cs.HCMar 30, 2026
SaveSelf++: Co-Determined Agency for Human--AI Symbiosis in Extended Reality
Organizations: School of Product Design, University of Canterbury, Christchurch, New Zealand
Abstract
Self++ is a design blueprint for human-AI symbiosis in extended reality (XR) that preserves human authorship while still benefiting from increasingly capable AI agents. Because XR can shape both perceptual evidence and action, apparently 'helpful' assistance can drift into over-reliance, covert persuasion, and blurred responsibility. Self++ grounds interaction in two complementary theories: Self-Determination Theory (autonomy, competence, relatedness) and the Free Energy Principle (predictive stability under uncertainty). It operationalises these foundations through co-determination, treating the human and the AI as a coupled system that must keep intent and limits legible, tune support over time, and preserve the user's right to endorse, contest, and override. These requirements are summarised as the co-determination principles (T.A.N.): Transparency, Adaptivity, and Negotiability. Self++ organises augmentation into three concurrently activatable overlays spanning sensorimotor competence support (Self: competence overlay), deliberative autonomy support (Self+: autonomy overlay), and social and long-horizon relatedness and purpose support (Self++: relatedness and purpose overlay). Across the overlays, it specifies nine role patterns (Tutor, Skill Builder, Coach; Choice Architect, Advisor, Agentic Worker; Contextual Interpreter, Social Facilitator, Purpose Amplifier) that can be implemented as interaction patterns, not personas. The contribution is a role-based map for designing and evaluating XR-AI systems that grow capability without replacing judgment, enabling symbiotic agency in work, learning, and social life and resilient human development.
Figures & tables
Figure 1: The nine Self++ role patterns organised across three concurrently activatable overlays, with co-determination principles (T.A.N.) scaling in strength with overlay scope and initiative. Overlay 1 (Self): Competence support. R1 — Tutor : reduces novice uncertainty through a safe, learnable corridor (e.g., a trainee electrician receives anchored directional arrows, step gating, and ghosted hand exemplars through XR glasses while working on a residential electrical panel). R2 — Skill Builder : calibrates and generalises skill through variability and augmented feedback (e.g., a training doctor receives real-time motion traces and a holographic accuracy heatmap overlaid onto a practice mannequin during a surgical procedure). R3 — Coach : builds robustness under stress and supports self-correction (e.g., a cellist receives intonation feedback, fingerboard pressure heatmaps, and metacognitive prompts during a live performance, with social comparison replaced by private progression tracking). Overlay 2 (Self+): Autonomy support. R4 — Choice Architect : shapes the decision context while preserving authorship (e.g., a person views a floating AR monthly calendar where recovery weeks are gently highlighted and a friction gate requests confirmation before overriding rest days). R5 — Advisor : externalises deliberation by making counterfactuals and trade-offs inspectable (e.g., an ER doctor sees a branching holographic decision tree with uncertainty bands, survival-confidence estimates, and provenance badges distinguishing AI prognosis from attending physician input). R6 — Agentic Worker : executes delegated tasks under a proposal-approval loop with rollback (e.g., an air traffic control shift manager oversees an AI-drafted routing queue where conflict items are flagged and rerouted back for manual handling, with any clearance reversible before transmission). Overlay 3 (Self++): Relatedness and purpose support. R7 — Contextual Interpreter : makes identity, norms, and downstream impacts legible to reduce social surprise (e.g., a firefighter arriving at an incident sees AR-labelled crew roles, building entry points, and provenance badges distinguishing dispatch-confirmed from AI-inferred information). R8 — Social Facilitator : improves human-to-human coordination and repair (e.g., diplomats at a round-table negotiation receive personalised, opt-in AR overlays including speaking-time balance, perspective-invitation prompts, and neutral micro-summaries of each delegation’s position, while embodied virtual agents surface shared precedents as common ground). R9 — Purpose Amplifier : supports long-horizon value coherence by making future trajectories legible and editable (e.g., a retiring athlete views a holographic value-map converging personal strengths toward an aspiration, with a future-self contrast between drift and purposeful mentorship and editable identity-narrative fields).
