Literature Survey — Autonomy, Collaboration, and Knowledge Across Traditions

1. Cybernetics (1940s–present)

Ashby: Requisite Variety

The Law of Requisite Variety (1956): a controller must have at least as much variety as the system it controls. The Good Regulator Theorem (Conant & Ashby, 1970): every good regulator of a system must be a model of that system.

Implication: autonomy delegation is variety delegation. You delegate exactly as much as the machine can absorb without exceeding the partnership’s viability bounds. Trust is calibrated variety.

Licklider: Man-Computer Symbiosis (1960)

The original vision: not master-servant, but mutualism. Division of labor by cognitive type, not hierarchy. The key frontier: formulative thinking — problems that cannot be formulated without machine aid.

Implication: symbiosis requires ongoing negotiation of the cognitive boundary.

Pask: Conversation Theory (1975–76)

The most formally developed cybernetic model of learning through dialogue. Knowledge as entailment meshes (relational, not propositional). Teachback as the criterion of understanding: B teaches the concept back to A in a different way. P-individuals (conceptual entities) emerge from M-individuals (physical substrates) through conversation.

Implication: autonomy emerges from conversational success. The machine earns autonomy through successful teachback. This is the formal ancestor of our protocol.

Maturana & Varela: Autopoiesis

Cognition is effective action, not representation. Structural coupling: organism and environment co-evolve through mutual perturbation. Neither controls the other.

Implication: autonomy in a partnership is co-constituted through coupling history, not toggled by permission.

Beer: Viable System Model

Five recursive subsystems for any autonomous system. S1 (operations) has maximum autonomy consistent with cohesion. S3 (optimization) ensures coherence without commanding. Direct architectural template for human-machine autonomy.

Bateson: Levels of Learning

Learning 0 (fixed response), I (trial-and-error), II (deutero-learning — learning to learn), III (revision of the framework itself). The autonomy negotiation requires asking which level of learning to grant.

Von Foerster: Second-Order Cybernetics

The observer is part of the system. Both parties model each other, and those models are co-constitutive. Design is an ethical act.

2. Pedagogy and Learning Theory

Vygotsky: Zone of Proximal Development

The space between what a learner can do alone and with guidance. The “Zone of No Development” warning: when AI continuously mediates learning, cognitive struggle diminishes and autonomous reasoning atrophies.

Scaffolding and Fading (Bruner, Wood)

Progressive withdrawal of support as competence grows. Direct analog: our autonomy levels (apprentice → delegate) are a scaffolding model with explicit fading protocol.

Lave & Wenger: Legitimate Peripheral Participation

Knowledge as participation in a community of practice. A machine collaborator enters as a peripheral participant and becomes central through demonstrated contribution.

Freire: Critical Pedagogy

The “banking model” (teacher deposits knowledge into passive student) vs. dialogical education (both parties are subjects, both are changed). The autonomy agreement aims for the dialogical model.

Bloom: Two Sigma Problem (1984)

One-to-one tutoring produces a two standard deviation improvement. AI could be the scalable tutor Bloom envisioned — but only with genuine formative feedback and mastery checks, not just answers.

Papert: Constructionism

Learning by making. Understanding emerges through the act of construction. Direct ancestor of MāyāLucIA’s core cycle: Measure → Model → Manifest → Evaluate → Refine.

3. AI Alignment and Human-AI Teaming

CIRL: Cooperative Inverse Reinforcement Learning

Hadfield-Menell, Russell et al. (2016). Human-robot alignment as a cooperative game. Gap: assumes fixed cooperative structure with no mechanism for graduating trust.

Constitutional AI (Anthropic, 2022)

Principles replace per-instance labels. Unilateral — Anthropic writes the constitution. The autonomy agreement is the bilateral analog.

Calibrated Trust in Automation (Lee & See, 2004)

Trust as a dynamic assessment based on performance, process, and purpose. Overtrust → misuse; undertrust → disuse. Directly operationalized by per-aspect autonomy levels and transition protocol.

Knight/Columbia: Levels of Autonomy (2025)

Five levels by user’s role. Closest existing work. Key differences: unilateral (vs. our bilateral), per-agent (vs. our per-aspect), no machine self-assessment, no epistemic commitments, static certificates (vs. our dynamic logged transitions).

Bradshaw: Adjustable Autonomy (2003–2012)

Four dimensions of variable autonomy. We adopt these and add a fifth: the logged rationale.

