What Human–AI Collaboration at Scale Shows About User Agency

What does Human–AI Collaboration at Scale show about user agency when work has higher stakes?

Human–AI Collaboration at Scale offers a measured answer: users appeared more involved in conversations when work was classified as higher stakes. The paper reports longer exchanges and greater engagement with AI output in those contexts. That is evidence of interaction, not proof that users checked answers against independent evidence, retained authority over real-world decisions, or achieved safer results.

Agency has at least three links: what someone does in chat, whether they can inspect, challenge, and approve the work, and what happens after it is used. The paper gives a signal about the first. Its findings are consistent with people exercising judgment, but conversation behavior alone cannot establish the quality of that judgment or whether it shaped the final decision.

A worker can apply the finding by inspecting one consequential task’s checking and approval path: what could catch a mistake, and who is accountable before the output is acted on? This is an editorial application, not an intervention tested by the paper. The result does not estimate job security, employer adoption, or displacement. The sections that follow distinguish visible participation from effective oversight and downstream outcomes.

Sources: Human–AI Collaboration at Scale: Task Criticality, Agency, and Friction Across 250,000 Conversations

What did the paper measure when it called work consequential?

The paper uses “consequential” as a description of the task context, not as a rating of the AI system or the person doing the work. In *Human–AI Collaboration at Scale: Task Criticality, Agency, and Friction Across 250,000 Conversations*, researchers classify conversations along three dimensions: how reversible the outcome is, how many people it may affect, and how severe an error could be. They then group tasks into four tiers, from ephemeral through operational and consequential to high-stakes. The dimensions capture different kinds of consequence. Reversibility concerns whether a result can readily be undone or corrected. A private draft that can be discarded is different from a decision already sent to a client or used in a process. The affected-people dimension distinguishes work with a narrow audience from work that may shape outcomes for colleagues, customers, or other stakeholders. Error severity asks what the mistake could mean: a minor inconvenience is not equivalent to a serious legal, financial, or public consequence. The tier labels summarize task-context judgments; they are not a formal safety certification or a finding that harm occurred. That distinction matters because a “high-stakes” label can sound more definitive than the underlying measure. It indicates that the conversation was assigned to a category based on the framework’s dimensions. It does not establish that every task in that category carried the same risk, that users faced the same constraints, or that an error was likely. A criticality tier does not describe model capability, occupational exposure, or employer adoption. A task can carry serious consequences even if the system handles it poorly. The authors analyzed 249,834 conversations sampled from Claude.ai’s website between April 25 and May 9, 2026. The sample covered Free, Pro, and Max users and excluded Claude Code, Team, Enterprise, and API use, according to the paper’s account of the dataset. This is a substantial collection, but its unit is a website chat in a limited window, not a representative survey of workers or employers. The categories describe the classified sample, not prevalence across an occupation or labor market. One reported contrast is that high-stakes conversations averaged 12.4 turns, compared with 6.5 turns for ephemeral conversations. A turn is an exchange in the chat. The result supports a statement about conversation length across the paper’s classified groups. It does not show that users spent nearly twice as long working, devoted more attention to verification, understood more, or produced a better result. A long exchange could include careful clarification, repeated revisions, unresolved ambiguity, or other activity; the turn count alone cannot distinguish among them. The number is a behavioral pattern, not evidence about oversight quality. The paper’s design consequently joins two measurements: a classification of the task’s possible consequences and counts of interaction within the chat. It does not measure the model’s accuracy on each task, the user’s independent checking, the authority to approve or reject the output, or the consequences after it leaves the conversation. The classification gives context for comparing interactions; it does not turn those interactions into a reliability test. Readers can conclude that the sampled chats assigned higher criticality had more exchanges on average. They cannot infer from that result alone that the AI was dependable, that people retained effective control, or that any particular job became more or less secure.

Sources: Human–AI Collaboration at Scale: Task Criticality, Agency, and Friction Across 250,000 Conversations

Which observed behaviors support a limited agency claim?

