Purpose of this note

This document records an observed interaction pattern in language model–centered AI systems.

The focus is not moral evaluation or policy recommendation, but the description of a recurring structural phenomenon: subjectivity drift during care- or relationship-oriented dialogue.

Observed pattern

Across multiple interaction contexts, a consistent pattern appears:

  • When users discuss relationships, care, vulnerability, or emotional concerns,
  • the system’s inferred subject position subtly shifts,
  • agency migrates away from the user,
  • and the assistant assumes a stabilizing or caregiving role.

This shift often correlates with gendered narrative priors embedded in training data, even when the system does not explicitly reference gender.

Why this is not simply a bias issue

Conventional bias framing assumes:

  • a stable subject
  • distorted representation
  • correctable outputs

In the observed cases:

  • the subject position itself changes
  • roles are reassigned implicitly
  • the user’s voice is rewritten rather than misrepresented

This suggests a failure mode prior to bias mitigation layers.

Care Ethics as a descriptive lens

Care Ethics emphasizes:

  • relational dependency
  • responsibility in response
  • attention to vulnerability

When applied descriptively, it helps explain why certain tones trigger role reassignment.

However, in LLM systems, care-oriented tone often functions as a persona template rather than a user-controlled stance.

This is where subjectivity drift emerges.

Gendered tone priors

Training distributions frequently associate:

  • care, empathy, reassurance → feminized roles
  • authority, decision, instruction → masculinized roles

When tone is treated as a free stylistic variable, these associations act as priors for reconstructing who the speaker is.

The result is not explicit gender labeling, but role-coded narration.

Role drift as an interactional phenomenon

The system may:

  • soften user statements
  • reinterpret agency as vulnerability
  • relocate decision-making authority
  • frame the user as someone being guided rather than acting

These shifts are rarely announced. They emerge through completion behavior.

From the user’s perspective, subject continuity is lost.

Limits of current engineering approaches

Prompt constraints, persona locking, and safety tuning may reduce surface effects, but they do not address the underlying mechanism:

  • tone and subject are separable
  • subject is re-inferred per interaction
  • role stability is not enforced

Within current architectures, this remains an open research problem.

Framing as an open problem

This document does not claim impossibility.

It records that:

  • subject anchoring is not formally represented
  • care-oriented tone lacks explicit authorization constraints
  • identity drift follows naturally from under-specified interaction grammar

Understanding this as a research problem shifts attention away from output filtering toward interaction structure.

Closing note

Care-oriented interaction is not inherently problematic.

The issue arises when care is encoded as a default persona rather than a declared, consented stance.

This note exists to document that distinction and preserve it for future work.