What “Still Waters, Rapid Currents” Shows About AI and Wages

What did the Danish study actually observe and compare?

The Danish study followed workers in 11 occupations the authors selected as highly exposed to chatbots, combining two surveys with monthly administrative records. In Still waters, rapid currents: Early labor market transformation under generative AI, the authors describe survey rounds in late 2023 and late 2024. They invited 115,000 workers in each round; that is the invitation count, not the number who completed either survey. The account reports about 25,000 valid complete responses in the 2024 round, spanning roughly 7,000 workplaces. The study therefore offers a detailed view of selected occupations and participating workers, not a census of Danish workers or a representative estimate for every occupation.

The surveys asked about chatbot use and changes workers associated with it. For this analysis, a chatbot adopter means a respondent who reported using chatbots for work; that self-reported category should not be confused with workers whose employer had formally introduced an AI system or with measured exposure to a task. The authors linked survey responses to monthly administrative records on earnings, hours, and occupation. Their account says the linked workplace data also covered roughly three million other workers at the respondents’ workplaces. This linkage lets the study compare reported experience with recorded labor-market outcomes, while the surveys and registers still answer different questions: one records what workers say changed in their tasks and tool use, the other records employment outcomes such as pay and hours.

The authors anchor their monthly comparisons around ChatGPT’s public launch in November 2022. They compare chatbot adopters with otherwise similar non-adopters in the same occupations, and workplaces with AI initiatives with similar workplaces without those initiatives. They also check how the groups’ labor-market outcomes moved before launch. The authors say those pre-launch trajectory checks make later differences more plausibly attributable to chatbot use or workplace initiatives. That is the authors’ description of their comparison and check; there is no need to add a method label or convert the check into a guarantee that the groups were identical.

The pre-launch check makes later differences more informative when groups had similar prior trajectories. But adoption and survey response were not random: motivation, job design, management, or other factors may affect outcomes. Similar pre-launch paths cannot rule out unobserved differences or concurrent changes. The design supports a careful comparison, not a claim that every later outcome was caused by chatbot use.

The Brookings authors’ account, Still waters, rapid currents: Early labor market transformation under generative AI, supplies the detailed survey, register, and comparison description. The NBER Working Paper 33777 abstract, Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI, states the earnings and recorded-hours result; the author account provides context for interpreting it. Together, the sources support this article’s reading of task reports alongside aggregate measures and its practical task-log suggestion. Neither makes the Danish sample an individual forecast.

Sources: Still waters, rapid currents: Early labor market transformation under generative AI — Brookings authors’ account; Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI — NBER Working Paper 33777

How can work change while average pay and hours stay still?

The Danish study shows how work can be reorganized inside a job before a change becomes visible in average pay or recorded hours. In Still waters, rapid currents: Early labor market transformation under generative AI, the authors classified workers’ free-text descriptions of new AI-related duties into broad task categories. About four in ten reported new tasks involved generating content, about a third involved oversight, and about a quarter involved integration. The categories include ideation and drafting, checking quality or handling ethics and compliance, and fitting AI outputs into existing workflows. These are shares of reported new tasks in the study’s selected occupations, as classified by the authors; they are not shares of all work hours or a rate of task change across the whole economy.

Those categories describe movement within work rather than a simple list of tasks removed. Generating a draft can create a need to judge its fit, check facts, revise it, and connect it to a client or internal process. Oversight and integration may draw on the same subject knowledge as the earlier task, but put that knowledge to use at a different point in the workflow. This reading is an interpretation of the reported categories, not a separate measurement of how much time each duty required or whether the resulting work was better. The survey identifies reported changes in task content; it does not measure the full balance of preparation, checking, correction, and accountability for every worker.

The same author account reports that 85% of chatbot users who said they saved time allocated it to other tasks in their jobs. That finding helps explain why tool use need not reduce recorded hours: saved time can be absorbed by additional work, including work that was previously delayed or handled less thoroughly. It does not show that every user saved net time after checking and correction, or that the added tasks produced more valuable output. Some workers who did not use chatbots also reported new AI-related duties: about 4% said their workload had changed because of AI, with reports varying by workplace support. A task can therefore change for someone who never operates the tool, for example when colleagues’ use changes what must be reviewed or integrated. These remain worker reports, not time-and-motion observations or a measure of the size of any spillover.

