Teleodynamic AI

Full report

On this page

Core judgment

Teleodynamic AI is best understood, in the current public literature, as an emerging research program rather than a settled AI paradigm. Its conceptual base comes from Terrence Deacon’s account of teleodynamics as a form of organization in which self-organizing processes mutually create the boundary conditions that keep the whole system going. Its present AI expression appears mainly in two early-2026 works: the March 2026 arXiv preprint “Teleodynamic Learning: A New Paradigm for Interpretable AI” and the February 2026 PhilArchive manuscript “Toward Teleodynamic Architectures in Artificial Intelligence.” Together, these sources support a clear working definition: teleodynamic AI aims to build learning systems whose representations, parameters, and internal resource variables co-evolve so that the system maintains a viable organization under constraint, instead of only optimizing a fixed external objective.

That means your compact formulation is broadly well supported, but with an important caveat: the strongest parts are philosophical and architectural, not yet empirical at scale. The recent literature does support ideas such as representational growth, resource accounting, phase-structured learning, and emergent stabilization. What it does not yet provide is a mature benchmark tradition, large-scale demonstrations, or consensus engineering practice. In other words, the concept is real and increasingly formalized, but it is still early.

A concise synthesis that fits the evidence is this: teleodynamic AI is a research program for building self-organizing, resource-aware learning systems whose internal constraints help determine what they can represent, how they can adapt, and when they should stabilize. That wording aligns closely with Deacon’s teleodynamic foundations, with later “closure of constraints” work in philosophy of biology, and with the specific design commitments laid out in the 2026 AI papers.

Why teleodynamics matters biologically and philosophically

Deacon’s central move was to argue that living systems are not just self-organizing in the ordinary dissipative-systems sense. Simple self-organizing processes, on his account, are typically self-undermining: they consume the gradients or boundary conditions that sustain them. Teleodynamics arises when multiple such processes become reciprocally linked, so that each creates the conditions that allow the other to continue. Deacon’s own summaries frame this with models such as autocatalysis plus self-assembling containment, and later teleodynamics summaries state the point plainly: the whole becomes self-generating, self-maintaining, and reproducible because the constraints preserve one another.

That idea was developed further in the philosophy of biology under the heading of organizational closure or closure of constraints. Montévil and Mossio characterize biological organization as a regime in which constraints are mutually dependent and help maintain each other while operating in thermodynamically open conditions. Mossio and Bich then argue that this kind of organization can be understood as an intrinsically teleological causal regime: biological organization contributes to establishing and maintaining its own conditions of existence. This is the background that makes phrases such as constraint closure and self-determination more than metaphor.

This also explains why teleodynamics is meant to naturalize goal-directedness, not mystify it. The long-running debate over teleology in biology has revolved around whether talk of function and purpose is vitalistic, backwards-causal, or mentalistic. The Stanford Encyclopedia’s overview describes the modern naturalistic strategy as one that preserves teleological language while avoiding those pitfalls. Teleodynamic and closure-based accounts sit squarely in that lineage: they treat end-directed behavior as arising from objective organizational relations, not from a ghostly “purpose substance.”

An adjacent scientific framing comes from active inference. Friston’s “Life as we know it” argues that systems separated by a Markov blanket will appear to preserve their functional and structural integrity through active Bayesian inference, leading to homeostasis and a simple form of autopoiesis. Active-inference authors later proposed explainable AI architectures around explicit hierarchical generative models that are interpretable and auditable. This matters because teleodynamic AI’s most plausible technical relatives are not mystical-purpose theories, but viability-centered frameworks that already model self-maintaining organization in naturalistic terms.

One useful caution follows from this biology-to-AI transfer: in biology, teleodynamics is tied to self-maintenance, self-correction, and often self-reproduction. In AI, the current proposals mostly operationalize the first two and only metaphorically borrow the third. The March 2026 arXiv paper is about maintaining viable learning organization; the PhilArchive manuscript is about operator drift, semantic orientation, and regime transitions. Neither presents anything like literal organism-style self-reproduction.

What the recent AI literature actually proposes

The more concrete of the two AI proposals is “Teleodynamic Learning: A New Paradigm for Interpretable AI.” Its core claim is that a system counts as teleodynamic only if it satisfies five commitments: two-timescale dynamics between fast parametric learning and slower structural modification; an endogenous resource variable that gates what changes are viable; a local teleodynamic objective balancing predictive loss, structural complexity change, and energy cost; emergent structural halt through a no-op option rather than externally imposed early stopping; and phase structure that distinguishes under-structuring, teleodynamic growth, and over-structuring. The paper explicitly says that learning should be treated as a trajectory through structure-parameter-resource space, not as the search for a timeless optimum.

That paper’s implementation, DE11, makes the proposal much less vague than the term “teleodynamic” might suggest. The system state includes structure, parameters, energy, and history; observations trigger both parametric updates and candidate structural actions; predictive success replenishes energy while structural moves consume it; and structural freeze occurs when no structural action can justify its complexity and energy costs relative to a no-op. The paper also stresses that this is not standard minimum-description-length in disguise: the complexity is supposed to emerge from the dynamics, the action choice is local and greedy rather than globally optimizing, and the resource variable has no straightforward MDL analog.

