From Model to Operation: Who Controls the Risk?

An openly released GPAI model becomes part of a different accountability problem when it enters a real workplace, hospital, bank, or public agency. This chapter focuses on the deployer—the person or organisation using an AI system under their authority—and asks what they control that the original model provider cannot. Following the system into operation shows why model-level relief does not settle context-specific risk, human oversight, monitoring, or intervention.

The crucial distinction is between capability and control. A model provider controls choices made in development and release. An integrator controls how the model is embedded and what the finished system is intended to do. A deployer controls the environment in which people use the system and decisions are made. Those forms of control overlap, but they are not interchangeable.

Deployment changes the question

At the model layer, a reader asks what the model is, how it was released, and what the provider must document. At the deployment layer, the questions become operational: who uses the system, what data enters it, what decisions does it influence, who checks its output, and what happens when it behaves badly?

A hospital deploying an AI diagnostic tool has access to patient data, clinical workflows, and trained staff. A bank using an AI system for credit scoring knows the products, customers, and decision process in which the output will be used. A public agency using automated decision support knows the administrative setting and the people affected by it. The original model provider cannot fully see or control those facts when releasing a general-purpose model.

That difference explains why the source analysis assigns important responsibilities to deployers of high-risk AI systems. It describes duties involving human oversight, staff training, performance monitoring, log retention, transparency, following the provider’s instructions, and suspending use when a system is non-compliant or presents a risk. These are not simply additional documents about the model. They are controls over the system as it operates in a particular place.

The qualification matters. The selected sources do not establish that every use of an open model is high-risk or that every deployer has the same duties in every situation. The examples show why the layer exists. Whether a specific system falls within a high-risk category requires a separate analysis of its intended use and facts.

The deployer’s position is therefore not a fallback for every upstream failure. A hospital cannot reconstruct all design decisions made by a model provider. A model provider cannot perform every clinical check for every hospital. The value-chain arrangement works only if each actor takes responsibility for the part of the system they can actually observe and change.

Generic hazards become contextual risks

A general-purpose model can have broad, foreseeable hazards before anyone chooses a particular application. The scholarly analysis describes upstream providers as better placed to consider risks such as disinformation, toxic content, cybersecurity problems, or misuse potential. Those risks concern capabilities or tendencies that can appear across many settings.