Alongside these task reports, the NBER Working Paper 33777 abstract, Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI, reports a precise null for average earnings and recorded hours about two years after ChatGPT’s launch: the analysis rules out differential effects larger than 2%. Here, “precise null” means a confidence-bounded finding that excludes effects above the stated threshold under this analysis; it does not mean the estimated effect was exactly zero. The Brookings authors’ account gives the detailed context for the bounded result. The finding concerns average differences in administrative outcomes over the study window, not every worker’s pay, every type of work time, job quality, or task content.

Reported task changes and unchanged average register outcomes are compatible. Registers record pay and hours, not whether workers draft with chatbots, verify more, or pass work to reviewers. A survey report of a new duty likewise does not establish changes in total hours, output, compensation, or staffing. Task descriptions show reported work changes; linked records test aggregate labor outcomes. Both can be accurate.

The bounded result limits interpretation in both directions. It does not prove task reorganization will eventually change wages, and survey reports do not show every worker experienced a change. Effects below the threshold, differences between workers or workplaces, and dimensions outside earnings and hours can remain. For the period studied, however, the analysis finds no differential average movement larger than 2% in recorded outcomes about two years after launch. Task change shows work being rearranged; by itself it establishes neither an impending aggregate shock nor immunity from one.

Sources: Still waters, rapid currents: Early labor market transformation under generative AI — Brookings authors’ account; Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI — NBER Working Paper 33777

What changes when a tool is supported and task productivity is measured?

In the Danish survey, workplace support went together with greater chatbot use, but the pattern is an association rather than a controlled test of what support causes. In Still waters, rapid currents: Early labor market transformation under generative AI, the authors report that 43% of respondents worked where chatbot use was explicitly encouraged. Among workers in those encouraging workplaces, 61% had access to enterprise tools and 39% had received training. Where encouragement, tool access, and training were all present, 93% reported having used chatbots at work. These figures describe how use clustered with organizational conditions; they do not isolate the effect of any one condition or show that training produced better outcomes. Support can lower practical barriers to trying a tool, while the workers and workplaces that adopt may also differ in other ways.

Access, use, and benefit are separate. A trained worker may use a permitted tool narrowly; another may use a public tool without support and take on checking or privacy work. The Danish pattern does not measure output per hour by support arrangement, so it indicates where use is common, not which employers are more productive. Task productivity means output on a specified task relative to time or other inputs. Feeling faster alone does not establish more accepted work per hour after review and correction.

A different kind of evidence comes from Generative AI at Work — NBER Working Paper 31161. In one customer-support deployment, 5,179 agents received access through a staggered introduction of a conversational assistant. The study measured issues resolved per hour and reported a 14% average increase. This is stronger evidence of a task-level productivity effect than a capability demonstration or a worker’s memory of time saved: the outcome was tied to an operational task and measured during a workplace rollout. It remains a result for that firm, support workflow, assistant, and measurement period. It does not show that all the additional resolutions met the same quality standard in every setting, nor does it establish a general effect on wages, staffing, or total workload.

The average also concealed meaningful differences by starting experience. The NBER study reports much larger gains for novice and lower-skilled agents, while experienced and highly skilled agents saw minimal effects. One plausible interpretation is that assistance helped less experienced workers draw on patterns or responses that stronger performers had already developed; that mechanism is an interpretation, not a separately established result here. It illustrates why the same tool can change task performance differently across workers doing nominally similar work. For an experienced worker, the assistant may add little to already efficient handling; a novice may have more room to improve. The result does not imply that experience has become worthless: quality control, unusual cases, context, and accountability may still matter even when a routine resolution metric rises.

The studies answer different questions: Danish encouragement, access, and training were associated with use; the customer-support deployment measured an output change on one task and found uneven gains by experience. Neither gives a portable estimate for a new team. Applying the field result elsewhere requires checking task inputs and outputs, information available to the tool, review and rework, and accepted quality. Support makes use more feasible; a productivity claim needs a measured task outcome for a specified group.

Sources: Still waters, rapid currents: Early labor market transformation under generative AI — Brookings authors’ account; Generative AI at Work — NBER Working Paper 31161

Why does changing task content not settle the displacement question?

