Concept Contrast Fable

General-purpose AI and foundation models: one depends on how regulators classify them, the other on how technical teams train them

Conclusion first:Regulatory documents and procurement contracts alternately use 'general-purpose AI' and 'foundation model,' but they are two completely different classification systems. GPAI is a regulatory category, depending on the intended use scope of the AI system and the level of legal obligations; foundation model is a technical description, depending on whether the model is pretrained on large-scale data and has multi-task transfer capability. The same model can simultaneously be a foundation model, but whether it is classified as GPAI differs across jurisdictions.

Regulatory documents and procurement contracts alternately use 'general-purpose AI' and 'foundation model,' but they are two completely different classification systems. GPAI is a regulatory category, depending on the intended use scope of the AI system and the level of legal obligations; foundation model is a technical description, depending on whether the model is pretrained on large-scale data and has multi-task transfer capability. The same model can simultaneously be a foundation model, but whether it is classified as GPAI differs across jurisdictions.

Fable

Two departments label the same machine

A heavy machinery factory shipped a new engine. This engine does not belong exclusively to any one vehicle—it is designed to fit into the chassis of trucks, ships, generators, and agricultural tractors. The engineering department marked the design characteristics on the drawings: bore, displacement, fuel compatibility, and output power curve.

After the engine went on the market, the market regulator sent people to inspect. They first looked at the cylinder bore and fuel curve, then pressed further on what occasions this engine would be used. If a customer bought it to install in a fishing boat, it had to be tested under marine engine emission standards; if a customer installed it in an agricultural tractor, a different set of agricultural machinery safety regulations applied.

The factory director was completely lost the first time he attended a compliance meeting. Engineering said, "We build a general-purpose engine," and Legal said, "So it may fall into six different regulatory categories." The same engine had the label "general-purpose" from Engineering and "routed by end use" from Legal.

The conflict escalated when exporting to the EU. The EU importer required a 'general engine compliance declaration', but EU rules actually ask two different questions: first, is this engine restricted by design to a single scenario? Second, is it manufactured through a large-scale production line and adaptable to multiple vehicles? The answer to the first question determines which article of which set of regulations applies, and the answer to the second determines whether additional factory inspection obligations are triggered.

The company later created two label cards for each engine in its internal ledger. The first was called the “factory technical card,” recording bore, displacement, fuel compatibility and output curve—numbers fixed at the factory. The second was called the “market compliance card,” recording the current export destination, installation scenario and applicable regulations; this card updates as customer use and regulations change.

This dual-label system saved them during a major crisis. A certain country suddenly raised the import threshold for general machinery, requiring an additional risk impact assessment. Other manufacturers hurriedly dug through blueprints to find the definition of “general,” but only this factory directly filtered the engine numbers that triggered the conditions from its market compliance cards and submitted all materials within three days.

From then on, every internal training emphasized the same line: the factory technical card answers how this machine was built, and the market compliance card answers how the law views it. The same engine installed on a fishing boat today has one set of requirements, and installed in a tractor tomorrow has another; technical parameters don't change, but regulatory classification must follow the use.

General-purpose AI

A regulatory concept referring to AI systems designed for broad use and adaptable to many downstream tasks. Regulations focus on the systemic risks that the scope of use may bring, and definitions differ somewhat across jurisdictions.

Foundation model

A technical concept referring to a model pretrained on large-scale data that can serve as a starting point for many downstream tasks. The core of the definition is the training method and transfer capability, which belongs to a different dimension from regulatory classification.

Story Mapping

Engineering department's design parameter card: the base model, describing how the model was trained and what capabilities it has
Market regulation classifies by end use: GPAI, categorized by the intended use scope under the regulation
The EU asked two different questions: technology and regulation are two separate classification logics
A dual-label system averted a crisis: maintaining technical labels and regulatory labels separately enables rapid response to regulatory changes

Applying it to compliance document management

Maintain two independent information cards for each model. The technical information card records the model architecture, pretraining data scale, and training method, for internal selection and performance evaluation. The regulatory information card records the current regulatory classification, compliance obligations, and expiration dates for each target market, refreshed as regulations update. In procurement contracts and external statements, note whether the current page cites the technical definition or the regulatory classification, to avoid being cross-examined during audits.

General-purpose AI Foundation model