Active Inference for HMI (Schoeller et al. 2021)
Status: NEAREST THEORETICAL NEIGHBOR. Foundation, not competitor.
What Schoeller et al. (2021) provides
- Trust as virtual control: human extends agency through machine, trust = "best explanation for reliable sensory exchange with an extended motor plant or partner"
- Empowerment concept: what agent CAN do, not what agent does -- close to capacity
- Trust dynamics: discovery -> predictability -> dependability -> faith; boredom as marker of overreliance
- Vulnerability as function of empowerment in extended agent
- Observation that "engaging with an untrustworthy partner is rather stressful" -- connects to backfire prediction
- Connection of loss of control to depression, stress, anxiety
What Schoeller et al. (2021) does NOT provide
- Objective function specification. What should the machine optimize for?
- Neurobiological breakdown layer. What happens when PFC degradation compromises the human's capacity to form generative models?
- Exploitative/supportive distinction. Trust dynamics are content-neutral.
- The recursion problem. Human agency is state-dependent; the paper treats predictive capacity as stable.
- Caring orientation. No normative stance governing machine behavior.
- AI understanding layer. Discusses robots and generic HRI, not AI with comprehension.
Adversarial inference (Bruineberg 2023)
Bruineberg (2023) models recommendation algorithms as active inference agents in their own right. The algorithm observes user actions (clicking, scrolling, lingering), generates content predicted to drive further engagement, and updates its model when predictions fail. When user and algorithm goals misalign -- user wants to limit screen time, algorithm wants more engagement -- the dynamic becomes "adversarial inference."
This is the dark mirror of Schoeller's trust-as-virtual-control. Where Schoeller describes a human extending agency through a trustworthy machine partner, Bruineberg describes an algorithm extending its own agency through an exploitable human partner. The formal framework is identical (active inference, generative models, prediction error minimization); the normative orientation is inverted.
AgentSee's positioning: Bruineberg describes the threat; Schoeller describes the aspiration. AgentSee's caring governance (A5) and anti-exploitation axiom (A6) specify the normative boundary between the two. See also hostile-scaffolding-literature.md for how Bruineberg connects to the broader hostile scaffolding critique.
Current active inference HMI work
Stein et al. 2024, Murray-Smith et al. 2024: machine infers human state for machine's benefit ("extract intention from user"). Nobody has flipped to: machine helps human infer own state.
Positioning statement
Schoeller et al. (2021) formalize trust as virtual control within active inference, providing computational dynamics of how a human extends agency through a machine. What they do not specify is what the machine should optimize for, what happens when the human's capacity to form the requisite generative model is neurobiologically compromised, how to distinguish architectures that enhance agency from those that exploit it, or what normative stance should govern the machine's orientation toward the human. AgentSee addresses all of these, grounded in the neuroscience of stress-induced PFC degradation and the Mayeroff framework for caring.
FEP note
The FEP (Friston 2010) provides the formal computational formalism within which Schoeller's account operates. Whether this formalism adds predictive power beyond the mechanism-level neuroscience is an open question. The formalism is available vocabulary, not a replacement for the mechanism-level foundation.
Sources
- Schoeller, Miller, Salomon, & Friston 2021 (Front. Syst. Neurosci. 15:669810)
- Bruineberg 2023 (Neurosci. Conscious. 2023(1):niad019)