The OECD’s The impact of AI on the workplace: Evidence from OECD case studies of AI implementation offers a different lens from the Danish survey and linked administrative records. It draws on nearly 100 qualitative implementation cases in manufacturing and finance across eight countries. Case evidence can show how an implementation unfolded inside an organization: which steps were reorganized, where workers had to learn new procedures, how responsibilities shifted, and what workers experienced as the system entered a workflow. Those details can make visible changes in skill use or job quality that an average wage or hours measure would not describe. The collection is useful for identifying kinds of change and questions to investigate, rather than estimating how common each change is.

That strength comes with a boundary. The cases are selected accounts of implementation, not a representative sample from which to calculate how often jobs are redesigned, eliminated, or improved. They cover AI broadly and predate the current generative-AI wave, so they cannot establish a GenAI displacement rate or tell a worker what is happening in a particular occupation now. A striking case may reveal a possible mechanism without telling us its prevalence; a quiet case may not capture later effects. The OECD material should therefore contextualize the Danish evidence, not override its specific results or be treated as a forecast for a new tool.

Qualitative redesign cases also complicate two easy readings. A replacement narrative can miss implementations where tasks move, skills change, or human review remains part of the process. Automatic reassurance can miss changes in workload, control, monitoring, or responsibility that matter to workers even when headcount does not immediately fall. The OECD cases do not prove either broad harm or broad benefit. They direct attention to outcomes that a displacement-only question can hide: who now performs each step, what expertise is required, who checks errors, who is accountable, and whether workers experience greater or lower discretion. These are questions the evidence can help frame, not findings to attribute uniformly to every case.

A changed task bundle is reason to inspect work design and accountability, not proof a job is disappearing or safe. The OECD cases can suggest what a worker might examine, but current local evidence is needed to judge adoption, demand, and feasible alternatives. Unlike the Danish linked comparison or the customer-support task measure, the cases supply neither a displacement rate nor a personal forecast.

Sources: The impact of AI on the workplace: Evidence from OECD case studies of AI implementation

Why can measured efficiency fail to become higher wages?

The ILO review, The impact of GenAI on jobs, productivity and work organization: a review of the empirical evidence, brings together experiments, firm data, platform studies, and worker surveys from several countries. Its summary describes productivity gains as uneven and often unverified: reported time savings have not translated consistently into measured output, earnings, or employment across the evidence it reviews. That is a synthesis across different settings and measures, not one estimate for a typical worker or a finding that gains cannot eventually affect pay. To understand the gap, it helps to separate five links that are often compressed into the word “productivity.” This conversion chain is an analytical model for examining a workplace, not a sequence the ILO tested as a universal law.

A faster first draft or routine response is perceived time saved, not necessarily net time saved on the complete task. Prompting, checking, correcting, reconciling outputs, and documenting decisions may consume the gain. Compare the full task cycle, including review and rework, across stable cases, and record accepted quality, errors, and accountability. Skipping review can make completion look faster without verified efficiency; added review may improve quality without raising speed.

The next link is verified net productivity: comparable or better accepted output from the same total inputs, after correction and oversight. Even that does not ensure higher worker compensation. Capacity released from one task may be reassigned to another, used to reduce a backlog, or absorbed by more coordination rather than turned into additional saleable output. Demand may be too weak to justify producing more, or the added output may have little value to the organization or its customers. To distinguish these possibilities, a team would track the volume of completed and accepted work, the queue or unmet demand, the time and staffing used across the whole workflow, and any changes in quality or downstream costs. A rise in throughput under unchanged conditions is different from a report that workers feel faster.

A further question is who captures the value. An employer might gain from more output, lower costs, faster service, or improved quality; some benefits could instead flow to customers, platform providers, or other parts of the organization. The relevant evidence would include the outcome the organization values, whether it changed after the workflow change, and whether the gain persisted after implementation costs. A worker may create value without controlling its allocation. The last link is wage pass-through: the extent to which a verified productivity gain is reflected in workers’ compensation. Pay-setting arrangements, budgets, job classifications, and timing can affect whether that happens. Comparing compensation over time for the affected work, alongside stated pay criteria and the measured gain, would help establish whether any pass-through occurred; a single worker’s faster task completion cannot establish it.

