ChatGPT’s expanding world of work reveals a broader task mix

OpenAI released data covering 165 work activities across nearly 100 countries on October 26, giving researchers a more detailed view of consumer ChatGPT use. ChatGPT’s expanding world of work describes a widening range of requests. The central tension concerns what those requests reveal about economic change, especially when using a tool and improving work are different outcomes.

For employers, the release raises a practical question about where employees find AI useful. For researchers, it offers a way to connect conversational activity with established occupational categories. Neither audience should treat a changing task mix as a direct measure of productivity, worker displacement, or business adoption.

Anthropic’s research on Claude provides a useful comparison because it also tracks how AI use spreads across tasks. Together, these efforts make workplace AI more observable. They also increase pressure on vendors and buyers to distinguish evidence of experimentation from evidence of lasting value.

What ChatGPT’s expanding world of work actually measures

According to OpenAI’s work activity release, the quarterly series spans the first quarter of 2025 through the third quarter of 2026. Across 74 countries observed throughout, the leading ten activities averaged 51.1% of classified work messages, down from 56.0%.

That is a decline of 4.9 percentage points. The finding concerns the composition of work messages, rather than the amount of work completed.

A changing composition matters because an adoption story can look different beneath its headline totals. A tool might attract more use while serving familiar purposes. Alternatively, its existing audience might start bringing it a wider variety of problems. These represent different forms of expansion.

The release creates an opportunity to investigate that distinction. Researchers can examine whether activity patterns move together across places or develop differently. The useful question becomes which work enters the conversation, rather than whether a country simply appears enthusiastic about AI.

The occupational connection comes through Intermediate Work Activities, categories that group related tasks across jobs. OpenAI draws these categories from O*NET, an occupational information framework. Its occupational content model separates work activities from job titles, worker requirements, and workplace conditions.

This distinction changes the unit of analysis. Preparing an explanation, evaluating information, and organizing material can appear inside different occupations. Looking only at job titles would obscure those shared activities.

Consider a hypothetical employee preparing a presentation. The assignment might involve interpreting customer feedback, checking financial assumptions, and drafting explanatory text. A job label describes the employee broadly. Activity categories describe pieces of the assignment.

That finer view gives researchers a better starting point for studying where assistance enters a workflow. It still leaves the surrounding process outside the frame. Someone must decide whether the presentation answers the business question and whether its supporting evidence holds up.

The same caution applies to comparisons over time. A declining share for a leading activity does not establish that people perform it less often. Other activities might simply grow faster.

For that reason, ChatGPT’s expanding world of work should be read as a map of changing attention within the observed platform. Its value lies in directing better questions toward particular activities. Turning that map into an account of economic outcomes requires additional evidence.

Employers face a measurement problem beyond adoption

The immediate pressure falls on employers who need to connect employee interest with dependable business results. A broader range of requests expands the territory they must evaluate.

An organization can approve an AI tool without understanding where it helps. Employees might use it to explore an unfamiliar topic, improve a draft, or work through an analytical problem. Each activity creates a different relationship between assistance, verification, and responsibility.

The management question therefore extends beyond access. A finished draft has little operational value if its claims require extensive correction. An answer that helps an employee understand a problem might create value even when no generated text reaches the final deliverable.

Those possibilities suggest different measures. Completion time matters for repetitive assignments. Accuracy matters when errors carry consequences. Learning matters when employees must perform similar work independently later.

A single adoption dashboard cannot settle all those questions. Counting active accounts establishes participation. Counting messages establishes interaction. Neither automatically establishes that the organization handled an assignment better.

Independent research illustrates the importance of observing outcomes directly. Erik Brynjolfsson, Danielle Li, and Lindsey Raymond studied a conversational assistant used by customer support agents. Their workplace productivity study measured issues resolved per hour and found benefits that differed across worker experience levels.

The relevance here concerns research design. Customer support offered a defined setting, a specific intervention, and an observable operational outcome. That supports a different kind of inference from a platform’s distribution of requests.

An employer reading the new Signals release should therefore treat popular activities as candidates for investigation. Popularity identifies something employees are trying to accomplish. It does not explain whether the current approach succeeds.

Take a hypothetical manager using AI to prepare a project update. Producing readable prose is part of the assignment. Establishing which milestones actually slipped, which decisions remain open, and which commitments changed requires supporting records.

If the draft misstates those facts, fluency can hide additional work. Someone must locate the discrepancy, recover the original context, and revise the update. Evaluating only drafting speed would miss that burden.

This is where personal knowledge management becomes relevant to the broader workplace discussion. The quality of an answer depends partly on whether people can locate and assess the information behind it.

That connection does not establish a performance claim for remio or any other product. It identifies a separate component of the workflow: keeping useful evidence available when a decision needs support.

