AI-mediated choice and identity

AI-mediated choice and identity tracks how AI systems shape the preferences, choices, exploration patterns, and self-defining behaviors people enact when AI recommends, filters, or acts on their behalf. The construct is adjacent to work but not identical to work experience: the risk is that an agent can preserve apparent preference satisfaction while changing the variety and distinctiveness of what a person actually chooses over time.

2025-matz-basic-effect gives the wiki its first direct behavioral evidence. In a study of 110,000 real choices from 1,000 U.S. Facebook users, both generic and personalized GPT-4o agents selected less distinctive pages than a human baseline drawn from users’ own followed pages. The design matters because the agents were not choosing alien options; they were selecting between things the person had already liked, so the flattening mechanism can operate inside the person’s existing preference space.

The most useful contribution is the distinctiveness-diversity trade-off. Generic agents exerted the larger pull toward population-popular options, while personalized agents better preserved between-person distinctiveness but more strongly compressed within-person topical and psychological diversity. Personalization therefore does not simply solve agency loss; it may optimize toward the person’s most typical self and reduce exploration of less central but still real preferences.

For I-O purposes, this topic is a warning against treating agentic delegation as a neutral convenience layer. If AI agents increasingly schedule, purchase, communicate, triage, or choose on behalf of workers, then identity, agency, and behavioral variety become design outcomes alongside efficiency and accuracy.

2025-jiang-artificial-hivemind extends the concern from delegated choice to generated expression. The Artificial Hivemind effect suggests that open-ended AI outputs can become homogeneous across models even when no single answer is correct, and the paper’s human-annotation results show that aggregate quality measures can miss idiosyncratic human preferences. For identity-relevant choices and expression, the risk is not only that AI chooses popular options for people, but that it repeatedly supplies a narrow menu of acceptable ways to think, speak, imagine, or prefer.

Connections

• Relates to agentic-delegation because the effect appears when an AI system chooses on a person’s behalf rather than merely advising them.
• Relates to ai-mediated-work-experience because identity, autonomy, exploration, and felt ownership are part of how delegated AI can change the experience of a role or workflow.
• Relates to ai-receptivity because people may welcome personalized convenience even when the downstream behavioral pattern narrows.
• Relates to human-ai-task-allocation because deciding whether the human or agent should choose may depend on whether the choice is identity-relevant, preference-forming, or worth preserving as human exploration.

Contradictions & open questions

• 2025-matz-basic-effect shows agent choices over Facebook page preferences, not workplace task choices. The source is therefore better evidence for a mechanism of preference flattening than for direct workplace effects.
• The study does not observe whether users would accept the agent-selected choices in live life or work settings. Actual identity effects depend on uptake, trust, and repeated delegation.
• Personalization appears protective for interpersonal distinctiveness but harmful for intrapersonal diversity. That trade-off needs to be separated from broad claims that personalization is either autonomy-preserving or manipulative.