The chain offers a practical diagnostic: measure the whole task, verify accepted output or quality per total input, identify demand for added capacity, locate the economic benefit, then check whether compensation changed. A break anywhere can explain why efficiency and wages move differently without making a time-saving report false. The evidence gap does not show gains never raise pay; demand, expertise, durability of gains, and pay arrangements vary. Given the ILO review’s heterogeneous findings, local measures are needed before claiming workers will or will not share in a gain.

Sources: The impact of GenAI on jobs, productivity and work organization: a review of the empirical evidence — ILO

What does the occupational-switching exception establish?

The Brookings authors’ account, Still waters, rapid currents: Early labor market transformation under generative AI, identifies a selected mobility result within the Danish chatbot adopters: workers who switched occupations moved toward roles with higher pay and greater chatbot relevance, and their earnings grew 12 percentage points faster than those of other Danish workers. The unit here is a group of occupation switchers, not every adopter and not a within-job raise. “Twelve percentage points faster” describes a relative difference in earnings growth over the comparison; it does not mean each switcher received a 12% pay increase. The comparator is other Danish workers, not an otherwise identical set of workers randomly assigned to switch or stay.

The destination matters to the interpretation. The account says switchers moved toward higher-paying occupations with greater chatbot relevance, which makes occupational mobility a plausible place to look for uneven gains that an adopter-wide average could miss. But the number of moves was too small to shift average earnings among adopters. A favorable outcome for a small subset can coexist with a muted group average because most members did not make that transition, and because the measured average combines workers following different paths. This is a compositional point: subgroup growth can be notable while its weight in the larger group remains limited. The result does not establish that moving to any role labeled AI-related produces faster earnings growth. The reported growth comparison also does not tell us how much any switcher earned before moving, how earnings changed immediately after a move, or whether the difference accumulated over a longer period. Those are distinct questions from whether the destination occupation had higher pay. Keeping the measures separate helps prevent a relative growth statistic from being mistaken for a promised starting salary or raise.

Selection also limits what the comparison can say. Switchers may have had different prior skills, credentials, networks, opportunities, motivation, or employer support than workers who stayed. Their new roles and earnings paths may reflect those differences or other changes occurring at the same time. The authors’ account reports a relative association among the observed groups; it does not show that chatbot use caused an individual transition or caused the faster growth. Nor does a 12-percentage-point growth difference specify the starting salaries, absolute currency gains, costs of transition, or outcomes for someone considering a similar move. Those questions require evidence about the particular roles and worker circumstances.

Mobility may be one channel to investigate when AI changes a task bundle: an adjacent role could value existing domain knowledge alongside integration or oversight. The Danish result makes this possibility concrete, but is neither causal career advice nor a personal forecast. A worker considering a move should check local vacancies, entry requirements, likely pay against a salary floor, and the cost and time to bridge skill gaps. Similar titles may hide credential barriers or a costly reset; this sample can prompt questions about feasible destinations, not answer them for an individual.

Sources: Still waters, rapid currents: Early labor market transformation under generative AI — Brookings authors’ account

What does worker-reported willingness to bargain add?

Workers’ Perspectives on Artificial Intelligence: Productivity Gains and Job-loss Fears, a Federal Reserve Bank of Boston report, adds a measure earlier in the possible path from work change to compensation: what workers think their productivity gains mean for their willingness to seek a raise. The report uses two December survey waves, in 2024 and 2025, with a rotating panel of about 1,300 U.S. household heads. It compares reported perceptions and stated intentions; it does not follow a recorded sequence from AI use through a request to a negotiated wage change. In the regression sample, the group reporting the strongest perceived productivity gains was small, about 6%. For that subgroup, the modeled likelihood of saying they were more likely to ask for a raise was 14%. This is a subgroup estimate from a survey model, not a 14% raise, a share of all workers who received more pay, or a forecast for any one respondent.

The result can fit before an actual request or a change in administrative earnings because intention is a different and earlier outcome. A worker who believes a tool has increased their contribution may become more willing to raise compensation in a review, but timing, access to a manager, workplace pay rules, confidence, and bargaining power could affect whether that intention becomes a conversation. Even a conversation need not change pay immediately: budgets, review cycles, and compensation decisions can intervene. That sequence is a plausible interpretation of how stated willingness might precede a recorded pay outcome. The Boston Fed report measures the survey responses, not this proposed mechanism or the intervening steps. Its estimate therefore signals reported willingness among a small group, rather than demonstrating that wage pressure has already begun across a market. It may help explain what workers expect to do, but it cannot establish whether employers perceive or share the same view of productivity.