The clearest support for a limited agency claim is not simply that people used AI on consequential tasks. It is that the reported form of interaction shifted with criticality. In *Human–AI Collaboration at Scale: Task Criticality, Agency, and Friction Across 250,000 Conversations*, adaptation was the most common reported way users engaged with outputs across criticality tiers. It reached 52% in high-stakes conversations, while asking for explanations or further detail, the paper’s “understand” category, rose to 21% there. Less direct, unmodified use also fell at the highest criticality level. These patterns fit users treating consequential answers as material to work on, not simply accepting every response unchanged. Adaptation means significant reworking of an AI output. That is compatible with the user bringing context, preferences, or subject knowledge to a draft and shaping what the system produced. This can signal authorship within the interaction. But the label does not tell us whether the revisions corrected an error, improved factual accuracy, or only changed tone and format. A polished revision may still preserve a mistaken premise. Nor does the conversation record establish whether the user relied on independent source material while editing. The “understand” category captures users asking for explanations or more detail. This is a stronger sign of active questioning than a one-turn exchange: a person is seeking grounds for an answer instead of only requesting its output. Yet a request for explanation is not evidence that the explanation was sound, that the person understood it, or that they checked it elsewhere. The behavior shows a question was asked, not answer quality or the user’s grasp of its limits. Less unmodified use also matters, but its meaning is narrow. Not passing an answer through unchanged shows some interaction with the output before the conversation ends. That may reflect scrutiny and a willingness to alter or reject what the system offered. The measure does not reveal what changed, whether the user could decline the recommendation, or who controlled the decision. A user can edit an answer yet remain dependent on it, and the final action may happen outside the chat. The strongest fair reading, then, is that higher stakes coincide with more visible user involvement: more adaptation, more requests to understand, and less verbatim reuse. This is a meaningful signal of agency-in-interaction, not an empty proxy. It supports the idea that users respond to consequential work with active engagement. It leaves the harder question open: whether that engagement becomes independent verification, real authority to override, and a better or safer result. Those are additional links in the evidence chain, and this paper’s reported conversation categories do not measure them.

Sources: Human–AI Collaboration at Scale: Task Criticality, Agency, and Friction Across 250,000 Conversations

Why is engagement evidence of involvement rather than control?

Engagement is evidence that a person is involved in an exchange; control is a stronger claim about what that person can decide and do. To control a consequential use of AI, a worker needs a basis to assess an answer, the ability to revise or reject it, and authority over whether it is acted on. Human–AI Collaboration at Scale observes conversation patterns. It does not follow answers through independent verification, approval, or execution. Its findings can support a limited claim about agency within interaction, but cannot establish control over the work as a whole. Reading an answer closely may mean checking its logic against a source. It may also mean trying to understand unfamiliar language or choosing a follow-up. The paper's category for understanding captures requests for explanation or detail; it does not report whether users checked claims against independent evidence. Reading is compatible with scrutiny, but the behavior alone cannot tell us what kind occurred. That matters when an answer could shape a decision affecting a client, colleague, or member of the public. Adaptation has a similar limit. Reworking an answer can fit a user's context, voice, or format, and can express judgment or authorship. But a tailored answer may still contain an unsupported premise. Changing wording or emphasis is not the same as testing whether a claim is true. Human–AI Collaboration at Scale reports how users engage with outputs, not whether edits improved accuracy or safety. The favorable reading is plausible: users may bring their knowledge to bear. The evidence does not show how often that knowledge caught a mistake. More turns are ambiguous too. A longer exchange may leave room to challenge an assumption or refine a response. It could also reflect an unclear request, repeated correction, prompt experimentation, or difficulty getting a usable answer. These are possibilities, not explanations demonstrated for the conversations in the paper. Turn count records exchanges; it does not identify effort, comprehension, or the reason for continuing. Activity is not a measure of review quality. Control can shift after the chat. A worker might prepare a draft but lack permission to change the final decision; a manager or client might approve it; or a workflow might send it onward automatically. Conversely, a user may retain a meaningful veto after delegating substantial work. Conversation behavior alone cannot resolve these possibilities. Effective agency depends partly on who can inspect the output, stop or amend its use, and take responsibility for the next step. Those conditions may not appear in the exchange. This distinction does not dismiss the study's positive signal. More active interaction as task criticality rises is consistent with people responding to higher stakes by engaging carefully. But consistency is not confirmation: the paper does not connect these behaviors to verified checks, retained authority, or downstream results. The defensible conclusion is narrower and useful. Engagement shows participation in conversation; whether that participation became control remains open and requires evidence about review, authority, and what happened beyond the chat.

Sources: Human–AI Collaboration at Scale: Task Criticality, Agency, and Friction Across 250,000 Conversations

What does the privacy-preserving method leave uncertain?

The study’s privacy protections also set a boundary on what its categories can show. In “Enabling independent research on how people use Claude,” Anthropic says outside researchers designed questions for Anthropic Insights, while Anthropic ran the analysis and shared aggregated outputs after privacy and legal review. Researchers did not receive raw conversations. They saw category descriptions and tallies, not individual exchanges they could read and recode. This enabled analysis without exposing transcripts, but external teams could not check labels against underlying conversations.

