Quick answer
Five sources give projections, not facts. One paper puts the 10-year total factor productivity (TFP) gain at no more than 0.66%; a budget model puts TFP and GDP levels 1.5% higher by 2035; the Forum's Future of Jobs Report 2025 projects 170 million new jobs by 2030, 92 million displaced, not from AI alone.
Key takeaways
- Every number here is a source's projection, not this site's. Each is reported with who published it, when, what it covers and the caution the source itself gives.
- Penn Wharton says it builds on Acemoglu's framework and that its exposure input is almost twice his. Read a projection's inputs, not only its headline.
- BLS says its projections are not a forecast, and says the uncertainty about AI's impact remains very high.
- This site's rule, which is ours and not a forecast: decide now on what is measured today, review on a fixed date, and do not pay in advance for a promise.
What is the future of AI?
The 5 sources read here give projections, not facts: one economist's paper puts the 10-year gain in total factor productivity (TFP) at no more than 0.66%; a budget model puts TFP and GDP levels 1.5% higher by 2035. In the Forum's Future of Jobs Report 2025, as its release describes it, global trends in technology, economy, demographics and the green transition "are projected to generate 170 million new jobs by 2030, while displacing 92 million others": not AI alone. This site's rule: plan on what you can measure now.
They measure different things. Where "will" appears in a quotation, it is the source's.
What do named sources project, and how sure are they?
Five sources, five different quantities. One is a statistical agency, one a university budget model, one an economist's paper, one an employer-data report and one a survey of researchers.
| Source and date | Quantity | Horizon | Method | Its own caution |
|---|---|---|---|---|
| US Bureau of Labor Statistics (factsheet, page modified August 2026) | Employment by occupation and industry | 10 years, rolled forward annually | Historical trends and relationships projected forward, with specific assumptions | Not intended to be a forecast; the uncertainty about potential impacts "remains very high" |
| Acemoglu, NBER working paper (May 2024) | Cumulative gain in total factor productivity over 10 years; an upper bound: "no more than a 0.66% increase in total factor productivity (TFP) over 10 years" | 10 years | Task-based model using existing exposure and cost-saving estimates | Even these estimates "could be exaggerated" |
| Penn Wharton Budget Model (8 September 2025) | Level of productivity (TFP) and GDP | By 2035, 2055 and 2075 | Acemoglu's task-based framework plus a projected adoption timeline informed by the historical diffusion of comparable technologies | Based on "limited data on AI's initial effects" |
| World Economic Forum, Future of Jobs Report 2025 (press release, 7 January 2025) | Jobs created and displaced by 2030 from several drivers together, not AI alone, as stated in the Forum's release | By 2030 | Drawing on data from over 1,000 companies in 22 industries and 55 economies | No caution stated in the release; it gives no margin of error and does not say how the counts were derived. |
| Grace et al., arXiv (posted January 2024) | Surveyed researchers' estimated chance of unaided machines outperforming humans in every possible task, by year, among other predictions | Varies by question | Survey of 2,778 researchers who had published in top AI venues | Timing estimates are stated "If science continues undisrupted"; the abstract's uncertainty sentence concerns the long-term value of AI progress |
Read the "Quantity" column before comparing numbers; none is stated for one small business.
Why do two estimates built on one framework give different numbers?
Penn Wharton says it follows Acemoglu's framework, and that some of its inputs differ. Acemoglu's abstract says "predicted TFP gains over the next 10 years are even more modest and are predicted to be less than 0.53%", a lower figure than its 0.66% upper bound. The abstract says the lower figure reflects that early evidence comes from "easy-to-learn tasks, whereas some of the future effects will come from hard-to-learn tasks". The abstract does not state which economy the figure is for.
The Penn Wharton Budget Model says it combined "the task-based framework developed by Acemoglu (2024) with a projected adoption timeline informed by the historical diffusion of comparable technologies like the commercial web and cloud computing services". Its published sentence is: "We estimate that AI will increase productivity and GDP by 1.5% by 2035, nearly 3% by 2055, and 3.7% by 2075." The page states these as levels: "TFP and GDP levels are 1.5% higher by 2035", with the yearly growth boost peaking at 0.2 percentage points in 2032.
Penn Wharton says its exposure estimate is "40 percent relative to his 20 percent" and that the difference comes from how it interprets another paper's automation exposure metric. This page does not judge which estimate is closer to right.
What do the labor projections measure?
BLS projects employment by occupation and industry over 10 years and says its projections are "not intended to be a forecast of what the future will be"; the Forum's Future of Jobs Report 2025, as its release describes it, counts jobs created and displaced by 2030 from several drivers together and not AI alone. They are not two estimates of one thing.
BLS says its projections cover 10 years and about 800 detailed occupations, and "are not intended to be a forecast of what the future will be but instead are a description of what would be expected to happen under these specific assumptions and circumstances". It adds that when the assumptions are not realized, actual values will differ. It says BLS "assumes that labor productivity and technological progress will be in line with the historical experience", while noting that recent developments in AI "have raised the prospect that the future rate of technological progress could be higher than in the past".
