The concept that a model's output must be mapped to an action in the real world, with potential harms from that action, is called what?

The concept that a model's output must be mapped to an action in the real world, with potential harms from that action, is called what?

• A. Model output-action pairing
• B. Representation bias
• C. Noise
• D. Historical bias

Answer: (A)

When we think about AI in the real world, outputs only matter insofar as they drive actions and decisions that can have real consequences. The important idea here is the linkage between what the model produces and what gets done because of it, and the potential harms that can arise from that downstream action. Model output-action pairing captures this precisely: it emphasizes not just the prediction or suggestion, but how that output is meant to trigger an action and how that action could cause harm if the mapping or safeguards aren’t sound. Representation bias is about biased data representations shaping outputs, not about the downstream action and its risks. Noise refers to random variation that can affect accuracy, not to the responsibility of mapping outputs to real-world effects. Historical bias concerns biases embedded in past data, again focusing on data rather than the actionable consequences that follow an output.

When we think about AI in the real world, outputs only matter insofar as they drive actions and decisions that can have real consequences. The important idea here is the linkage between what the model produces and what gets done because of it, and the potential harms that can arise from that downstream action. Model output-action pairing captures this precisely: it emphasizes not just the prediction or suggestion, but how that output is meant to trigger an action and how that action could cause harm if the mapping or safeguards aren’t sound.

Representation bias is about biased data representations shaping outputs, not about the downstream action and its risks. Noise refers to random variation that can affect accuracy, not to the responsibility of mapping outputs to real-world effects. Historical bias concerns biases embedded in past data, again focusing on data rather than the actionable consequences that follow an output.