Model
A trained algorithmic artifact that performs tasks using learned parameters.
How the laws define this
Regulators consistently treat a model as a mathematical, statistical, or algorithmic artifact that processes input data to generate outputs such as predictions, classifications, decisions, or generated content, serving as a discrete component subject to governance, testing, and oversight. The central division lies in whether a model must be strictly quantitative or trained. The Supervisory Guidance on Model Risk Management restricts its scope to quantitative methods processing data into quantitative estimates. Conversely, Guideline E-23 — Model Risk Management (Office of the Superintendent of Financial Institutions) explicitly expands coverage to both quantitative and non-quantitative systems, while the Policy for the Development of Artificial Intelligence in Poland to 2030 – draft includes symbolic approaches alongside statistical and machine-learning techniques. Furthermore, instruments like the Draft Law (Bill) 'On Artificial Intelligence', the Interim Generative AI guidance for the public service, and the National Artificial Intelligence Strategy — Working Document for Citizen Participation 2021–2026 focus specifically on constructs produced or configured through training on datasets.
Synthesised from the 6 statutory definitions below. Each one is quoted in full, with its source.
How jurisdictions differ
What is distinctive in each definition, binding law in force first. Showing 5 of the 7 instruments that define this term; all are listed under Definitions below.
| Jurisdiction | Instrument | Status | What is distinctive |
|---|---|---|---|
| Kazakhstan | Draft Law 'On Artificial Intelligence' | In Force (Amended) | Treats the model explicitly as a discrete legal object for determining regulatory obligations under Kazakhstan's law. |
| Canada | Guideline E-23 — Model Risk Management (Office of the Superintendent of Financial Institutions) (final) | Awaiting Entry | Covers both quantitative and non-quantitative systems, including AI/ML, used for analytics, decisioning or operational support in Canadian financial institutions. |
| Singapore | Guidelines on Artificial Intelligence Risk Management | Awaiting Entry | Notes that models may be embedded within broader systems and can be internally built or supplied by third parties in Singapore's financial context. |
| United States | Supervisory Guidance on Model Risk Management | Superseded | Specifies three components—information input, processing, and reporting—and describes model risk management. |
| Peru | National Artificial Intelligence Strategy — Working Document for Citizen Participation 2021–2026 | Draft | Treats models as distinct artifacts for documentation and governance purposes within Peru's draft national AI strategy. |
What it triggers
Definitions (7)
A mathematical or algorithmic construct trained or configured to process input data and generate predictions, classifications, decisions, or other outputs; includes parameterized representations and learned weights and is treated as a discrete legal object for determining regulatory obligations.
The trained algorithmic artefact (including architecture, parameters/weights and behaviour) that produces outputs—predictions, classifications or generated content—based on input data and its training; treated as a discrete component requiring testing, versioning and governance.
An application of theoretical, empirical, judgmental assumptions or statistical techniques (including AI/ML) that processes input data to generate results; this definition covers both quantitative and non-quantitative systems used for analytics, decisioning or operational support.
A model is defined as a quantitative method, system, or approach that applies statistical, economic, financial, or mathematical theories, techniques, and assumptions to process input data into quantitative estimates. It includes three components: information input, processing, and reporting, and is broad enough to cover traditional models and AI/ML systems.
Algorithms or mathematical representations produced through training on datasets (e.g., ML models, statistical inference engines) that are used to perform inference or support decision-making within AI systems; treated as distinct artifacts for documentation and governance purposes.
Defined broadly to include foundation models and task‑specific models, encompassing statistical, machine‑learning and symbolic approaches used to produce AI capabilities. The draft treats 'models' as artefacts subject to governance, deployment inventories, testing and conformity pathways aligned with EU processes.
A model is a method or approach that converts assumptions and input data into estimates, predictions, recommendations, or decisions. In the MAS framework, models may be embedded within broader systems and can be internally built or supplied by third parties.
Not to be confused with
An AI model is specifically the mathematical engine within an AI system, while model refers more broadly to trained algorithmic constructs.
Questions and answers
What makes a model a legal object in Kazakhstan?
Kazakhstan's Draft Law 'On Artificial Intelligence' treats models as discrete legal objects. This determines regulatory obligations for AI systems.
Do non-quantitative systems count as models under Canadian guidelines?
Guideline E-23 from Canada's OSFI covers both quantitative and non-quantitative systems as models. This includes AI/ML used for analytics, decisioning or operational support.
How are models governed in financial institutions?
Singapore's Guidelines describe models as potentially embedded in systems from third parties. Canada's and US guidance set expectations for governance and risk management.
Related Terms
Models
Trained algorithms used to generate inferences or predictions....
AI model
The mathematical engine within an AI system....
Trained model
The model artifact produced by training on a dataset....
Model owner
Entity that owns or controls an AI model's rights and deployment....
AI systems
Systems using automated models to make/assist decisions or generate content....