate empirical questions.

This edition reviews 40 candidate and coverage records, selects 10 main papers, registers 32 source records, and develops six figures. Publication records that refer to the same paper are identified as such; they are not counted as independent corroboration. A paper’s rank reflects its analytical importance under a fixed editorial rubric, not a probability that its conclusion is true.

August’s edition, The Interface Became an Economic Institution, examined how access becomes attention, judgment, permission and action. Its central concern was the design of the connection between a user and an effective intervention. September adds a question about the stock of capability left behind. A good interface may help someone complete a task; it does not automatically tell us whether that person can recognize a different problem tomorrow. S11S11 · SchymAugust 2026 Economics RadarPrior edition; historical scores frozen

The change is an extension of August’s argument. If interfaces allocate attention and decide which actions are easy, they also influence opportunities to practice. A system can keep the user engaged in diagnosis, ask for a reason before revealing an answer, or move the user directly to acceptance. Those design differences could affect learning even when immediate output looks similar. The September experiment does not randomize all these interface choices, so this is a mechanism to test rather than an identified explanation of its results.

Three broader developments give the issue economic weight. First, several new contributions move beyond exposure measures toward outcomes with a clearer decision meaning: expert-rated work, interview invitations, unemployment, transaction prices. Second, the aggregate literature is becoming more explicit about the bridge from technical improvement to economic value. That bridge may run through capital markets, historical production functions or calibrated scenarios, but each route imposes different assumptions. Third, distribution appears at several scales: between junior and senior workers, majority and underrepresented applicants, low- and high-capability countries, and nearby residents and a wider local economy.

The weak point remains the middle of the transmission chain. September has relatively informative evidence on selected tasks and several ambitious ways to discuss aggregate effects. It has less direct evidence that connects a measured task gain to sustained firm value after integration, supervision, rework and opportunity costs. The gap is especially relevant when an organization measures activity because the intended business outcome arrives later.

Two distinctions should therefore organize the reading. A result can be economically important without being causal, as with a market signal about expectations. A result can be credibly causal without being economy-wide, as with a randomized task experiment. Treating either distinction as a reason to dismiss the evidence wastes information. Treating the distinctions as irrelevant produces unwarranted conclusions.

What changed since August

Historical August scores remain unchanged. The prior article and live page were recovered, but a separate historical component-score ledger was not available. This edition therefore makes no numerical claim about an increase or decrease in the quality of the literature.

The search covered NBER, CEPR, IZA, CESifo, IMF, OECD, university and author repositories, arXiv, and primary publication searches in major economics and science outlets. Coverage was deliberately broader than papers with “AI” in their titles: employment, measurement, human capital, software delivery and infrastructure incidence all belong in the economic account. Searches of speeches, book releases and conference listings supplied leads, not an automatic presumption of new findings.

Selection followed three screens. The contribution had to change an economic question, supply a useful method, or challenge an important interpretation. Its date and version had to fit the edition, with older lineage disclosed. Finally, the available evidence had to support the claim made here. Full text was reviewed for all ten selected studies. Abstract-only items remain in frontier watch or the exclusion log; they do not receive detailed causal appraisals or extracted effect sizes.

This is a curated research census, not an exhaustive bibliometric census. Journal and conference coverage is incomplete, and some inaccessible or poorly versioned items remain leads. The candidate ledger records these gaps alongside substantive exclusions. Its forty entries include some coverage leads and publication-lineage records; they should not be described as forty fully appraised new papers.

The fixed score weights are methodological rigor 25%, originality 20%, economic significance 20%, long-run importance 15%, policy or business relevance 10%, and surprise 10%. Components are judgmental 0–100 assessments. The package retains the unrounded weighted value and displays one decimal. Differences of a few tenths are not meaningful precision, and a high theoretical score does not make a scenario empirically established.

Several exclusions matter for interpreting the month. The Measurement Revolution? entered the NBER series in September but was already on arXiv in August. AI and the Market for Signals has an August draft date and a September public posting. Voice AI in Firms and two NBER production/infrastructure papers have older public lineages. Their September versions should not be presented as first discoveries. The accessible CEPR record for Everyone Is a Manager carries an October 1 revision, so its substantive claims are held for the next edition. S16S16 · Authors / arXivVoice AI in FirmsFull revised text; older lineage S19S19 · CEPREveryone Is a Manager: AI Use as an Act of ManagementExcluded: accessible revision beyond cutoff S22S22 · Authors / arXivThe Measurement Revolution? arXiv recordVersion-lineage evidence; same work as S04 S28S28 · CEPRAI and the Market for Signals: CEPR recordSeries date; same work as S09

An organization earns a return from AI through at least two distinct objects. One is the flow of usable output generated while the system is available. The other is the stock of capability that remains in people, routines and institutions. The first can rise while the second grows slowly, remains unchanged or declines. A business case that records only the first object cannot determine which path it is on.

