The OECD Framework for the Classification of AI Systems, published in February 2022 as OECD Digital Economy Papers No. 323, describes any AI system along five dimensions: People and Planet, Economic Context, Data and Input, AI Model, and Task and Output. It was written for policymakers, regulators, and legislators who need to compare very different systems, such as a virtual assistant, a self-driving vehicle, and a video recommender for children. The framework describes a system; it does not score, rank, or certify one.
Key takeaways
- Five dimensions organize the description of an AI system: People and Planet, Economic Context, Data and Input, AI Model, and Task and Output.
- The framework fits specific applications, such as a credit-scoring tool, better than generic labels such as "chatbot."
- In the OECD's own testing, people classified the first two dimensions more consistently than the technical ones.
- Some criteria apply before deployment ("in the lab"), and all apply after it ("in the field").
- The framework does not cover corporate governance, mitigation measures, or compliance and enforcement.
Why one word cannot describe the size of AI
"AI" covers a spam filter, a protein-structure predictor, and a loan-approval model. They differ in who is affected, what data they use, how they are built, and what they do with their outputs. The OECD paper starts from that observation: different systems bring different benefits and risks, so policy needs a shared way to say which kind of system is under discussion. The framework was approved and declassified by the OECD Committee on Digital Economy Policy on January 13, 2022, and it links system characteristics to the OECD AI Principles first adopted in 2019.
The five dimensions
Each dimension has its own set of properties, which the paper phrases as survey questions. The table paraphrases the core ideas.
| Dimension | Question it answers | Example properties |
|---|---|---|
| People and Planet | Who uses the system and who is affected? | User skill, impacted stakeholders, whether people can opt out or challenge outputs, effects on rights, well-being, environment, and work |
| Economic Context | Where is it deployed? | Sector, business function, for-profit or public service, criticality, breadth of deployment, technological maturity |
| Data and Input | What does it learn from and take in? | Provenance, collection method, structure, whether data are personal, static or real-time, quality and scale |
| AI Model | How is it built? | Rule-based, statistical, or hybrid; open or proprietary; generative or discriminative; whether it keeps learning in the field |
| Task and Output | What does it do? | Recognition, forecasting, personalization, or optimization; level of action autonomy; how output is evaluated |
The Economic Context dimension usually describes an application, not a generic system, which is why a model with no named use cannot be fully classified.
Lab versus field
The paper separates AI "in the lab", meaning conception and development before deployment, from AI "in the field", meaning use and change after deployment. Data and Input, AI Model, and Task and Output matter in both stages. People and Planet and Economic Context become meaningful mainly in the field, when real users and real organizations are involved. The distinction matters for risk management: checks done before release need different facts from monitoring done afterward.
A worked case from the paper: credit scoring
One of the systems the OECD classified is a credit-scoring tool. Per the paper, its users are bank employees, its impacted stakeholders are consumers and the bank, and applicants cannot opt out. It sits in the financial and insurance activities section of the international industrial classification, with broad deployment and the highest technology readiness level (an operational system in final form). Its model is hybrid, using human-written rules together with learning from repayment data, and it works as a forecasting tool with low action autonomy. The same method was applied to AlphaGo Zero, which differs on nearly every dimension.
What the OECD's own test found
The OECD invited stakeholders from government, business, civil society, and academia to classify sample systems through an online survey in June 2021. More than 850 surveys arrived, and 151 were fully completed. Four conclusions followed:
- The framework suits specific applications more than generic systems.
- Answers were more consistent for specific systems, such as credit scoring, than for general ones, such as voice assistants.
- People and Planet and Economic Context were classified more consistently than the technical dimensions.
- The technical dimensions required more information than is typically available, producing many uncertain or blank answers.
That last point is a practical warning: if a vendor cannot answer the Data and Input or AI Model questions, a classification will have gaps.
What the framework leaves out
The paper states that the framework does not address governance at the corporate, institution, or system level, nor mitigation, compliance, or enforcement across the lifecycle. It names uses such as algorithm registries, sector-specific criteria lists, and risk assessment work. It also notes that the UK Medicines and Healthcare products Regulatory Agency and NICE were adapting it to triage technologies for health technology assessment. Because it dates from 2022, it predates many generative AI products, and readers should apply its questions to newer systems with care.
This article summarizes a policy document. It is not legal or regulatory advice.
FAQ
What are the five dimensions of the OECD AI classification framework?
They are People and Planet, Economic Context, Data and Input, AI Model, and Task and Output. Each has sub-criteria that describe a system and its policy implications.
Is the framework a law or a certification?
No. It is a descriptive tool published by the OECD for policymakers and others. It does not create legal duties and does not rate a system as safe or unsafe.
Can it classify ChatGPT or a similar model in one pass?
Not well. The OECD found that generic systems yield less consistent answers than specific applications, so a chatbot would need to be classified for a particular deployment, user group, and task.
Who can use it?
Anyone describing a system, including regulators building algorithm registries and organizations preparing internal inventories. The paper names these uses explicitly.
Related reading
More in AI Basics.
Last verified: October 2026
Sources
- OECD Framework for the Classification of AI Systems, OECD AI Policy Observatory, 2022
- OECD Framework for the Classification of AI Systems, OECD Digital Economy Papers No. 323, OECD Publishing, February 2022