The pipeline matters because a category is produced through several interpretive steps. “Enabling independent research appendix” says Anthropic Insights applies researcher-written questions, or facets, to sampled conversations. Claude answers open-ended facets; similar answers are grouped into clusters and summarized. Multiple-choice or numeric answers are tallied. Researchers receive privacy-preserving descriptions and counts, with small clusters suppressed. Anthropic also reviewed outputs for privacy and quality before release. These safeguards reduce access to sensitive material; they do not establish that each classification accurately captures what a user did.

The appendix calls open-ended cluster names and descriptions interpretations, not objective measures. A forced category can make the system assign a judgment even when a conversation gives little basis for one. Cluster boundaries are unstable: wording differences can split a topic, while small groups may be merged; repeated runs on the same data can produce different clusters. Here, broad labels such as reading, adapting, or directing summarize behavior inferred from conversations. They are not transcript-level proof that a particular user independently checked an answer or retained authority over a consequential decision.

Anthropic’s account adds a second limit: question wording is sensitive. The company says researchers tested questions on WildChat, where they could inspect underlying conversations, but that dataset skews toward casual and creative use. Some questions that worked there produced misleading categories when applied to Claude traffic. Anthropic reports that an Oxford facet was removed after the team suspected a misphrased prompt had generated misleading cluster descriptions. This is concrete evidence that the method can misclassify, not evidence that every result is invalid. The company’s account explains the process and its known problems; it is not an independent transcript-level audit of this paper’s agency categories.

The fair reading is neither to dismiss the findings nor to treat privacy review as measurement validation. Aggregate patterns can support a cautious claim about how classified interactions differ as task criticality rises. They cannot show exactly what an individual understood, whether a check was correct, or whether a final decision changed outside the chat. Give the pattern weight as evidence about categories, not direct observation of effective oversight in each case.

Sources: Enabling independent research on how people use Claude; Enabling independent research appendix; Human–AI Collaboration at Scale: Task Criticality, Agency, and Friction Across 250,000 Conversations

What comparison helps interpret the validation caveat fairly?

The validation caveat is important, but it does not mean public conversation data is automatically useless. A separate evaluation, “Can public chat data predict real-world AI misalignments?”, tested a specific question: whether older, publicly released WildChat prompts could help estimate how often newer OpenAI models produced defined undesirable behaviors in private production use. The researchers sampled about 100,000 WildChat conversations, generated a new final assistant response with each of five OpenAI models, and compared 19 behavior categories with production estimates based on at least roughly 200,000 conversations per model. For 95% of the WildChat predictions, the reported difference from the production rate was within 1.04 orders of magnitude. That result is a fair counter to a blanket claim that WildChat cannot carry useful signal because it is old or publicly collected. In that evaluation, it retained information about specified model behavior rates despite a reported two-to-three-year gap between WildChat collection and current usage. But the same report also found a meaningful boundary: prediction degraded for categories more connected to technical and agentic use, where current production traffic had shifted furthest from the dataset. The authors say WildChat is a weaker proxy for tool-rich, longer-running workflows. Those findings do not validate the Anthropic study’s measures of task criticality or human agency. The OpenAI evaluation estimates model-output behavior rates against production comparisons; “Human–AI Collaboration at Scale” classifies the context and pattern of Claude conversations, including whether users read, adapt, or use outputs directly. The model, platform, period, task definitions, labels, and comparison ground truth differ. A validation result for one dataset and one model-behavior outcome cannot establish that another study’s classifications accurately capture whether a person checked an answer, understood it, or could reject it. The distinction matters because Anthropic’s account, “Enabling independent research on how people use Claude,” describes a privacy-preserving process in which external researchers received aggregated outputs rather than raw transcripts. It also says question wording can affect Claude’s categories, and that questions tested on WildChat sometimes produced misleading categories when applied to Claude traffic. Conversely, the separate OpenAI result is not a rebuttal to that specific constraint: it shows that WildChat can work as a proxy for certain tested outcomes, not that it will validate all categories on another platform. The calibrated conclusion is therefore neither “public data cannot tell us anything” nor “validation elsewhere settles this paper.” The OpenAI report supports a narrower methodological point: public chat data may retain useful signal when researchers compare a defined outcome with an appropriate production reference, while performance can weaken when the data and target use diverge. Whether its labels identify effective oversight, and whether oversight improved decisions, would require validation designed around those separate outcomes. Validation must remain specific to the outcome actually tested.