This distinction is familiar in economic terms. A machine can substitute for a task while complementing the judgment required to specify and evaluate it. An employee can become more productive in a narrow workflow without becoming more adaptable across workflows. An organization can improve delivery by concentrating review in a few experienced workers, even if that arrangement reduces the learning opportunities available to the next cohort.

The patent experiment gives the distinction empirical content, but it does not supply a universal production function for expertise. The plausible mechanism is that prior knowledge changes the quality of engagement with the tool. Someone who knows what a good answer requires can identify a weak premise, test an implication and learn from a correction. Someone with less knowledge may lack the basis for choosing among plausible alternatives. Assistance can still be useful to both people; the route from use to learning need not be the same.

There is a competing explanation that should remain open. The independent task may favor experience in a way that differs from the assisted drafting task. The trial may be too small or too short to detect junior learning. A training intervention could alter the pattern. These possibilities do not dissolve the distinction between production and capability. They determine how widely the particular estimates should travel.

A two-account evaluation

The practical response is to maintain two accounts. The production account measures quality-adjusted delivery, including review, rework, coordination and the cost of errors. The capability account measures independent diagnosis, transfer to a new problem and the ability to detect a failure. Both accounts need a baseline and a comparison; neither is adequately represented by a satisfaction survey or a raw activity count.

Consider a proposed rollout to an analytical team. A production evaluation could follow a defined class of assignments from intake to accepted output, recording total paid time and material revisions. A capability evaluation could periodically ask the same workers to diagnose a related unfamiliar case without assistance. Randomizing the rollout or training design would separate the intervention from selection into use. The independent assessment should be designed before results are visible, so it does not become a convenient test chosen to vindicate the rollout.

These are proposed measurements, not estimates supplied by September’s papers. Their purpose is to reveal a trade-off that an ordinary productivity dashboard can hide. If delivery improves while independent performance does not, the organization may still choose to scale. It would then need to decide how to maintain resilience, who owns verification and how junior staff will acquire the knowledge needed for future review.

Why verification can become scarce

If generating an acceptable first draft becomes cheap, attention can shift toward checking, integrating and taking responsibility for it. That does not imply that verification must always expand proportionately. Better models can reduce some errors, and automation can assist review itself. The relevant margin is the residual error that matters for the decision and the cost of detecting it.

A firm that evaluates only first-pass output will miss the possibility that the bottleneck moved downstream. It might complete more drafts but place a larger burden on a small group of reviewers. Alternatively, the tool could reduce routine review and release experienced staff for difficult cases. These alternatives have different consequences for throughput, training and pay, even if the initial task benchmark is identical.

The economic question is therefore about the organization of attention. How many decisions require experienced judgment? How much evidence must a reviewer inspect? Does the tool make errors easier to detect, or merely make the output more fluent? Does the user know when to escalate? A good evaluation traces these questions through completed work, rather than assuming that a faster upstream task produces an equal reduction in total cost.

The apprenticeship margin

Junior work often has a joint product: the current assignment and the skill acquired while doing it. A firm can rationally eliminate low-value tasks while unintentionally weakening a training process that was never recorded as an investment. Conversely, it can waste time preserving an old task that no longer teaches anything valuable. The existence of learning-by-doing does not justify keeping every traditional step.

The appropriate test asks which experiences build the capability the organization will still need. Error diagnosis, deliberate practice, explanation and feedback may matter more than manually reproducing an entire routine workflow. Assistance could help create those experiences, but that benefit requires design and measurement. It should not be inferred merely from access to a powerful system.

Here the signaling model adds a further complication. A junior assignment may also certify effort or ability to a supervisor. If AI makes that artifact less informative, the organization must find a better signal. Requiring more output of the same kind could intensify a contest without improving learning. A targeted assessment of reasoning or correction may be cheaper and more informative, although its validity must also be established.

From expertise to diffusion

At the country level, the analogous issue is absorptive capacity. Access to a model is only one input into productive use. Local knowledge, institutions, organizational routines and infrastructure influence which problems can be solved and which solutions can be implemented. The historical development paper makes persistent capability differences visible, but it does not show that current AI must widen them.

AI could lower the cost of obtaining some complements. It might provide usable technical guidance, support management or improve access to education. If those effects are large and durable, they would weaken the bottleneck. The question is whether assistance substitutes for the need to learn, helps people learn, or does both on different margins. A credible development agenda should test these alternatives directly.

The strongest version of September’s thesis is therefore conditional: as output becomes easier to generate, the economic value of judging, integrating and learning from it can rise. This is an inference from the evidence, not a new measured elasticity. It would be weakened by widespread findings that verification costs fall faster than output expands and that independent capability improves across experience levels without additional support. It would be strengthened by evidence that rollout returns systematically depend on the organization of review and learning.