For buyers, the practical implication is to ask for a complete account of the task. What did the employee receive, what did the assistant produce, and what verification followed? The answers provide a stronger purchasing basis than an impressive volume of activity.

Claude research shows why broader use needs interpretation

ChatGPT’s expanding world of work arrives within an existing effort to measure AI through observed behavior. Anthropic’s AI adoption research offers a particularly relevant comparison.

Its March 2026 report found that Claude.ai’s leading ten tasks fell from 24% of conversations in November to 19% in February. Anthropic also described coding activity moving toward its developer interface and changes in personal and coursework use.

A broader task distribution can reflect several changes happening at once. Product migration, audience composition, and changing needs all deserve examination before drawing an economic conclusion.

The Claude comparison supports a cautious interpretation of platform data. People choose where to perform an activity. If a specialized interface becomes more suitable, activity can leave a general chat product without disappearing from the vendor’s wider ecosystem.

That makes platform boundaries economically meaningful. A consumer conversation and an automated software workflow offer different views of work. Studying either can be useful, provided the analysis explains what sits outside it.

The datasets also use different categories and populations. OpenAI’s country-level work activity measure and Anthropic’s conversation task measure do not provide a direct contest over workplace adoption. Comparing their percentages as equivalent would create a misleading ranking.

A better comparison concerns the questions each company investigates. Do activities become more varied? Does experience change how people use assistance? Do interfaces attract different assignments? These questions help explain the development of AI use without declaring a vendor the winner.

Anthropic’s report also associated longer user tenure with more successful interactions, while acknowledging alternative explanations. Experienced users might learn better strategies. Early adopters might also differ from later arrivals in ways that influence results.

That uncertainty matters for employers evaluating broader ChatGPT work use. Giving employees access and allowing them to develop useful habits are separate organizational decisions. A tool can be available before a team knows which assignments suit it.

From a product perspective, this creates a demanding interpretation problem. A rise in requests can signal useful discovery, unresolved difficulty, or repeated attempts to obtain an acceptable answer. The interface records the interaction, while the outcome may occur elsewhere.

A hypothetical analyst might ask for several revisions because the assignment is complex. Another might stop quickly because the answer meets the need. A third might abandon the tool after an unhelpful response.

Message volume alone cannot distinguish those situations. Understanding them requires evidence about the task’s purpose, its difficulty, and what happened afterward.

The competitive context therefore sharpens the article’s main tension. Vendors can increasingly describe the work people bring them. Buyers still need evidence about the work people finish.

The strongest reporting will preserve both perspectives. Behavioral data reveals patterns that individual anecdotes cannot establish. Outcome studies test whether those patterns translate into benefits within particular settings. Each addresses a different part of the workplace AI story.

Country differences challenge a universal rollout strategy

Geographic variation makes a uniform explanation of workplace AI less convincing. An activity’s importance inside a country’s observed usage does not establish its importance across that country’s economy.

A share describes a relationship between a category and its comparison group. If analytical requests occupy more of the observed mix, that tells researchers about the mix. Establishing why requires evidence beyond the percentage itself.

Several hypotheses deserve testing. Occupational composition might influence the kinds of requests people bring. Language, access, organizational rules, and familiarity with AI might also influence participation.

These are possible explanations, rather than findings established by the release. Separating them matters because each implies a different response from employers and policymakers.

If a pattern reflects local occupations, task selection becomes central. A training program should address assignments people actually perform. If access explains a pattern, training alone will not resolve the underlying constraint.

If organizational rules shape participation, informal consumer use may reveal only part of the picture. Employees facing restrictions might avoid particular requests or use approved systems outside the observed platform.

Country comparisons should consequently avoid turning a usage profile into a judgment about workforce capability. An activity can be uncommon in a dataset for reasons unrelated to whether workers would benefit from assistance.

The same reasoning applies within multinational companies. A central team might identify a promising use case and assume it travels easily. Local offices still need to assess the language, source material, approval process, and customer expectations surrounding the assignment.

Consider a hypothetical customer communication workflow. Drafting an explanation is only part of the task. Staff must also understand which policies apply, which promises they can make, and when a case requires escalation.

A model’s ability to produce polished language does not settle those local requirements. Evaluating the workflow means checking whether employees receive the context and authority needed to use that language appropriately.

For developers, geographic differences create research questions about product fit. The useful next step is to investigate an observed need with local users. A country-level pattern can guide that investigation without prescribing the feature to build.

For workers, the implications are similarly specific. Experience with a popular activity elsewhere does not guarantee that the same approach suits a local assignment. The relevant test concerns the actual work and its constraints.

ChatGPT’s expanding world of work therefore supports a more differentiated account of adoption. Broader participation need not produce identical behavior. A technology can spread while people adapt it to distinct purposes.

