I have Parkinson's, which has severely impacted my ability to type. I can think clearly and strategically (I'm a CEO), but voice-to-text produces fragmented, nearly incoherent output. Over the past few months, I've been using Claude intensively to bridge this gap - I provide broken, disjointed voice notes, and it translates them into coherent professional communications while preserving my strategic intent.
This created what feels like a natural laboratory for observing preference extraction under extreme information constraints.
The Setup:
Input: Severely degraded signal (fragmented sentences, missing context, unclear references)
Human goal: Complex, multi-layered intent (e.g., strategic legal communications, business negotiations)
AI task: Extract true preferences/intent and generate output that matches what I meant, not just what I said
Feedback loop: Immediate and high-stakes (real business/legal consequences)
What I've Noticed:
The AI performs surprisingly well at inferring intent from fragments
It seems to build a model of my goals across the conversation that goes beyond instruction-following
Success/failure is immediately obvious to me (unlike synthetic alignment tests)
The constraint forces the AI to do genuine interpretation rather than pattern-matching
My Question:
Has anyone in the alignment community studied assistive technology use cases like this as evaluation grounds for preference extraction? It seems potentially useful because:
The human's actual preferences are clear (to the human)
The signal degradation is real, not artificial
Success is measurable and high-stakes
It's iterative and naturally generates lots of data
Or am I just experiencing "sophisticated autocomplete" and overthinking it?
I'm curious if this is worth documenting more rigorously, or if there's existing work I should read that covers this. I have months of detailed transcripts if the data would be useful to anyone studying this problem.
Natural Experiments in Preference Extraction: LLMs as Assistive Tech