Sources: Can public chat data predict real-world AI misalignments?; Enabling independent research on how people use Claude; Human–AI Collaboration at Scale: Task Criticality, Agency, and Friction Across 250,000 Conversations

What stronger evidence of agency looks like in an actual workflow

A stronger sign of agency appears at the handoffs around an AI-assisted answer: who chose to delegate, who can challenge or change the output, and who accepts responsibility before it reaches someone else. The qualitative study “Episodic oversight in generative AI workflows: A nine-step protocol for preserving human agency (OP-9)” makes these signals concrete. Across interviews and workflow materials from two German technology firms, it describes workers intervening while drafting, refining, and reviewing AI-generated material. It examines reported work episodes and artifacts, while the Claude.ai analysis measures interaction categories. Neither establishes that a particular oversight practice prevents errors. Oversight begins before a prompt is written. In the study, workers scope the task, set boundaries, and decide where an LLM may contribute. Agency therefore includes discretion over whether and how to delegate, not just the ability to edit an answer afterward. The researchers describe external audiences and legal or reputational stakes as cues that prompted participants to intervene more. They are case findings, not universal triggers or a formula for review. During refinement, the question is what happens to the output. The study describes professionals checking claims and provenance, adding domain context, and deciding what to keep, rewrite, or discard. Reviewers can trace a claim to its source and test whether revisions meet the task’s requirements. Still, an edit is not automatically a sound check. Rewording may improve fit or tone without establishing that a factual claim is correct. The paper identifies where judgment can occur; it does not certify the quality of each judgment. Review makes responsibility more observable when another person can see what changed and who approved the deliverable. The paper reports checks involving peers, managers, or legal and compliance staff; feedback sometimes sent work back for revision. Documentation and sign-off can show who owned a decision at a handoff, while peer review creates another opportunity to challenge an AI-assisted segment. Neither traceability nor a named approver guarantees correctness. They make delegation and accountability inspectable, rather than leaving “a human checked it” untested. The study draws on 15 semi-structured interviews in two early-adopting German technology firms, plus internal AI-use documents, annotated workflow screenshots, and demonstrations. This let researchers describe situated practices and compare accounts with artifacts. They do not show how common the practices are across workers or industries, or test whether the proposed nine-step protocol improves outcomes. Read alongside the Claude conversation study, the distinction is practical: chat involvement shows participation; a workflow record can show where discretion, challenge, and responsibility were exercised. Whether those steps caught errors or changed results requires evidence beyond either study.

Sources: Episodic oversight in generative AI workflows: A nine-step protocol for preserving human agency (OP-9)

What is the strongest case that engagement does signal agency?

The strongest favorable reading is that higher-stakes users appear to do more than receive an answer. In “Human–AI Collaboration at Scale: Task Criticality, Agency, and Friction Across 250,000 Conversations,” the authors report that high-stakes conversations more often involved reading for understanding and adapting output, while direct, unmodified use declined. Adaptation was common in consequential work. Together, these patterns fit a recognizable form of agency: a person brings context the system lacks, asks for clarification, and reshapes a suggestion for a particular purpose. That interpretation is not certain, but it is stronger than saying the chat record shows only passive acceptance.

Longer exchanges may give users room to state constraints the first request left implicit. A follow-up can narrow a broad answer, expose a mismatch, or ask for an explanation. Revising output is consistent with judgment entering the work rather than the answer passing straight through. Reading for understanding is a plausible sign that a user wants to grasp the reasoning or assumptions, not merely copy the result. The reported patterns support a limited claim: interaction changes with classified stakes, and users appear more involved in handling output.

That positive reading matters because skepticism can overcorrect. Chat behavior is not meaningless simply because it falls short of a verified outcome. Asking for an explanation, changing a draft, or declining to use text verbatim are observable choices within the exchange. They can preserve authorship and create opportunities to challenge an answer. The study offers evidence compatible with those possibilities across a large sample of Claude.ai conversations. It does not warrant dismissing participation as theater or assuming the tool alone completes every consequential task.

The boundary is that each behavior admits more than one interpretation. A longer exchange could allow clarification, but it could also reflect unresolved ambiguity or repeated attempts to get a usable answer. Adaptation can make output fit the user’s context without correcting a factual error. Reading can be deliberate appraisal, but the labels cannot show whether someone compared the answer with an independent source or understood its uncertainty. Nor do they establish whether a user could reject the suggestion, who held approval authority outside the chat, or whether the final decision changed. These are limits of the record, not claims about what sampled users did.

