America Begins Counting AI in Hours
Every national statistic begins by deciding what deserves a question. Time diaries once made unpaid care, commuting and fragmented work visible in ways that payroll records could not. On 10 July 2026, the U.S. Bureau of Labor Statistics took an early step toward adding artificial intelligence to that tradition. Its Federal Register notice seeks comment on proposed AI questions for the American Time Use Survey, the federal diary that asks people how they spent the previous day.
The notice occupies only two Federal Register pages. It is a request for comment under the Paperwork Reduction Act, with submissions due by 8 September 2026. No new workplace right, reporting duty or AI rule takes effect. Yet measurement choices often arrive before policy choices. Once a category enters a national statistical instrument, researchers can compare it across occupations, households and time. AI may be moving from anecdotes about adoption into the slower, more consequential machinery of public measurement.
2003: a diary for work that payrolls cannot see
The American Time Use Survey began regular collection in 2003. Its distinctive method asks respondents to reconstruct one day, activity by activity. That produces information that employment records miss: time spent caring for children or adults, traveling, studying, exercising, shopping and doing household work. The survey has helped economists examine gender gaps in unpaid labor, the lived shape of unemployment and the difference between having a job and having usable time.
That history matters for AI because adoption is an incomplete measure. A company can buy an assistant without workers using it. An employee can use a public tool without a procurement record. A model can save ten minutes drafting and create twenty minutes of checking, correction or disclosure work. License counts and executive surveys will rarely show that exchange. A time diary can, provided its questions distinguish the activity with enough care.
Practical takeaway. The proposed survey questions could turn AI productivity from a broad adoption claim into a distribution question: whose time changes, which tasks disappear or expand, and where the gains and checking costs actually land.
10 July 2026: AI becomes a proposed survey category
The BLS notice is precise about its legal stage. The agency is conducting pre-clearance consultation under the Paperwork Reduction Act and requests comments on the proposed new collection, “American Time Use Survey Artificial Intelligence Questions.” It asks whether the collection is necessary, whether the burden estimate is accurate, how the quality and clarity of the information might improve, and how respondent burden might be minimized. The public comment period closes on 8 September 2026. BLS plans to field the questions from January 2027 for two years, producing the first nationally representative activity-level data on AI use.
Those procedural questions carry substantive weight. “Using AI” may describe writing with a generative assistant, accepting an automated schedule, receiving an algorithmic recommendation, monitoring workers through predictive software, or correcting machine output. Some uses are chosen by the worker; others are embedded in a platform. Some happen during paid hours; others shift preparation or verification into personal time. One broad yes-or-no item would compress these experiences into a category too coarse for serious analysis.
Time is also a distributional measure. An AI system may shorten a senior professional’s research while adding review work for a junior employee. It may speed an insurance decision while requiring a claimant to spend longer contesting an error. In medicine, an ambient documentation tool may return minutes to a clinician or introduce a new proofreading duty after the consultation. In public administration, automation may reduce processing time while moving explanation and appeal costs onto citizens. The same system can produce a productivity gain in one ledger and a time loss in another.
What the diary may reveal that adoption surveys miss
The first gain is occupational detail. National debate often treats “AI exposure” as a property of whole jobs. Daily work is made of tasks, and tasks change at different speeds. Lawyers may use AI for first-pass review while keeping client counseling human. Engineers may generate code quickly and spend longer testing dependencies. Teachers may draft materials with a model while fielding more disputes about authorship. Time-use questions can expose these mixed patterns if they connect tool use to the activity underway.
The second gain is a clearer view of unpaid correction. Automated systems create invisible labor when people verify hallucinated citations, repair formatting, repeat a failed application or explain a machine-generated decision. That work can fall on employees, contractors, patients, customers and family members. Economic accounts tend to reward the organization that records the saved minutes. A diary has a chance to find the minutes exported elsewhere.
The third gain concerns inequality. Access to high-quality tools, training and permission differs across workplaces. So does the power to refuse monitoring or challenge an automated instruction. If the survey can connect AI-related activities with occupation, employment status, caregiving and demographic variables under appropriate privacy safeguards, it could show whether the technology broadens autonomy or concentrates it. The answer is likely to vary by sector and by the kind of task.
Measurement design is already governance
A federal time diary carries unusual trust obligations. Respondents need to understand the question, remember the activity and feel safe answering it. Questions about workplace tools may touch employer policies, confidential work or fears of monitoring. The survey should collect enough detail to distinguish meaningful forms of AI use while avoiding requests for sensitive content, employer secrets or identifiable model interactions. That balance belongs to statistical design and to privacy governance at once.
Definitions will age quickly. Product names change, AI features disappear into ordinary software, and workers may not know when an automated system uses machine learning. A durable instrument will probably need functional language tied to activities, assistance and automated decisions. It should also preserve comparability across years. A question that tracks today’s chatbot vocabulary may become useless just as the most valuable trend line begins.
The problem resembles the measurement of remote work. A binary category can hide frequency, location, control and spillover into domestic life. AI use has the same texture. Minutes saved during paid work may reappear as after-hours checking. An employer-mandated tool and an employee’s informal assistant create different questions about responsibility. Statistical categories do not settle those disputes, but they decide whether later labor research can see them.
From productivity claims to public accountability
The proposal arrives amid aggressive claims about AI productivity. Vendors report faster drafting, coding and analysis; employers announce adoption programs; workers describe both relief and new supervision. Each account may be true in its own setting. National time-use data could add a common frame by showing whether changes persist beyond pilots and where saved time goes. Higher output, shorter hours, more checking and work intensification are different outcomes even when a dashboard labels all four “efficiency.”
This connects labor statistics with AI law and competition. Regulators assessing marketing claims need credible baselines. Employers negotiating workplace rules need to know which activities changed. Unions and worker representatives need distributional detail. Investors need a better distinction between purchased capacity and realized use. Public institutions need to understand whether automated services save citizens time or make access harder. A well-designed survey cannot answer every question, but it can make weak claims easier to detect.
The commercial implication sits in claim quality. Quentir’s Signature Brief adds fixed scope, a dated source spine, executive synthesis, an internal-use checklist and the complete Article 50 disclosure map. The connection here is narrow: measurement language, disclosure language and product claims all depend on categories that another person can understand and test.
The next important document may be the questionnaire
The Federal Register notice starts a public design period, not a settled methodology. The questionnaire that follows will determine whether the survey captures meaningful differences among assistance, automation, monitoring and correction. It will also determine whether unpaid and displaced time appears in the record. Those choices deserve attention before a clean national percentage gives them an authority they may not have earned.
How Quentir Reads It: measurement infrastructure is arriving before doctrine. Earlier Quentir analysis examined when an AI accuracy claim becomes part of the product. The BLS proposal addresses the upstream problem: public institutions need categories capable of testing what adoption and productivity claims mean in lived time. By 8 September, the most useful comments will likely focus on the boundaries of the category and the burdens hidden outside paid work.
A time diary cannot decide whether AI makes work better. It can show who received an hour, who lost one, and which tasks filled the gap. That modest power has civic importance. Democratic policy works better when technological change can be described in units people recognize from their own lives.
Sources: U.S. Department of Labor, Bureau of Labor Statistics, “Proposed Information Collection; ATUS Artificial Intelligence (AI) Questions”, 91 Fed. Reg. 42775 (10 July 2026), comments due 8 September 2026; Bureau of Labor Statistics, American Time Use Survey; Quentir public-source snapshot: 10 July 2026.
Published intelligence, built to inform your own decisions. Published: July 10, 2026.