The World Economic Forum's Future of Jobs Report 2025 is described in the Forum's press release of 7 January 2025. The release says global trends in technology, economy, demographics and the green transition "are projected to generate 170 million new jobs by 2030, while displacing 92 million others", and that job disruption "will equate to 22% of jobs by 2030".
Two limits apply. The release names several drivers (technology, economy, demographics and the green transition), not AI alone, so these totals are not attributed to AI alone. It does not say how many of the companies are small.
What do the researchers' surveys say?
A survey of 2,778 AI researchers reports their predictions, not facts, and its abstract states the timing estimate with a condition: "If science continues undisrupted". In the survey by Grace and co-authors, the researchers had published in top-tier AI venues and gave predictions about the pace of AI progress. The abstract reports the respondents' aggregated answers.
The abstract says: "If science continues undisrupted, the chance of unaided machines outperforming humans in every possible task was estimated at 10% by 2027, and 50% by 2047." The quantity is the surveyed researchers' estimate of that chance by those years, with the condition attached. The abstract calls them "aggregate forecasts" and does not say there how the answers were combined. It is what surveyed researchers answered, not a forecast by the authors or by this page. The abstract says "The latter estimate is 13 years earlier than that reached in a similar survey we conducted only one year earlier".
On a separate question, the abstract says "Most respondents expressed substantial uncertainty about the long-term value of AI progress", and goes on to figures about good and bad outcomes. That sentence concerns the value of outcomes, not when milestones arrive.
What is measured today, and what does it not tell you?
In the Census Bureau's Business Trends and Outlook Survey, 24% (95% CI 23% to 24%) of US employer businesses with 1 to 4 employees reported AI use for the reference period 24 August to 6 September 2026. It is self-reported, with a response rate of about 10%. It is a baseline for that survey period, not a trend and not a target. Detail is on AI use by business size and industry.
How should a small business plan when the sources are uncertain?
This site's rule, which is ours and not a forecast: decide now on what you can measure, set a date to look again, and do not pay for a promise.
- Decide now, on facts you can check. Pick one task you already do, such as drafting replies. Try it on a monthly plan you can cancel. Judge it on your own time and error counts, not on a projection. This site's suggestion: decide before the pilot what you will count (tasks done, outputs kept after review, corrections needed), and keep the tool only if those counts justify its cost to you.
- Set the next look. Put a review on the calendar each quarter and at every renewal date. BLS updates annually, so a yearly check of its factsheet is reasonable.
- Keep data portable and do not prepay. Store customer lists, prompts and documents in exportable formats. Prefer monthly billing; if you sign annually, note the cancel date. Buy what works on your data this month, not what a seller says AI "will" do.
To pick that first task, see AI automation for small business: a practical 2026 guide.
What should you ask a seller who sells "the future"?
Ask for what exists today, dated, and for the source of every statistic. Use the columns of the table above.
- Which features can I use today on my own data, and which are "coming"?
- Who measured the statistic in your deck, in what population, over which years, and under what assumptions?
- Is it a projection, a survey of opinions or a measured result?
- What can I export if I leave, and what does cancelling cost?
A seller who cannot name source, date and assumption is giving an opinion.
Our data
In the Census Bureau's Business Trends and Outlook Survey, 24% (95% CI 23% to 24%) of US employer businesses with 1 to 4 employees reported AI use (reference period 24 August to 6 September 2026).
Method: National share for the 1-4 employee class, from the Census-published Business Trends and Outlook Survey table, period 202619, as registered in this site's reviewed study of 103 published sector-by-size cells. Self-reported use, employer businesses only, response rate about 10%, current question wording only and never joined to earlier periods. One cross-section: it is a measurement, not a trend.
FAQ
Does any of these sources say when AI will match people at every task?
Not as a finding. The Grace abstract says: "If science continues undisrupted, the chance of unaided machines outperforming humans in every possible task was estimated at 10% by 2027, and 50% by 2047." That is the estimate of 2,778 surveyed researchers who had published in top-tier AI venues, stated with that condition, not a forecast by this page.
Why do projections about AI disagree?
They measure different things over different periods with different methods. One estimates a productivity gain over 10 years, another reports jobs figures to 2030, another surveys researchers.
Should a small business wait for AI to settle before using it?
This site's view, not a forecast: waiting is a choice too. A small, reversible pilot on one task costs little to stop. This page advises against paying in advance because a seller describes the future.
How often should I revisit my AI plan?
This site suggests a fixed review date, for example every quarter and at each contract renewal.
How this page was made
Figures come from the five sources listed, opened on October 7, 2026 from outside the United States, and from this site's reviewed study of Census data. This page was drafted with Claude, an AI model made by Anthropic. Anthropic is not among the sources cited here, and no AI vendor's statement about its own plans is used.