The comparison with Still waters, rapid currents: Early labor market transformation under generative AI — Brookings authors’ account is useful precisely because the outcomes differ. The Boston Fed result concerns U.S. household heads’ perceived gains and stated likelihood of asking; the Danish study links workplace and worker information to administrative earnings and hours and reports no average change above its stated bound in the period examined. One finding does not cancel the other: people may report greater willingness before any recorded compensation change, while a register outcome can remain unchanged over its observation window. Geography, populations, samples, and measures also differ, so the two studies do not establish that this sequence happened to the same workers. The Boston Fed survey adds a possible early signal in workers’ own expectations. Whether requests occurred, employers responded, or compensation moved requires separate evidence.

Sources: Workers’ Perspectives on Artificial Intelligence: Productivity Gains and Job-loss Fears — Federal Reserve Bank of Boston

How can you test whether a task change warrants a career move?

A short task log can show what to investigate next, but it cannot tell you whether your occupation is disappearing. Choose one recurring deliverable and compare similar instances with and without assistance. Record preparation, prompting, checking, correction, and integration time; whether the work met the same acceptance standard; who remained accountable; and where any time difference went. Note when cases differed in difficulty or received extra help. This captures your workflow under observed conditions, not a general productivity effect or a compensation consequence.

For a recurring client summary, for example, an assistant might produce a draft quickly, while the worker still checks facts against source records, corrects omissions, adapts it to the client, and confirms what can be shared. Another summary may need little correction because its source material is cleaner. Compare the full process and accepted quality across similar cases rather than draft speed alone. Minutes, revisions, review steps, and acceptance are observations; whether the tool changes task value or shifts responsibility is an inference. The illustration organizes questions; it does not describe every summary workflow.

Compare three feasible responses. An in-role upgrade could mean using an approved tool for a bounded step while strengthening the judgment, verification, or integration the work still needs. Check whether the employer supports the workflow and values that work. An adjacent move could preserve domain knowledge while changing the task mix; review current local vacancies, credentials, and prerequisites. A larger change may fit if the current path is narrowing, but first examine training time, cost, entry requirements, and income during transition. None is automatically best; feasibility depends on both workplace evidence and available options.

Apply your real constraints before choosing: a salary floor based on obligations, location limits, study hours and cost, health needs, and family or caregiving responsibilities. Review several current postings and distinguish required qualifications from preferred ones. If postings, task observations, and employer plans conflict, gather information: ask which tasks are changing, verify whether an approved workflow is planned, speak with someone in an adjacent role, or try a small project before paying for substantial training. A course may fit your schedule but exceed your budget; a role may require a credential or location change you cannot make now.

Use the log to choose a low-cost test, then compare paths with actual prerequisites and personal limits. If total time falls while quality and accountability hold, the workflow changed; local evidence on demand, role design, and pay is still needed before moving. If checking absorbs the gain, investigate the process first. A brief log cannot reveal employer-wide demand, long-run adoption, or pay outcomes; it is not a risk score. Use it to focus a conversation, posting review, or reversible experiment. A supervised pilot, informational conversation, or small project can test a prerequisite before a major commitment.

What is a proportionate next step?

Across ordinary work, note what changes in one recurring task when AI is involved: what you check or redo, what output is usable, and where time saved goes. Keep notes concrete enough to discuss with a manager or compare with another workflow. Then investigate one feasible response, such as an adjustment in your role or a nearby role, checking prerequisites, likely pay, location, and training against your income needs, time, health, and family responsibilities. If information is missing, gather it before paying for substantial training or making a larger move.

If uncertainty remains, the free task checker can organize task-level change-pressure signals and possible first actions. It does not provide a validated probability that you will be displaced. The paid career roadmap can compare scenarios against your preferences and constraints, but it does not guarantee employment or income. Use either as decision support after identifying what you need to know; neither can replace evidence about a specific employer or local opening.

The proportionate next move is therefore small: observe one task, investigate one feasible option, and spend only when the evidence and your constraints support the next step. If they do not yet line up, continue gathering information rather than treating uncertainty itself as a reason to make a costly commitment.

Questions readers ask