AI Influence: Mechanisms, Amplifiers, and Consequences

TMLR Paper6442 Authors

Abstract: AI influence refers to AI's impact on the knowledge and values of individuals by acting as producers, mediators, and receivers of information. As a result, it impacts our collective processes of creating and spreading knowledge, forming beliefs, and reaching consensus. We argue that there are mechanisms of inconspicuous influence in AI development and deployment pipelines, which, when amplified by societal dynamics, could lead to dangerous outcomes that we may reverse by early interventions. We detail those mechanisms, amplifiers, and potential long-term consequences.

Submission Type: Regular submission (no more than 12 pages of main content)

Changes Since Last Submission:

The largest structural changes include the following.
- Newly created **Section 2** that specifies the scope of systems we consider, formally explains the absence of AI Influence in the formalisms of these systems, and decomposes the impact of such missing influence into three components, each corresponding to mechanisms, amplifiers, or consequences.
- Newly created **Figure 1** that illustrates the precise and formal distinction between mechanisms, amplifiers, and consequences, and the reason why they form a complete picture.
- Newly created **Table 1** that compares the formalism of language models training (supervised/reinforced), RecSys training, and the construction of knowledge-based systems. They serve as canonical examples in the scope of systems that we consider.

In our response to each reviewer, we have explained other targeted improvements, and how these changes address their concerns.

Assigned Action Editor: ~Pin-Yu_Chen1

Submission Number: 6442