The short answer: the Anthropic Economic Index describes what conversations with Claude are about, matched to O*NET task descriptions. It does not say which professions use AI, and its methodology states outright that it supports no inference about displacement or job security. It is a good dataset for one question and a misleading one for two others.
What the index actually does
The Anthropic Economic Index takes a large sample of conversations, classifies what each one is about, and matches that content against the task descriptions in O*NET, the same federal catalogue the AI Job Risk Check score is built on. The result is a picture of which kinds of tasks people bring to the model, and how often.
In the May 2026 release, the occupational breakdown says that tasks commonly done in computer and mathematical occupations account for about 24% of matched usage. Content creation and copywriting is the largest single request category at about 23%. Education and learning is second at 13%.
The first misreading: usage as adoption
Twenty-four percent of matched usage does not mean programmers are 24% of users. The matching runs from what a conversation is about to the task catalogue, not from a person to a job. A student debugging a script, a marketer writing a formula, and a professional developer all land in the same bucket, because the bucket is about the task, not the person.
Anthropic says this plainly in the documentation. The chart still gets quoted as professional adoption, usually with a headline about which industries are "using AI most." The data cannot support that sentence.
The second misreading: usage as job risk
The methodology note is explicit that the data describes observed usage and supports no inference about displacement or job security. Three separate things get collapsed here routinely:
- How often a tool is used for a task
- How much of that task the tool could do
- What happens to the people who do it
The index covers the first one. The Science exposure ratings address the second. Nothing in either dataset addresses the third, and the gap between exposure and outcome is filled with adoption, regulation, quality bars, and demand.
Anyone handing you a job-risk claim sourced to a usage dataset has told you they did not read the methodology.
What it is good for
The index is a strong answer to the question it was built for: which kinds of requests are common. Content creation at 23% and education at 13% tell you where the model is being reached for today. The work, personal, and coursework split tells you how much of that reaching happens on the clock. Over time, the releases show which categories are growing.
That is useful context for a worker. If your occupation's tasks are heavily represented in usage, the tools are already in people's hands for work like yours, which says something about how quickly your employer's peers are moving. It still says nothing about your job's future, and it was never meant to.
How to read AI statistics in general
Before accepting any number about AI and work, ask which of the three things it measures: usage, capability, or outcome. Then ask what unit it uses: conversations, tasks, occupations, or people. The Anthropic index measures usage at the task level. The Science paper measures capability at the task level. Almost nothing credible measures outcomes at the person level yet, which is why the confident job-loss figures in circulation are worth treating with suspicion. For a task-level look at what people actually bring to these tools, see what people use AI for at work.
The dataset is published under CC BY 4.0 at anthropic.com/economic-index. It rewards reading in full.
Frequently asked questions
Does the Anthropic Economic Index show which jobs are at risk from AI?
No. Its methodology states that the data describes observed usage and supports no inference about displacement or job security. It measures what conversations are about, matched to O*NET tasks.
Does 24% of usage in computer and mathematical tasks mean programmers are 24% of users?
No. The index matches conversation content to task descriptions, not people to jobs. Anyone whose request resembles a programming task lands in that category regardless of their occupation.
What is the Anthropic Economic Index useful for?
Seeing which kinds of requests are common and how they change over time. In May 2026, content creation and copywriting was the largest category at about 23%, with education and learning second at 13%.
Get your free AI Job Risk Score. Tell us your job title and how you actually spend your time, and we will show you which of your tasks are exposed today. Free. 60 seconds.
Get my AI risk now