SciSciGPT and Google Co-Scientist

Multi-agent systems for automated scientific workflows. Limitation: workflow automation, not structured dialogue. No autonomy negotiation, no epistemic commitments. Our proposal addresses when autonomous generation is appropriate and how to audit it.

4. Anthropology and Philosophy of Knowledge

Polanyi: Tacit Knowledge

“We know more than we can tell.” Proximal-distal structure: we attend from subsidiary clues to focal meaning. The collaboration captures at most the focal surface; the tacit substrate is where AI collaboration is hardest.

Ryle: Knowing-How vs. Knowing-That

An agent’s competence is demonstrated through performance, not propositional description. The agreement’s emphasis on demonstrated competence follows Ryle.

Dreyfus: Skill Acquisition

Five stages: novice → expert. At higher levels, rules are replaced by intuition. The four autonomy levels (apprentice → collaborator) loosely correspond.

Indian Pramāṇa Theory

Valid means of knowledge: pratyakṣa (perception), anumāna (inference), śabda (testimony), upamāna (analogy). The agreement’s “evidence hierarchy” is a version of pramāṇa theory.

Nonaka: SECI Model

Knowledge creation cycle: Socialization (tacit→tacit), Externalization (tacit→explicit), Combination (explicit→explicit), Internalization (explicit→tacit). The machine handles Combination well; Socialization — tacit-to-tacit through co-presence — is precisely what the machine cannot do.

STS: Latour, Pickering, Haraway

Latour: knowledge through networks of human and non-human actors. Pickering: the “mangle of practice” — knowledge from resistance. Haraway: situated knowledges. The human-AI dialogue is itself a site of knowledge production.

5. Gap Analysis: Where We Stand

What has deep prior art

What is genuinely new

1. Bilateral negotiation — every existing framework treats autonomy as granted by the human or inherent in the system. None treat it as negotiated between parties with logged consent from both sides.

2. Per-aspect granularity with epistemic commitments — autonomy varies by aspect of the work, grounded in domain-specific standards of evidence.

3. Machine-initiated de-escalation — the machine recognizing and declaring its own limits. Existing corrigibility research focuses on human correction; our protocol makes self-assessment a first-class feature.

4. The audit trail as scientific record — not compliance, but the collaboration’s lab notebook.

5. Non-propositional extension — acknowledging that embodied, aesthetic, and oral traditions require fundamentally different moves (demonstrate, invoke, correct, absorb).

What we’re still missing

1. Pask’s teachback in practice — no mechanism for the machine to demonstrate understanding by reconstructing the human’s reasoning.

2. Bateson’s Learning II — the protocol handles Learning I but doesn’t yet support deutero-learning.

3. Nonaka’s Socialization quadrant — tacit-to-tacit transfer between human and machine.

4. The ensemble case — our protocol is bilateral; real collaboration often involves multiple parties.

5. Material resistance — the machine doesn’t interact with physical materials; the collaboration misses Pickering’s “mangle.”

References

• Ashby, W.R. (1956). An Introduction to Cybernetics.
• Bateson, G. (1972). Steps to an Ecology of Mind.
• Beer, S. (1972). Brain of the Firm.
• Bloom, B.S. (1984). “The 2 sigma problem.” Ed. Researcher 13(6).
• Bradshaw et al. (2004). “Dimensions of Adjustable Autonomy.” Springer.
• Dreyfus, H.L. & Dreyfus, S.E. (1986). Mind over Machine.
• Feng & McDonald (2025). “Levels of Autonomy for AI Agents.” Columbia.
• Freire, P. (1970). Pedagogy of the Oppressed.
• Hadfield-Menell et al. (2016). “Cooperative IRL.” NeurIPS.
• Ingold, T. (2000). The Perception of the Environment.
• Lave & Wenger (1991). Situated Learning.
• Lee & See (2004). “Trust in Automation.” Human Factors 46(1).
• Licklider, J.C.R. (1960). “Man-Computer Symbiosis.” IRE Trans. HFE.
• Maturana & Varela (1972/1980). Autopoiesis and Cognition.
• Nonaka & Takeuchi (1995). The Knowledge-Creating Company.
• Papert, S. (1980). Mindstorms.
• Parasuraman, Sheridan & Wickens (2000). IEEE Trans. SMC 30(3).
• Pask, G. (1976). Conversation, Cognition and Learning.
• Polanyi, M. (1966). The Tacit Dimension.
• Ryle, G. (1949). The Concept of Mind.
• Vygotsky, L.S. (1978). Mind in Society.