| Lvl | Role | Role objective | Example XR-AI behaviours | Transparent | Adaptive | Negotiable |
| Overlay 1 (Self): Competence support (seconds to minutes) | ||||||
| R1 | Tutor | Reduce novice uncertainty; establish safe learnable corridor | Anchored arrows and ghosted exemplars with step gating; clutter suppression; completion detection with attention-aware pacing and corrective feedback | Cue provenance; disclose suppression; show limits | Fade prompts; retarget errors; adjust pacing | Pause/skip; show all vs minimal; override highlights |
| R2 | Skill Builder | Calibrate + generalise; variability with feedback, not scripting | Ghost tracks and shadow end-states with partial hints; performance analytics with adaptive hinting and controlled variability | Explain feedback basis; show comparison model | Increase task variability; withhold hints; change modality | User-set difficulty; toggle ghosts; consent for perturbations |
| R3 | Coach | Robustness under stress; self-correction; prevent brittle mastery | Fault injection and overlay removal with altered timing; safety/quality monitoring; targeted debrief with fall-back to R1/R2 | Disclose perturbation intent; disclose role/agency shifts | Adjust challenge intensity; adapt thresholds; taper monitoring | Opt-in for stress tests; emergency stop; hand-off confirmation |
| Overlay 2 (Self+): Autonomy support (minutes to days) | ||||||
| R4 | Choice Architect | Shape decision context (salience) while preserving authorship | Lightweight cueing with route salience (alternatives remain selectable); multi-criteria filtering; attention reweighting with trade-off previews | Mark nudges; link to goals; label optimised criteria | Update weights; fade as user internalises; reduce during load | Opt-out slider; consent for high-stakes; unnudged view |
Table 1: Self++ role patterns across overlays (concurrently activatable overlays) with example XR-AI behaviours and co-determination principles (T.A.N.).
| P | Proposition (what must be true) | Evaluation checks (what to test/measure) |
| P1 | Concurrency: Overlays act concurrently (not a pipeline) and can interfere. | Test overlap interference and recovery: (i) run Overlay 1 guidance while Overlay 2 deliberation UI is present (e.g., motor task + counterfactual dashboard) and measure errors/time-on-task; (ii) measure reclaim-time (time to pause/override after an AI-led phase) and success rate of taking back control. |
| P2 | Timescale Alignment: SDT needs map to uncertainty targets across temporal scales. | Evaluate on the right horizon: Overlay 1 with immediate sensorimotor metrics (errors, collisions, smoothness); Overlay 2 with decision quality and goal-alignment/endorsement (regret, confidence, stated-goal match over days); Overlay 3 with longitudinal drift indicators (relationship repair, wellbeing, dependence, value-consistency) over weeks/months—not only short task scores. |
| P3 | Inspectability: Legitimate augmentation requires an inspectable, contestable AI voice. | Probe legibility and ownership: users can say what was influenced (evidence/salience/delegation), why , and how to undo . Include a behavioural test: can users find and use “show alternatives/unnudged view”, “undo”, and “turn-off this support” controls successfully? |
| P4 | T.A.N. Scaling: Co-determination strength must scale with scope and initiative. | Audit T.A.N : higher-scope, Overlay 3 (R7–R9) must provide (i) stronger reason/provenance and incentive disclosure, (ii) explicit consent gates and clearer boundaries, and (iii) easier reversal/audit trails than lower-scope, Overlay 1 (R1–R3). Example check: an R9 narrative nudge must offer “see unframed view” and “disable this class of intervention”. |
| P5 | Transition Legibility: Shifts in agency between role patterns must be perceptible and reversible. | Test hand-offs and failures (Advisor Agentic Worker; escalation/de-escalation): users must correctly answer “who is acting now?”, “under what authority?”, “how do I stop/revert?” Measure situation awareness during transitions and recovery time/success after errors. |
| P6 | Endorsement over Compliance: Autonomy support preserves endorsed priors, not compliance. | Check internalisation, not just performance: users endorse outcomes as their choice and can explain “because…” in terms of their goals/values. Compare nudged vs unnudged conditions: if outcomes improve but endorsement drops or users cannot justify choices, autonomy support failed. |
Table 2: Self++ propositions (P1–P8) with brief evaluation checks. Use as a lightweight audit: map system features to Self++ role patterns, then verify T.A.N at the required strength and test transitions and drift under realistic overlap.