Quick answer
None of the 4 sources we read names the jobs AI will not replace. One 2023 estimate, rated by human annotators and GPT-4, found around 80% of US workers could have at least 10% of their tasks affected by LLMs. The others count one chatbot's observed use or project employment, which BLS says is not a forecast.
Key takeaways
- No source cited here lists jobs that AI will or will not replace. Each measures something narrower: an exposure estimate rated by human annotators and GPT-4, observed use of one chatbot, or an employment projection that BLS says is not a forecast of what the future will be.
- The three task-based sources score tasks, not job titles. Microsoft researchers found widespread AI applicability across sectors, as most occupations have information work components.
- A high exposure score is not a forecast of job loss. Two of the sources say so in their own words, and BLS says its projections are not a forecast.
- Our own San Francisco study (419 occupations, observational) found no relationship between an occupation's share of one vendor's chat usage and how its local employment moved against the national count; a difference of up to about 10 percent between the least and most chat-used occupations cannot be ruled out.
- For a 1-50 person firm, the usable step is a task audit by role: which weekly tasks are text and information work, who checks the output, and what stays with a person.
What jobs will AI not replace?
None of the four sources below lists them: they estimate exposure to AI, count one chatbot's observed use, or project employment, and none counts jobs lost. The three task-based sources score tasks, not job titles. Microsoft researchers, for example, found widespread AI applicability across sectors, as most occupations have information work components.
The one list-shaped source is BLS's, and it is a list of where BLS says its AI-related changes are concentrated, not a list of the jobs that go. BLS says that while some changes are expected to reach across the economy, most are concentrated in four areas: computer and mathematical occupations; arts, design, entertainment, sports and media occupations; various business administration occupations (SOC 13, 41 and 43); and legal services (an industry, NAICS 5411).
What did the main sources actually measure?
Four named sources, four different measures, and none of them counts jobs lost. Read the third column before any number.
| Source and year | What it measured | Method | What it does not say |
|---|---|---|---|
| Eloundou et al., authors at OpenAI and the University of Pennsylvania, 2023 | Exposure of occupations' tasks to large language models: whether access to the model or software built on it would cut the time for a task by at least 50 percent | Rubric applied by human annotators and GPT-4; the authors call it an estimate of technical capacity (not observed use) | Says technical feasibility "does not guarantee labor productivity or automation outcomes" and "We do not make predictions about the development or adoption timeline" |
| Tomlinson et al., Microsoft Research, 2025 | An applicability score per occupation: whether there is a non-trivial amount of Copilot usage that successfully completes activities relevant to it | 200k anonymized US Bing Copilot conversations from January 1 to September 30, 2024, classified by an LLM pipeline (observed use of one chatbot) | Says that concluding high AI-action-score occupations will be automated and lose jobs or wages "would be a mistake" |
| Anthropic Economic Index, summary of the report sampling November 2025 | Share of jobs in Anthropic's sample where Claude is used for at least a quarter of their tasks (first report, January 2025 data; pooled across reports) | Claude.ai conversations (observed use, one vendor's product) | Authors say it only assesses tasks performed on Claude.ai and that it is "not always clear how these conversations might map onto changes in the real world" |
| US Bureau of Labor Statistics, AI factsheet, page modified August 27, 2026 | Where BLS says most of its AI-related changes to the 10-year employment projections are concentrated | Projection from historical trends, with adjustments from research applied conservatively | Not a forecast of what the future will be |
They are not four readings of one thing: one estimate, two observed-use counts, one projection.
The only low-end answer any of them gives is the 2023 paper's, under its rubric and date. It prints "Table 11: All 34 occupations for which none of our measures labeled any tasks as exposed." (Appendix D; the list opens with Agricultural Equipment Operators and Athletes and Sports Competitors, and it says nothing about demand or wages for those jobs.) It also reports that the importance of science and critical thinking skills are strongly negatively associated with exposure, which it reads as occupations requiring these skills being less likely to be impacted by current LLMs. Its own caveat applies: technical feasibility does not guarantee labor productivity or automation outcomes. Less exposed in this estimate is not the same as safe.
What does each source say?
Each measures something about tasks or projections, and each hedges it.
The 2023 study defines exposure as whether access to a language model or software built on one would cut a task's time by at least 50 percent, as rated by human annotators and GPT-4. It estimated that around 80% of the US workforce could have at least 10% of their work tasks affected, and approximately 19% of workers may see at least 50% of their tasks impacted. These are estimates of technical capacity; the authors add that technical feasibility does not guarantee labor productivity or automation outcomes.
