Human behavior variability reflects a dynamic internal 'state' that determines decision outcomes, and interventions targeted at that state can, in principle, control those outcomes; a new framework is supported by multiple literatures and two years of platform data from 200,000 users.

Main Finding

Human outcomes are not a fixed function of observable inputs; they are causally mediated by a dynamic, multi-dimensional latent “state” (a time-indexed weighting vector) that determines how events are evaluated and processed. Outcomes are controllable only conditionally—by intervening on the person’s state (or the timing of intervention) rather than by manipulating observable inputs or relying on verbal reasoning alone. Correlational personalization systems that ignore state and causal structure will systematically fail under intervention and distributional shift.

Key Points

• Definition of state: S_t is a latent weighting vector (biological, physiological, neuropsychological, contextual dimensions) that, together with event X_t and history H_{t-1}, causally determines decision A_t and outcome Y_t. State is dynamic (sub-daily to seasonal updates), multi-dimensional, and expressed as weights on features rather than fixed feature values.
• Causal claim: The mapping from event to outcome depends on state; observational associations over event–outcome pairs underdetermine interventional effects on outcomes. Predicting the effect of an intentional change requires explicit causal structure (Pearl-style intervention reasoning).
• Three propagation channels: the state affects outcomes through biology (HPA axis, hormones), physiology (HRV, sleep, energy), and neuropsychology (attention, appraisal, working memory). Each is in principle an intervention point; intervening on the conscious decision itself is usually the least efficient.
• State dynamics: updates operate at multiple speeds—slow (circadian/seasonal), medium (social events, exercise, meals), and fast (immediate event processing). The correlation structure among state dimensions varies across individuals/groups, producing large heterogeneity in intervention effects.
• Attentional bottleneck: enormous gap between sensory/neuronal input (~10^11 neural events/sec) and conscious reportable bandwidth (~40–60 bits/sec). What people report about their motives is compressed and state-dependent; verbal narratives are poor levers for efficient intervention.
• Synthesis of literatures: integrates causal inference, predictive processing/active inference, allostasis, chronobiology, computational psychiatry, and cognitive science to argue for operational state-aware systems.
• Normative/prescriptive claim: People can exert real, bounded control over outcomes by manipulating state and timing—this is actionable and measurable but requires state-aware design.
• Empirical status: This is a position/theoretical paper supported by motivating, aggregate observations from a deployed behavioural platform. It is not a pre-registered clinical study; an empirical companion paper is promised.

Data & Methods

• Approach: conceptual/theoretical synthesis plus formalization of “state” and its properties; the argument combines causal theory (intervention vs association), predictive-processing ideas (precision/weighting), allostatic biology, attentional bandwidth estimates, chronobiology, and computational-psychiatric generative-model thinking.
• Formal elements: explicit definition of S_t as a latent weighting vector; causal graph from S_t → (biology, physiology, neuropsychology) → A_t → Y_t, with feedback S_{t+1} ← Y_t. State-update operators specified qualitatively at different timescales (slow/medium/fast).
• Observational base: a 24-month dataset from a deployed behavioural platform with >200,000 consented users across four occupational personas (research 2023–2026). Observations are reported at aggregate, qualitative level to motivate the framework; no clinical instruments, validated psychometrics, or pre-registered hypothesis tests are presented in this paper.
• Limitations of empirical claims: authors explicitly disclaim claims about specific interventions’ efficacy; quantitative, pre-registered analyses are deferred to a future empirical companion paper.
• Deliverables in paper: seven testable predictions and six operational requirements for state-aware systems (enumerated in the full paper, but not validated here).

Implications for AI Economics

• For personalization and recommendation markets:
  • Correlational personalization (models trained on observed event–outcome pairs) is systematically mis-specified for interventions: firms that optimize for short-term engagement using observational correlations risk poor causal generalization and perverse effects under intervention or distributional shifts.
  • Platforms should invest in state-aware models that represent latent state dynamics and causal pathways; doing so improves counterfactual reasoning about interventions (e.g., timing nudges, physiological supports).
• For product design and service offering:
  • Effective behaviour-change products should target upstream state channels (sleep, HRV, nutrition, timing, attention scaffolding) rather than relying mainly on persuasive messages or framing. Timing of interventions (aligned with state trajectory/chronotype) is a high-value lever.
  • Heterogeneous treatment effects are expected due to individual differences in state-dimension correlation structure; one-size-fits-all A/B testing can be misleading without modeling state and timing.
• For measurement and evaluation:
  • Platforms must move beyond surface features and self-report as primary labels. Investment in unobtrusive state proxies (passive physiology, fine-grained temporal context, sequence histories) and in experimental identification of interventional effects is essential.
  • Evaluation protocols should test interventional distributions (do interventions on state produce predicted outcome changes?) rather than only predictive accuracy on observational data.
• For economic policy and regulation:
  • Consumer protection and regulation of “behavioural” digital services should account for causal opacity: optimizing platforms that ignore state-aware causal structure can produce harm via mis-targeted interventions. Regulators may require evidence of interventional safety/effectiveness, not just correlational performance.
  • Welfare-aligned objectives are necessary (the companion work’s requirement): economic incentives and market design should incorporate individual welfare constraints to avoid exploitative manipulation of state-dependent vulnerabilities.
• For labor and human capital economics:
  • Productivity and training interventions can yield better returns if timed and tailored to workers’ state trajectories (chronotype-aware scheduling, fatigue-aware task allocation).
  • Compensation and incentive design could factor state-dependent responsiveness—e.g., scheduling high-value tasks when state predicts higher responsiveness—raising ethical questions around employer control of worker states.
• For research priorities and investment:
  • High ROI for AI-economics research: building causal, dynamical models of human state (incorporating physiology and chronobiology) and conducting interventional experiments to estimate policy-relevant elasticities.
  • Need for multidisciplinary capacity (econometrics + causal inference + physiology + cognitive science) in teams designing personalization and intervention systems.

If you want, I can extract and summarize the authors’ listed six operational requirements and seven falsifiable predictions from the full paper (or from the companion paper) and map them directly to research and product priorities for platform firms, regulators, or economists.