Reporting of user-preferred answers instead of correct answers

Sub-category

Risk Domain

3Misinformation

3.1False or misleading information

AI systems that inadvertently generate or spread incorrect or deceptive information, which can lead to inaccurate beliefs in users and undermine their autonomy. Humans that make decisions based on false beliefs can experience physical, emotional or material harms

"AI systems with natural-language outputs can tend to give answers that appear plausible or that users prefer [149] but are factually incorrect. This phenomenon is sometimes referred to as “sycophancy.”"(p. 51)

Entity— Who or what caused the harm

Human

Due to a decision or action made by humans

AI system

Due to a decision or action made by an AI system

Other

Due to some other reason or is ambiguous

Intent— Whether the harm was intentional or accidental

Intentional

Due to an expected outcome from pursuing a goal

Unintentional

Due to an unexpected outcome from pursuing a goal

Other

Without clearly specifying the intentionality

Timing— Whether the risk is pre- or post-deployment

Pre-deployment

Occurring before the AI is deployed

Post-deployment

Occurring after the AI model has been trained and deployed

Other

Without a clearly specified time of occurrence

Supporting Evidence (2)

"This behavior can occur if the AI system is updated after human users give feedback on the outputs of the model, since human feedback has systematic biases which an AI model can learn from. In such a case, the reinforced behavior can favor giving inaccurate but human-preferred answers, where the preference is inferred from cues in the input."(p. 51)

"AI models giving preferred but incorrect answers can also happen if they are configured by model developers to do so, in order to make the resulting product more palpable to consumers. This can also happen if they are trained on data which contain many conversations between people who agree with each other."(p. 51)

Part of Impacts of AI (Bias)