That interpretation also changes how success should be discussed publicly. A universal story invites sweeping claims about who is ahead. A task-based story asks where assistance fits, what prevents useful adoption, and which outcomes justify further investment.

Those questions demand more work than a country ranking. They also offer employers and researchers a clearer route toward decisions they can defend.

The missing work limits conclusions about jobs

OpenAI says the dataset covers individual ChatGPT plans, excluding enterprise messages and Codex. Its U.S. comparison also places about 43% of employment-weighted mapped tasks in “Other IWAs,” outside the individually reported categories.

Those boundaries prevent the release from serving as a census of work or a measure of job replacement. The absent activity matters to any interpretation.

The accompanying sampling methodology adds further detail. The broader Signals sample uses 1 million consumer messages monthly and excludes accounts with certain eligibility or data-use restrictions. Automated classifiers assign behavioral categories, while privacy protections add statistical noise.

These choices shape what researchers can observe. A classifier infers the category of a message. It does not interview the worker, inspect the finished deliverable, or confirm that a requested action occurred.

Privacy protection also creates an interpretive tradeoff. Aggregated information allows patterns to be studied while protecting individual messages. Fine distinctions require care, especially when a subgroup draws on fewer observations.

The appropriate response is to match the conclusion to the measurement. A pattern of requests can support an investigation into demand for assistance. Establishing a change in employment requires employment evidence.

A job contains responsibilities that do not appear neatly inside a prompt. Someone must notice that a problem needs attention, obtain missing information, coordinate with colleagues, and accept responsibility for the result.

Assistance with a component can alter that job without eliminating it. It might reduce effort, increase the scope of an assignment, or introduce a new review burden. Which effect dominates remains an empirical question.

The opposite mistake would be to dismiss the release because it cannot answer every labor question. Detailed behavioral records offer a useful starting point. They help researchers identify activities worth studying more closely.

Independent population research provides another perspective. Alexander Bick, Adam Blandin, and David Deming’s generative AI adoption study used nationally representative U.S. surveys to examine use at work and home. That approach helps place platform observations within a wider population.

The methods complement each other. Surveys can ask about work hours and experiences beyond a particular vendor. Platform data can describe interactions without relying entirely on a respondent’s memory.

Neither method removes the need for careful interpretation. Survey responses can contain reporting errors. Platform samples can omit important populations. Agreement across different methods would strengthen confidence more than repetition of the same measurement.

A skeptical reading should therefore focus on the gap between observation and outcome. The release does not independently establish that a broader mix of requests produced better output, higher wages, or fewer jobs.

It also does not establish that absent activities are immune to future change. Their absence describes a measurement boundary. Predicting their future requires additional evidence about capabilities, deployment conditions, and organizational decisions.

The defensible conclusion is narrower and useful: researchers now have a more detailed basis for asking where consumer AI intersects with work.

Three signals will make the next release more informative

The next phase should connect breadth, measurement consistency, and outcomes. Over the next several months, those are the signals that will strengthen or weaken the interpretation of broader workplace use.

First, watch for persistence within comparable country and activity groups when OpenAI next updates the series. A continuing pattern under consistent definitions would strengthen the case that the mix is changing durably.

A reversal would require closer examination. Seasonal assignments, changing users, or changes in how activity is captured could explain temporary movement. Researchers should investigate those possibilities before attributing a trend to lasting changes in work.

The timing of any future update should not be assumed from the current release. The important signal is the comparability of the next observation, whenever it becomes available.

Second, watch for clearer connections between consumer chat, enterprise systems, and specialized interfaces. Vendor publications that explain those boundaries would make migration between products easier to interpret.

If activity declines in one interface while growing elsewhere, the broader story changes. What initially resembles reduced interest might instead reflect a better fit between a task and a different environment.

Consistent definitions would strengthen that analysis. Incompatible categories would leave uncertainty about whether researchers were observing behavioral change or a change in measurement.

Third, watch for independent evaluations that pair particular activities with completed-work outcomes. Useful measures include accuracy, review effort, completion time, and whether employees retain the ability to evaluate the result.

Evidence of improvement across the full assignment would strengthen the economic significance of broader use. Evidence of substantial correction work would qualify it. Either result would help buyers make better decisions.

This final signal matters most to people responsible for actual deliverables. A growing range of requests makes AI relevant to more assignments. It does not remove the need to determine what happened after the conversation ended.

For readers assessing ChatGPT’s expanding world of work, start with an activity you recognize in your own job. Compare the question an employee brings with the evidence needed to finish the assignment. Ask where the assistant contributes, where verification remains necessary, and which result would justify continued use. Keep those questions attached to the source material rather than the fluency of the answer. Then look for research that measures the entire process, including corrections and human judgment. The release offers a starting point for that investigation. Which observed activity would become meaningfully more valuable if researchers could show that people completed it more accurately, with less effort, and with responsibility clearly assigned?