RESEARCH PROGRAM
The Fluent Interaction Conjecture
Toward a unit of benefit for – human AI interaction.
Download the prospectus (PDF)The AI field measures capability with extraordinary rigor — and measures benefit to the individual person hardly at all. Where benefit is measured, the instrument is the thumbs-up: approval ratings and engagement. Approval is not benefit. It correlates with feeling agreed-with, not with being helped — and optimizing an invalid instrument produces flattery, not help. The remedy for a measurement artifact is not a constraint. It is a better instrument.
B = γ (A × L)
- A — Accuracy (-1 to + 1).
- How well the system models the specific person, scored against validated psychometric instruments. + 1 is a colleague who truly gets you; 0 is a stranger; -1 is someone confidently wrong about everything you are.
- L — Load (0 to 1).
- The share of a task's outcome that depends on communication style rather than raw correctness — measured, not asserted.
- γ — The context gate (0 or 1).
- Some communications should not be tuned to their recipient at all. The gate is the model telling operators where not to use personalization.
- B — Benefit (-1 to + 1).
- The measured difference between the personalized system and the same system running generic, for the same person on the same task.
- Where load is zero, accuracy buys nothing.
- Where accuracy is zero, load buys nothing.
- The sign of the benefit follows the sign of accuracy — a wrong model is harmless on a spreadsheet and expensive in a negotiation.
- The benefit of accuracy grows steepest where load is highest.
A flat, uniform benefit across task types would falsify the model, and we would report that.
| Level | The Work | Examples |
|---|---|---|
| FD0 | No human recipient; a verifiable right answer exists | Data transforms, tested code, reconciliation |
| FD1 | Human recipient, low stakes, conventional formats | Status updates, scheduling, routine summaries |
| FD2 | Routine persuasion and service; outcome depends on the recipient's reaction | Support resolutions, standard sales follow-up, onboarding |
| FD3 | Goal conflict present; style variance dominates outcomes | Negotiation, pushback, critical feedback, retention saves |
| FD4 | High-stakes relational work; trust is the deliverable | Leadership communication, conflict resolution, key accounts |
Score the work. Score the model's accuracy for the people involved. Deploy where accuracy meets demand — and never deploy unverified personalization into FD3 or FD4.
Stage 1
Calibrate the Load scale. Identical content rendered in varied styles, judged by blind recipients; a validated task taxonomy the field can use with or without us.
Stage 2
Blinded pilot. Participants work under three conditions — no profile, their true profile, a foil profile belonging to someone else. The foil is the control: if a wrong profile hurts where load is high, the benefit cannot be placebo.
Stage 3
Pre-registered confirmatory study. Protocol frozen and published before data collection; independent methodological review; results published regardless of direction.
This prospectus is circulated for critique before the protocol is frozen. Collaborators are sought in person perception, psychometric methods, interpersonal communication, and machine behavior.