Mimetic Alignment with ASPECT: Evaluation of AI-inferred Personal Profiles
New pipeline infers personal communication styles from behavioral data, achieving moderate alignment with self-assessments.
A team of researchers from Microsoft, including Ruoxi Shang and Denae Ford, has introduced ASPECT (Automated Social Psychometric Evaluation of Communication Traits). This novel pipeline addresses a core challenge for AI agents that communicate on behalf of individuals: capturing a person's authentic style without expensive, individualized model training. ASPECT directs large language models (LLMs) to assess established communication traits against behavioral evidence extracted from a person's workplace data, such as emails or messages, creating a detailed profile without any per-person fine-tuning.
In a case study with 20 participants, involving 1,840 paired item ratings and 600 scenario evaluations, ASPECT-generated profiles achieved moderate alignment with the individuals' own self-assessments. More importantly, responses generated based on ASPECT profiles were preferred over both generic AI outputs and responses based on simple self-reports. The system includes a crucial review phase where users can inspect the evidence behind each trait inference, helping them identify mischaracterizations and recalibrate their own self-ratings. This process enables negotiation of a context-appropriate AI representation.
The research, detailed in a 20-page arXiv preprint, demonstrates a shift towards inspectable and user-controllable AI personas. Instead of opaque models that generalize or require massive datasets for personalization, ASPECT offers a transparent, evidence-based method for scoping an AI agent's communication style. This has significant implications for building trustworthy AI assistants in professional settings, where accurate personal representation is key to effective delegation and collaboration.
- ASPECT creates communication profiles by analyzing workplace behavioral data with LLMs, eliminating the need for per-person model fine-tuning.
- In testing, ASPECT profiles showed moderate alignment with self-assessments and were preferred in 600 scenario evaluations over generic and self-report baselines.
- A review interface lets users inspect linked evidence for each trait, enabling them to correct AI inferences and control their agent's representation.
Why It Matters
Enables professionals to delegate communication to AI agents that authentically mirror their personal style, fostering trust and effectiveness in workplace AI adoption.