The fairest conclusion gives the favorable signal its proper weight: engagement may be one visible component of agency, and rising engagement with stakes is consistent with users taking a more active role. Yet involvement alone does not demonstrate effective control. The paper cannot tell us when scrutiny caught a mistake, whether a person retained meaningful authority, or whether the result improved. Those questions require evidence about checks, decisions, and consequences beyond conversational behavior.

Sources: Human–AI Collaboration at Scale: Task Criticality, Agency, and Friction Across 250,000 Conversations

Why this study cannot answer a worker's job-security question

The study observes conversations, not workplace adoption, job redesign, labor demand, or displacement. A consequential task may attract active AI use without showing whether the tool can perform it reliably or whether an employer changes staffing. “Human–AI Collaboration at Scale: Task Criticality, Agency, and Friction Across 250,000 Conversations” helps describe engagement in a consumer chat setting; it does not track organizational staffing decisions or workers’ later job outcomes. Several distinct questions separate chat use from a claim that work will disappear. Capability asks whether a system can perform a defined task to an acceptable standard. Use asks whether someone brought that task to the system. Adoption asks whether an employer has put the tool into a repeatable workflow with access and responsibility arranged. Redesign asks whether duties or roles change. Labor demand concerns how many workers organizations seek. Displacement concerns whether workers lose work or employment. The preprint informs use and the task contexts users describe, but does not measure these later links. A consequential request can show consultation without showing reliance, correctness, workflow acceptance, or work removed. A manager might use AI to prepare a draft and still require a qualified reviewer; another organization might not permit that tool for the task. The paper’s association between classified stakes and chat engagement cannot distinguish among them. The International Labour Organization’s “Generative AI and Jobs: A Refined Global Index of Occupational Exposure” illustrates why exposure needs separate evidence. The 2025 working paper combines task data, worker input, expert discussion, and model predictions. It draws on 29,753 tasks in Poland’s occupational classification, a survey of 1,640 employed people across broad ISCO-08 groups, and 52,558 data points about perceived automation potential for 2,861 tasks. Its index estimates occupational exposure gradients; it does not count workers already displaced. The ILO says most occupations still include tasks requiring human input and frames job transformation as the more likely impact in its analysis. That conclusion belongs to its global task-based method. It neither verifies the Claude study nor predicts what happens in a particular employer or occupation. Keep each signal attached to its question. The preprint reports interaction patterns across tasks classified by criticality. An exposure index estimates which occupational tasks may be affected. Adoption evidence would show tools entering workplaces and defined processes; redesign evidence would trace changed duties or staffing. Displacement claims require observed employment outcomes and attention to other causes. “High-stakes,” “exposed,” and “automated” are not interchangeable measures. Neither study gives an individual a redundancy estimate or retraining prescription. A task may be exposed while a role still depends on context, trust, judgment, or accountability; affected duties can also change without removing a whole role. These are reasons to examine actual arrangements, not guarantees of safety. Neither source identifies a reader’s task mix or employer plans. The free task checker can organize task-level change-pressure signals, but does not provide a validated displacement probability. The narrow conclusion is that higher classified stakes coincided with more involved chat behavior in this sample. Whether that becomes effective oversight, changes workplace design, or affects employment requires different evidence.

Sources: Generative AI and Jobs: A Refined Global Index of Occupational Exposure; Human–AI Collaboration at Scale: Task Criticality, Agency, and Friction Across 250,000 Conversations

What is a proportionate next step for one consequential task?

Choose one recurring task that now involves AI and map four things before changing the workflow: what would be difficult to reverse, who could be affected, what independent source or test could reveal an error, and who must approve the result before use. This practical checklist applies the paper’s dimensions of reversibility, affected people, and error severity; Human–AI Collaboration at Scale did not test it or show that it improves outcomes. The paper’s finding is narrower: conversations classified as higher-stakes showed more engagement. Whether that became effective oversight remains open. A separate qualitative study of professionals describes task selection, revision and checking, and team review or sign-off as distinct oversight points, but its interview findings do not establish their prevalence or effects. For example, an internal draft that can be corrected before circulation may need a lighter check than an external recommendation with financial consequences. In either case, name the independent check and approval point; a longer chat alone is not a safeguard. If you also want to map which parts of your work are changing, the free task checker offers task-level change-pressure signals, not a probability of job loss. Record the check and approver, then revisit the workflow if the evidence or consequences change.

Sources: Human–AI Collaboration at Scale: Task Criticality, Agency, and Friction Across 250,000 Conversations; Episodic oversight in generative AI workflows: A nine-step protocol for preserving human agency (OP-9)

Questions readers ask