📄 Prompt Stylometry for On-Device Affect-Adaptive AI: A Feasibility Study in Linguistic Signal Detection and Response Steering

Summary

The paper asks whether an AI assistant can read emotional and cognitive cues from how a prompt is written, not just what it asks, and adapt its reply without exposing the user to psychological profiling. The author calls this "Prompt Stylometry." It covers affect cues, such as sadness, anxiety or stress rated on a 1–7 Likert scale, and cognitive-style cues: self-critical, indecisive, exploratory, analytical or confident. Server-side processing could silently profile users this way, so the whole pipeline runs locally on a consumer device. The study uses 600 synthetic prompts covering 30 stylometric profiles, generated by Llama 3.3-70B. It compares lexicon-based (VADER), neural (GoEmotions) and generative (small LLM) annotation. Response steering is done at inference time through deterministic prompt assembly, with no fine-tuning, and is tested on two small model families under 5B parameters. Annotators and model families behaved differently in systematic ways. The results suggest on-device affect-adaptive AI is feasible but immature. Annotation sensitivity and the mismatch between human and synthetic text remain open problems.

Key points

• The paper names a privacy threat, "Implicit Profiling": any cloud LLM that handles prompts can infer users' mental states as a side effect. Fully on-device inference is proposed as the most direct countermeasure.
• The benchmark has 600 LLM-generated prompts across 30 profiles, annotated with VADER, GoEmotions and small LLMs. VADER has a single sentiment axis, so it cannot separate fine-grained negative emotions.
• The rule-based cognitive-style classifier was heavily skewed: it labelled 383 prompts analytical and 216 indecisive. The author calls this a "stylometric transfer gap," meaning psycholinguistic markers validated on human text do not carry over cleanly to LLM-generated text.
• Steering effects are hard to measure because of a floor effect. Baseline negative affect was about 1.1 on the 1–7 scale, and 84% of prompts were routed to low-distress quadrants.
• Stated limitations:
  • synthetic, single-turn data only;
  • models annotate their own inputs;
  • affect-shift is used as a proxy metric, not a measure of user benefit;
  • evaluation on a single hardware device.
• Proposed future work includes session-level tracking, IRB-approved tests with real users, and local layers that strip affect signals before prompts reach commercial APIs.