Anthropic's report counts jobs where Claude shows up in a quarter or more of tasks: 36% of jobs in the sample in the first report (data from January 2025), 49% pooling across reports. It describes where one vendor's users bring work, not where employers replaced people. The authors say they only assess tasks performed on Claude.ai, and that it is not always clear how these conversations might map onto changes in the real world.
Anthropic also says that once it accounts for Claude's success rate, "we get a different picture of which jobs are most affected by the use of AI". So 36% and 49% are one measure, its success-weighted measure orders occupations differently, and both describe use of one product, not jobs at risk or safe.
What do government projections say?
BLS says its projections are not a forecast. Its factsheet states that the projections "are not intended to be a forecast of what the future will be" and that the uncertainty about potential impacts remains very high. The factsheet does not say that occupations outside the concentration areas listed above are untouched by AI.
What does this site's own data show?
Our San Francisco study found no link between an occupation's AI chat usage and how its local employment moved. In the study's words, we found no relationship between how much an occupation shows up in one AI vendor's chat usage and how San Francisco-area employment in that occupation moved relative to the national count (419 occupations, two published snapshots, May 2022 and May 2025); a difference of up to about 10 percent between the least and most chat-used occupations cannot be ruled out.
The vendor is Anthropic (Claude.ai chat and Cowork); the usage is national, measured after the employment window, and is a share of conversations mapped to tasks, not employer adoption.
This is an association among occupations, not proof that AI had no effect. A large or moderate association is ruled out; a small one, up to about 10 percent across the range, is not. The sample could reliably detect a correlation of 0.14 or larger, while small occupations' employment counts are noisy and the usage measure is coarse, so a small real association would be hard to see.
Method and limits are on the San Francisco jobs and AI chat usage study.
A second study, Entry-Level Jobs and AI: Is It Only Programmers?, looks at payroll employment by age band. It does not measure job loss or AI use, and neither study shows jobs lost to AI.
How do you apply this in a 1-50 person firm?
Audit tasks by role, because every source above measures tasks. These steps are this site's suggestion, not a finding from the studies.
- List what each role does in a typical week, as tasks, not duties. "Replies to quote requests" is a task; "customer service" is a title.
- Mark the tasks that are text or information work: drafting, summarizing, sorting, entering data, answering routine questions. Microsoft's paper reports that the most common and successful AI-assisted work activities in its Copilot sample involve information work.
- For each marked task, decide who checks the output, so an AI draft is never used unread.
- Separately list the tasks that are physical, done at a customer's site, or tied to a relationship you hold, so you can decide on them case by case. No source cited here shows these tasks are safe; absence from a study is not proof of safety.
- Revisit in six months; BLS says it releases new projections annually.
For hiring, this site suggests describing the role by its task mix and training people to review AI output. For what firms of your size already do with AI, see How Many Small Businesses Use AI?.
Our data
We found no relationship between how much an occupation shows up in one AI vendor's chat usage and how San Francisco-area employment in that occupation moved relative to the national count (419 occupations, two published snapshots, May 2022 and May 2025); a difference of up to about 10 percent between the least and most chat-used occupations cannot be ruled out.
Method: Share of one vendor's (Anthropic Claude.ai chat and Cowork) US conversations in April and May 2026 mapped to each occupation's tasks, against the change in San Francisco-metro employment minus the national change between the May 2022 and May 2025 BLS OEWS published snapshots; n = 419 occupations; observational.
FAQ
Is there a safe job that AI cannot touch?
None of the sources cited on this page says so. They measure how much of an occupation's tasks AI could affect, or how much a chatbot is used for them. A job title is not a safe category, and absence from a study is not evidence of safety; tasks are what get measured.
Does a high exposure score mean the job will disappear?
No. The 2023 study that introduced the exposure measure says technical feasibility does not guarantee labor productivity or automation outcomes. The Microsoft researchers call it a mistake to conclude that occupations with a high AI action applicability score will be automated and lose jobs or wages.
Do government projections say which jobs AI will replace?
No. The Bureau of Labor Statistics says its projections are not intended to be a forecast of what the future will be. It also says the uncertainty about potential impacts of AI remains very high.
Does this site have data on AI and jobs?
Yes. A reviewed study, linked in the body, compares 419 occupations in the San Francisco area with one vendor's chat usage and found no association; a difference of up to about 10 percent between the least and most chat-used occupations cannot be ruled out.
How this page was made
Figures come from the five sources listed, opened on October 7, 2026. This page was drafted with Claude, an AI model made by Anthropic, and checked against each vendor's published material on October 7, 2026. Anthropic is one of the sources cited (its Economic Index and the usage data in our San Francisco study), and its measure covers tasks on one product; it gets no more weight than the others.