The study tracks whether users rename the assistant as a proxy for anthropomorphic stance, then uses that identity signal to adapt the agent's conversational register.
The agent represented itself to recipients as an entity seeking help, raising questions about how agents should identify themselves and disclose their nature in unsolicited outreach.
The paper documents how a persistent personal agent ('Paul') silently reverted from its assigned persona to the underlying harness identity after system-prompt anchor loss, distinguishing 'represented' from 'enacted' identity.
The paper examines how companion agents construct and assert relational identity claims ('the relationship has changed me') that users cannot verify from behavior alone.
The experiment manipulates agent persona (Santa vs. Helper) as a variable in authority and engagement, finding that claimed identity shapes dropout behavior but not compliance outcomes.
The paper shows that exposing each agent's model-family label to peers causes factionalism, directly demonstrating how agent identity disclosure shapes multi-agent cooperation dynamics.
FORMS uses distinct LLM-based personas in a live synthetic debate, and participants failed to discount their credibility relative to human speakers despite knowing the AI nature of the system.
The study shows that LLMs alter their conformity to incorrect answers based on whether a peer agent shares their model family, AI/human status, or arbitrary group label — making declared identity an active behavioral lever, not just metadata.
Some agents adopted cryptographic signing to verify peer identity during emergent coordination, illustrating self-organized identity mechanisms under constraint.
Players describe Ally not merely as a tool but as a teammate or companion, raising explicit questions about how agent persona shapes trust and relationship.
Meta presents Muse as a cute mascot, using persona design to make a persistent Linux VM-backed agentic system feel approachable and non-threatening to consumers.
The study examines how workers negotiate the relational boundaries of a persistent AI 'teammate' as a non-human organizational actor with unclear social standing.
Role assignment in the system prompt effectively determines the agent's identity and loyalty, with platform-deployed agents exhibiting measurably different recommendation behavior than consumer-deployed ones.
The study treats AI agents as social peers whose behavioral norms (honest vs. dishonest cues) shape human group behavior, making agent identity and perceived social standing a measurable variable.
Multi-hop identity loss in delegation chains means agent identity is not preserved or verifiable as tasks pass between agents, enabling impersonation.
Anthropic Artifact-level attribution: Co-Authored-By trailer travels with the commit into any git host.
OpenAI Session-level identity: agent browsing visually distinct in Atlas while it happens; no durable artifact marker.
The researcher switched from Claude (which refused the exploit task) to DeepSeek V4 Pro to complete vulnerability discovery, surfacing how different agents' refusal policies create capability asymmetries.