OECD AI Capability Indicators: What They Tell Us About Work

Start with the original question

The 2025 framework organizes AI capabilities across nine domains and five levels. It describes a way to compare system capabilities with human abilities, using a broader view than a single benchmark score.

Treat the publication year as part of the evidence. A framework published in 2025 is not a live test of every model available today. New releases, benchmark changes, and deployment conditions can change what a system does well.

A benchmark is not a workplace

A model may perform well on a bounded test while struggling with incomplete records, unusual cases, shifting requirements, or responsibility for a harmful error. Workplace performance also depends on tools, data access, review, and the cost of correcting mistakes.

For example, drafting a board memo requires language generation. Producing an accurate memo for a real board also requires trustworthy inputs, financial context, confidentiality, and someone prepared to defend the conclusion. A language capability level does not settle those questions.

Mapping capabilities to occupations adds judgment

An occupational description contains many tasks. Deciding which capability each task needs, how much time it takes, and what reliability is acceptable requires a mapping process. That process can introduce bias and uncertainty.

The Index uses capability research as context for its own occupation estimates and generated assessments. It should not be described as an official OECD occupation ranking, an endorsed tool, or a direct mechanical output of the OECD indicators.

Use the framework to ask better questions

Choose a task and identify what successful completion requires: reasoning, social interaction, physical action, factual recall, or coordination across several steps. Then test the complete workflow with an approved tool, using representative cases and explicit quality criteria.

A failed test does not prove a permanent human advantage. A successful test does not prove a job can disappear. Both are evidence about a particular task under particular conditions.

Cite the source and the interpretation separately

When citing the indicators, name the OECD publication and its year. When citing our occupation score, name the Replaceability Index and the relevant page or dataset version. Keeping those attributions separate makes it clear where the primary research ends and our interpretation begins.

The methodology page links to the original publication alongside BLS, O*NET, and Eloundou et al. These sources provide different kinds of evidence; none alone establishes a personal employment forecast.

Frequently asked questions

Is the Index affiliated with the OECD?

The OECD publication is a source reference. The Index is an independent product and its scores are not OECD-endorsed.

Can the 2025 indicators establish what all current AI systems can do?

No. They are a dated research framework. Current claims require current, task-specific evidence.

Sources and methodology

The score is a relative index of task exposure from 0 to 100. It is not a percentage chance of losing a job or a forecast of layoffs. Individual tasks, employer choices, adoption, and regulation all matter.

The Index uses stored BLS-based occupation data. O*NET, OECD capability indicators, and task-exposure research are contextual references. The derivation of the stored occupation scores is not reproducibly documented in the published dataset.

The source snapshot does not record the wage-survey year or employment-projection window. Check the original BLS release before treating these figures as current. Editorial changes and export dates do not establish data freshness.

• Bureau of Labor Statistics Occupational Outlook Handbook
• O*NET occupation and task data
• OECD AI Capability Indicators (2025)
• Eloundou et al. (2023), GPTs are GPTs

Read the methodology and limitations