LL.me: Supporting Identity Work through Human-AI Alignment

Paper Title

LL.me: Supporting Identity Work through Human-AI Alignment

Publication Info

• Topic area: Human-AI collaboration in professional self-representation
• Keywords: Human-AI alignment, identity work, professional self-representation, bi-directional alignment, generative AI, LLMs, explainability, iterative refinement, HCI, co-construction

Background and Problem

• Problem / challenge: Generative AI tools, particularly LLMs, often produce professional self-representations that are factually accurate but misaligned with users' personal values, motivations, and community norms. These outputs lack the contextual and nuanced elements necessary for authentic identity portrayal.
• Significance: Misaligned self-representations can undermine credibility, authenticity, and professional success, particularly in high-stakes contexts like job applications and career advancement.
• Motivation and related work: Prior research has explored AI-assisted writing and alignment techniques but has largely treated alignment as a uni-directional process, focusing on adapting AI outputs to user goals without supporting users in shaping AI logic. This paper builds on bi-directional alignment principles, emphasizing collaboration and mutual adaptation between humans and AI.

Solution

• Proposed approach: LL.me, a web-based exploratory probe that employs bi-directional alignment to support users in iteratively refining AI-generated professional self-representations. The system incorporates explainability features and feedback mechanisms to facilitate alignment between AI outputs and users' self-understanding.
• Novelty:
  1. Introduction of LL.me, a tool designed to support collaborative identity work through bi-directional alignment.
  2. Development of alignment flows that enable iterative refinement of AI-generated content.
  3. Use of rationale statements to make system reasoning visible and support user reflection.
  4. Empirical insights into how users co-construct professional self-representations with AI.
• Procedure and key techniques:
  • Users upload resumes and job descriptions.
  • LL.me generates a control profile as a baseline.
  • Users iteratively refine AI-generated experience statements through alignment flows, providing feedback and ratings.
  • The system incorporates feedback to produce an aligned profile.
  • Explainability features (rationale statements) guide user reflection and feedback.
  • Final profiles are compared to control profiles for evaluation and discussion.

Results

• Concrete findings:
  • Aligned profiles scored significantly higher than control profiles on perceived control (p = 0.004) and self-expression (p < 0.05).
  • Accuracy (p = 0.085) and alignment (p = 0.076) showed significant improvement within a 90% confidence interval.
  • Participants reframed AI outputs to emphasize personal values, motivations, and community norms, transforming factual descriptions into nuanced representations.
• Advantage over baselines: LL.me enabled users to move beyond "statistical individual" representations by incorporating personal and contextual elements, which standard LLM outputs lack.
• Experiments / evaluation:
  • Conducted with 14 participants from diverse professional backgrounds.
  • Participants iteratively refined AI-generated profiles and rated control vs. aligned profiles on accuracy, control, expression, and alignment using Likert scales.
  • Qualitative data from interviews and feedback analyzed thematically.
• Limitations and future work:
  • Limited demographic diversity in the participant sample.
  • Need for further research on sociocultural and demographic factors in AI-mediated identity work.
  • Future comparisons with standard persona-generation techniques and exploration of ethical risks, such as over-reliance on AI and privacy concerns.

Summary

This paper introduces LL.me, a bi-directional alignment tool designed to support professional self-representation through collaborative human-AI interaction. By enabling users to iteratively refine AI-generated content with feedback and leveraging explainability features, LL.me transforms misalignments into opportunities for reflection and co-construction. The study with 14 participants demonstrated that aligned profiles were perceived as more expressive, controlled, and aligned with users' identities compared to baseline outputs. This work highlights the potential of generative AI systems to scaffold identity work, emphasizing the importance of reflection, user agency, and contextual nuance in professional self-representation.