The nine links below are an analytical map. They are not stages in a measured funnel, and the source populations cannot be multiplied together to calculate an economy-wide conversion rate.

For business evaluation, the critical missing link is often between task change and firm effect. For public policy, it is between aggregate value and distribution. For long-run growth, it is between current assistance and the stock of complementary capability. These are different research priorities; a single headline about adoption cannot resolve all three.

Does AI equalize expertise or reward existing expertise?

Assisted work can improve most for less experienced users while independent gains accrue more clearly to experienced users. These statements refer to different outcomes. The disagreement becomes testable when studies assess both, use comparable tasks and follow workers long enough to observe retention. A production gain alone cannot decide whether the skill distribution has converged.

Are young workers already being displaced?

Evidence of pressure in a defined payroll or hiring segment and a null in broader graduate unemployment can both be correct. The relevant test aligns age, occupation, education, timing and labor-force status before comparing estimates. Neither a narrow exposure correlation nor a broad null supplies a complete account of entry-level opportunity. September reduces confidence in sweeping claims while increasing the value of more specific monitoring.

Do market valuations reveal productivity or enthusiasm?

Prices can contain information about real future gains and can also reflect changing beliefs, rents and discount rates. A structural mapping clarifies what must be true for the productivity interpretation. It cannot make alternative explanations disappear by definition. Later production and cash-flow outcomes provide a way to evaluate the signal, especially when the exposure measure is fixed in advance.

Are data centers a local benefit or a local cost?

The answer depends partly on “local.” A county-wide outcome can be positive while households next to a facility experience a loss. This is a distributional fact to measure, not an inconsistency to resolve by choosing one geographic scale. A complete appraisal needs a common counterfactual, careful treatment of transfers and explicit standing for affected households. S03S03 · UC Berkeley Energy InstituteData Centers Lower Nearby Housing ValuesFull text reviewed; working paper S12S12 · NBERData Centers and the Local Economies in the Age of AIFull text; revision delta unverified

Will cheap intelligence narrow international gaps?

The optimistic mechanism is lower-cost access to knowledge and capability. The pessimistic mechanism is stronger returns to complementary skills, capital and institutions that are already unevenly distributed. Historical mobility estimates show why the second mechanism matters; they do not rule out the first. The decisive experiments compare access alone with interventions that help users build and apply the complements.

Can an AI credential reduce exclusion while making assessment harder?

Yes, as a logical possibility. A signal can improve access to an interview while the underlying capability remains uncertain. The next stage of the hiring process then matters more. Evidence on invitations should be followed through offers and performance, with attention to the cost of obtaining a credible signal. Otherwise an initially useful credential could become another expensive screening requirement.

Electrification and organizational complements. The useful analogy is that an enabling technology may require changes in production arrangements before its full value appears. For AI, the relevant complements include workflow design, review and learning. The limitation is equally important: software can diffuse and change far faster than factories can be rebuilt. A historical lag is not a calibrated forecast for the present.

Calculators, spreadsheets and the location of expertise. When routine calculation becomes cheaper, value can shift toward problem formulation, checking and interpretation. The analogy helps specify the likely task reallocation. It does not prove that independent skill must deteriorate or that every profession will follow the same path. Language-model output has different error and verification properties from deterministic arithmetic.

Apprenticeship as joint production. An assignment can produce both current output and future human capital. Automation can change the private incentive to provide that experience, especially when its learning value is not measured. The analogy suggests evaluating training substitutes; it does not justify preserving inefficient tasks merely because they were once formative.

Industrial siting and concentrated costs. Large facilities have long produced benefits and burdens at different geographic scales. The lesson is to measure incidence and mitigation under the same counterfactual. A data center has its own resource use, employment intensity and amenity profile; borrowing the net result from another industry would be misleading.

These analogies organize hypotheses. None contributes a numerical estimate to this edition, and none substitutes for the current evidence.

Sources: S01, S03, S05–S08, S10. Exact values, source-table locators and interval definitions accompany every figure.

Deployment authority inside organizations. The OECD agentic-AI overview is relevant to the question of who can authorize, monitor and reverse an automated action. Its qualitative scope differs from a causal productivity evaluation. The next useful evidence would link the allocation of authority to completed work, errors and escalation, rather than infer value from the number of agents deployed. S15S15 · OECDAgentic AI in OrganisationsOfficial overview; frontier watch

Beliefs across countries. The multinational expectations paper can help identify where people anticipate facilitation and where they fear displacement. Beliefs can influence adoption and policy, but the accessible primary abstract does not justify a detailed numerical verdict here. The full paper and its survey design should be reviewed before using country comparisons as a measure of readiness. S14S14 · CEPRWill AI Help or Hurt Work and the Economy? What People Around the World ExpectPrimary abstract only; frontier watch

