The Model Instinct: The Use Sets the Risk and the Value

All courses / Landscape / Module 2 / Chapter 7

Learning objectives

After the lesson the participant can:

• Explain the chapter's comparative claim: the model bounds what a use can do, while the use and usage mode set the risk and value.
• Describe what the AI Incident Database and the OECD monitor record, and three limits of the database's counts.
• Relate the roles of provider and deployer and the list in Annex III of the AI Act to use, as the chapter reads them.
• Draw two units that share a model provider in the supply and sovereignty view.

Key concepts

The model instinct

The habit of treating the model as the thing that carries the risk and the value: asking which model first and expecting the answer to say how dangerous, how useful and how well governed an AI use is.

The model bounds, the use decides

Models differ in training, openness of weights, where they are served and what they can do, so the model bounds what a use can do. The use and its usage mode decide what it does do, to whom and with what consequences.

Alleged deployer and developer

The fields the AI Incident Database carries instead of a model column, attributed by volunteer editors from press reports. The most-named deployers are scammers in 240 incidents, deepfake creators in 129 and synthetic media creators in 106.

Annex III as a list of uses

Under Article 6(2) of the AI Act the systems listed in Annex III are high-risk, and the annex lists uses, such as recruitment and the evaluation of eligibility for essential public benefits, not models. Article 6(3) narrows the class again by use.

The model chip

In the supply and sovereignty view the model appears as a small chip in its provider's container, joined by a thin line to each unit that depends on it. The model's name stays in the AI system register as text.

Interactive exercises

Do them here and get feedback at once. Choose a card and then where it belongs, or drag it.

Teaching plan

Discussion

1. The chapter grants the instinct a kernel of truth: some differences between models matter. For one use in your organisation, which property of the model bounds what the use can do?
2. Counts by year in the incident database reflect media attention and editorial throughput as much as the frequency of harm. What can such a record still tell a reader, and what can it not?
3. At Hansa Mutual a change of mode in one unit may be read as no change because the model is the same. How would a register of models and a register of uses each show that change?

Written exercise

Self-test

Choose an answer to see whether it is right, or open the answer below each question.

1. Under Article 25(1)(c) of the AI Act, as the chapter reads it, what follows when a deployer modifies the intended purpose of a system not classified as high-risk so that it becomes a high-risk AI system?

   1. The original provider keeps every obligation for the new purpose
   2. The system falls outside the Act until the next version of the model is released
   3. The deployer is then considered to be the provider of that high-risk system
   4. The general-purpose AI model inside it becomes a model with systemic risk
   Show the answer

   C. The deployer is then considered to be the provider of that high-risk system
   Article 25(1)(c) closes the loop: a deployer that changes the intended purpose so that the system becomes high-risk is considered its provider. The roles follow what a body does with a system, not which model is inside it.

2. In Figure 7.4, why does the mark of the conversational assistants HM-CONV sit inside a model provider's dashed container, while HM-CA and HM-AWT stay inside Hansa Mutual's own boundary?

   1. Because that provider operates the HM-CONV service, while the group operates HM-CA and HM-AWT itself
   2. Because HM-CONV uses a larger model than the other two units
   3. Because HM-CONV is a high-risk AI system under Annex III, and the notation places such systems with their provider
   4. Because HM-CONV has no seat and the other two have one
   Show the answer

   A. Because that provider operates the HM-CONV service, while the group operates HM-CA and HM-AWT itself
   A mark is placed with the supplier that operates it. HM-CONV sits in the second provider's container because that provider operates the service; HM-CA and HM-AWT are operated by the group and only run a line to the model chip.

3. The Stanford AI Index 2025 reported 233 incidents for 2024, and the snapshot of 2 November 2026 holds 313 for the same year. What limit of the AI Incident Database does this illustrate?

   1. The database counts each incident once per harmed party
   2. The database only records incidents that its editors can locate in the United States
   3. The database merges incidents and hazards into one count
   4. The record is revised retroactively, as reports for earlier years arrive
   Show the answer

   D. The record is revised retroactively, as reports for earlier years arrive
   The chapter lists three limits: counts reflect media attention and editorial throughput, the record is revised retroactively, and its one geographic field says little about Europe. The two figures for 2024 show the retroactive revision.