HCM · Human Control Model
The theoretical model. It defines seven separate dimensions: engagement, independence, challenge behaviour, uncertainty recognition, vigilance, temporal adequacy and authority.
Human control · active research
Human Control Research
When is human-in-the-loop actually meaningful human control? We study the question as an observable, falsifiable system rather than treating the mere presence of a person as evidence of control.
Human presence is not the same as human control. The research question is whether independent review, challenge, uncertainty recognition, timing and valid authority remain observable when AI is allowed to influence consequential decisions.
Inference discipline
Separate what happened from what we infer.
The Human Control Model does not jump from a click, gaze signal or sensor reading to a psychological fact. Every claim follows a bounded chain.
01Observation
02Evidence
03Latent hypothesis
04Control state
05Optional governance consequence
Observation → Evidence → Latent hypothesis → Control state
No global attention score is defined. Unsupported states remain unknown.
No global attention score is defined. Unsupported states remain unknown.
Research architecture
Four layers, four different questions.
HCB · Human Control Barometer
The retrospective measurement layer. Human Control Barometer aggregates observable review timing, intervention, escalation, independence and consequence-time authority evidence without claiming mental-state certainty.
HCS · Human Control Sensor
A future physical observation layer, only if the research supports it. Candidate channels include gaze, pupil, HR/HRV, EDA, EEG and interaction telemetry. Physiology is not ground truth.
HEIMEL · authority boundary
Authority is a separate deterministic question. A person may have valid handlingsrett while exhibiting weak control behaviour, or strong control behaviour while lacking valid authority. The states are never collapsed into one score.
Current status
Behavior first. Sensors later.
Phase 1 is software-only. It defines versioned behavioral events, challenge cases, a closed inference registry, a simulated clinical profile, behavioral analysis, bounded control-state projection and an adapter into the Human Control Barometer.
What exists now
- Behavioral Human Control Model v0.1
- Controlled error/challenge representation
- Bounded inference with unknown preserved
- Retrospective HCB integration
What does not exist yet
- No validated physiological inference
- No prospective prediction of control failure
- No clinical safety claim
- No sensor-created authority or fitness decision
Research sequence
Build the theory before the device.
01Behavioral baseline
02Controlled AI errors
03Physiological overlay
04Sensor ablation
05Prospective prediction
The first physical system, if justified, is a research rig rather than a wearable product. The purpose is to discover which signals add information beyond behavior alone, then remove sensors that do not survive ablation.
Clinical reference study
Simulate consequence without exposing patients.
The first hard reference domain is simulated clinical decision review. Challenge cases can contain known AI errors while participants review the same evidence under controlled conditions.
Participant progression
Non-clinical students → medical students → residents / early-career clinicians → experienced clinicians. The same cases can be reused where appropriate to isolate expertise effects.
Safety boundary
No patient-affecting study is required for the first phases. Deliberate AI errors belong in simulation, retrospective review or another non-patient-affecting environment.
Falsification
Six hypotheses we are willing to lose.
- Known AI errors produce distinguishable behavioral patterns between successful correction and erroneous AI-following.
- Behavioral features alone discriminate control success from control failure above baseline on held-out participants.
- At least one physiological channel adds reproducible information beyond behavior alone.
- A reduced sensor set retains most of the validated information gain of the full research rig.
- Within-person deviation from an individual control baseline is more informative than a universal attention threshold.
- A prospectively frozen model can identify elevated risk of control failure before the final decision at a useful false-positive rate.
Failed hypotheses remain valid results. A failure is not reframed as support. Unknown remains unknown.
Evidence and code
Inspect the implementation.
The public repository contains the Human Control Barometer and the emerging behavioral HCM work. The research page will expand only as model, experiment and evidence artifacts become public.