Independent audit of AI-assisted research. Work on auditing pre-analysis plans points toward a practical complement to AI measurement. The open question is whether an audit improves error detection against an independent reference, including false positives and disagreement across models. A convincing audit should survive a change in the system that generated the original analysis. S18S18 · NBERThe Utility of AI in Auditing Pre-Analysis PlansFull text; methods context

The cost of the surrounding system. Session repricing and studies of the development workflow make clear why a model’s advertised token price is not a firm’s total cost. Verification, repeated attempts, orchestration and task completion need a common denominator. The promising research design measures cost per accepted output under a stable quality criterion, then tests how that changes at scale. S13S13 · NBERWriting Code versus Shipping CodeFull text; revision delta unverified S31S31 · Authors / arXivControl the Harness, Control the CostAbstract reviewed; no realized firm ROI inferred

Investment before uncertainty resolves. Sequential investment models of AI support schemes raise a useful policy question: when should evaluation precede subsidies or infrastructure commitments? The frontier item remains theoretical in this edition. A policy recommendation would require verified program details, the relevant market failure and evidence on additionality. S20S20 · CEPRSubsidizing the Unknown: Europe’s AI Support SchemesPrimary abstract; frontier watch

October watch. Revisit the pre-cutoff version of Everyone Is a Manager if it becomes independently verifiable; inspect full survey evidence behind the cross-country expectations study; and look for replications that connect assistance to retained expertise. The review should record what changed in a revision rather than treating the date on a repository page as the date of discovery.

The next useful paper would estimate the joint return to production and learning. Randomize teams across ordinary access, access with structured explanation and error diagnosis, and a suitable comparison condition. Define completed output and independent capability before treatment. Follow the same workers through the rollout and a later retention assessment, while recording review time and rework.

Its central figure would place quality-adjusted delivered output on one axis and change in independent capability on the other, with uncertainty for each randomized arm. That chart is not produced here because the required joint data do not exist in this package. A scatterplot assembled from unrelated papers would falsely imply a common population and a measured trade-off.

The economic contribution would be to distinguish a tool that substitutes for a skill from a tool that helps build it, and to estimate whether the learning-oriented design pays for itself. Heterogeneity by baseline skill should be prespecified. Attrition, spillovers between workers and the cost of the training design should be part of the primary analysis, not left to an appendix after a favorable task result.

For a firm, the result would support a concrete choice among deployment designs. For labor economics, it would identify whether the organization of assistance changes the accumulation of human capital. For growth, it would supply a better microfoundation for a complement that is currently assumed more often than measured.

The most useful economic question has moved beyond whether AI can produce an impressive answer. It is whether an organization can repeatedly turn that answer into value while maintaining the ability to judge it. September’s evidence gives a reason to measure those two achievements separately.

I would make independent capability part of the evaluation of a major rollout, particularly where experienced staff must verify work or train the next cohort. That does not require withholding a useful tool. It requires knowing whether the production gain is accompanied by learning, dependence or a shift in who carries responsibility. A firm can make a sensible decision under any of those outcomes if it measures them honestly.

The broader distributional lesson is similar. Aggregate gains do not answer who receives them. A credential can help at one hiring gate while a gap remains. A facility can contribute to a regional economy while imposing a concentrated local loss. A scenario can deliver much more output alongside difficult adjustment for a particular class of workers. The appropriate response is to improve the account of incidence and the design of the intervention.

My provisional conclusion is that judgment will remain a consequential complement to cheap output. The strongest challenge would be evidence that independent capability improves broadly, verification becomes reliably cheaper, and gains survive the transition from individual tasks to completed organizational work. Those are observable claims. They deserve the next round of experiments.

The edition is a research synthesis with original interpretation, not a meta-analysis. It preserves incompatible denominators rather than pooling them. Working papers may change, and the dates below distinguish publication from data vintage and earlier circulation. The cutoff is a publication-evidence cutoff; it does not imply that every dataset reaches September 2026.

Approximate 95% intervals in the patent and hiring figures use the published point estimate plus or minus 1.96 times the rounded standard error. They can differ from intervals calculated from unrounded estimates or the authors’ preferred finite-sample procedure. No covariance matrix is available to test differences across separate task estimates. For the hiring comparison, the paper’s reported Wald tests are retained. The housing figure uses the published main interval and explicitly leaves the nearest-mile interval unavailable in the extracted table.

The GDP comparison uses rounded abstract values from a single clearly identified vintage. The scenario table transcribes Table 6 and assigns no probabilities. The development comparison uses Table 5’s estimated threshold, whose reported bootstrap interval is wide; it is not mixed with the separate operational-cutoff results. Historical scores are frozen, and no monthly aggregate “AI progress” score is constructed.