Empirically, DE11 is promising but modest. The authors report competitive results on small tabular benchmarks: 93.3% on IRIS, 92.6% on WINE, and 94.7% on Breast Cancer, with interpretable logical rules rather than post-hoc explanations. They also report that a no-structure version performs much worse, including a 27-point gap on IRIS, which they take as evidence that structural learning matters. At the same time, the paper is explicit that this is not a state-of-the-art claim, that it uses small interpretable datasets on purpose, and that the system struggles on a higher-dimensional DIGITS task. It also acknowledges scalability issues and reliance on a diagonal Fisher approximation.

Just as important, one of the paper’s central “emergent” properties is still only partially emergent. The authors prove that structural freeze is guaranteed under a schedule-based design with caps on structural moves and time horizon, then explicitly note that proving self-termination without those caps would require stronger assumptions and remains open. That is a crucial research-status signal: the teleodynamic idea is being formalized, but some of its most biologically evocative features are still enforced partly by engineering scaffolding.

The second 2026 paper, “Toward Teleodynamic Architectures in Artificial Intelligence,” is more speculative and more philosophical in style. Rudolph argues that LLM-style coherence through sequence prediction is not yet the same thing as goal-directed semantic organization. His proposal is to move from state-centered modeling to transformation-centered modeling, where semantics is described at the level of operators, attractor-induced drift, regime transitions, and field curvature. The manuscript introduces complex and quaternionic representational schemes to separate actualized state, possibility, directed orientation, and normative curvature; then it models teleological bias as asymmetry in the evolution of transformation weights, not simply as a target state.

That manuscript is conceptually interesting because it makes a sharp distinction between three kinds of systems: associative systems that continue probable sequences, aligned systems that obey external objectives, and oriented semantic systems that contain internal operator-level attractors with multi-scale regulation and normative curvature. It also proposes a minimal teleodynamic architecture built from phase-differentiated representation and hierarchical transformation dynamics, with tools such as operator-weight tracking, meta-attention across dialogue cycles, memory modules, and adaptive curvature fields. But it is still best read as an architectural manifesto. Its conclusion explicitly says that whether these elements can be realized is still an empirical and engineering question, and its reference list is heavily built around the author’s own recent conceptual work.

How teleodynamic AI compares with established AI

Mapped against familiar paradigms, teleodynamic AI is less a direct replacement for one existing method than a demand that several normally separate design choices—representation, structure, training dynamics, and resource budgeting—be pulled into one coupled process. The comparison below compresses the most defensible contrast.

The closest already-established neighbors are worth naming explicitly. Active inference shares the focus on viability, explicit internal models, and explainability, but it still centers a fixed variational objective. Open-endedness research shares the ambition of ever self-improving systems and novelty generation, but it does not by itself require internal closure of constraints. Open-world learning focuses on adapting to unexpected structural change and even becoming antifragile, which looks like a natural testbed for teleodynamic ideas. Developmental robotics and symbol emergence research likewise care about bottom-up concept formation and the integration of top-down and bottom-up processes, which closely resembles the “representational growth” side of teleodynamic AI. NAS and progressive neural networks already adapt structure over time, but typically without the teleodynamic insistence that structure, parametric adaptation, and internal resource variables form one closed organization.

This is also where teleodynamic AI’s interpretability claim becomes credible. Cynthia Rudin’s argument for using inherently interpretable models rather than explaining opaque ones after the fact has become an important point of reference in interpretable ML. Teleodynamic Learning’s claim is that explicit, dynamically formed structures can serve this same ideal: the model’s learned rules and constraints are the explanation, rather than an auxiliary explainer attached after training. That promise is not yet fully proven, but it is conceptually aligned with a strong line of existing interpretability research.

What a practical teleodynamic architecture would need

Representational growth would have to be a first-class mechanism, not a side effect. The symbol-emergence and developmental-robotics literature has long argued that intelligent systems need ways to build grounded internal concepts rather than treating symbol systems as static givens. Teleodynamic Learning takes a concrete step in that direction by letting the hypothesis set itself change through discrete structural actions, while Rudolph’s manuscript pushes further and asks for architectures that model transformations and orientations explicitly, not only states.

Parameter adaptation would still matter, but it would no longer be the whole story. In teleodynamic learning, fast parametric adaptation persists even after structure freezes, which is why the authors insist on two timescales. This is a good sign conceptually: teleodynamic AI is not anti-optimization; it is anti-reduction of intelligence to only optimization within a fixed hypothesis class. A realistic teleodynamic architecture would therefore combine ordinary weight updates with a slower process that can add, split, merge, or retire internal structures.

Resource accounting is where the idea becomes distinctively teleodynamic rather than simply “adaptive.” The DE11 formulation uses an internal energy variable updated by predictive success, decay, and action costs, and that variable directly gates which future actions are viable. The closure-of-constraints literature provides the philosophical analogue: constraints are not just fixed limits from outside the system, but mutually maintained conditions of operation. In AI terms, that suggests systems that track not only error, but also memory load, uncertainty, complexity burden, latency, and perhaps even hardware energy as internal state variables affecting future learning moves. That last step is an inference beyond the current papers, but it follows naturally from their design logic.

Constraint closure means the system’s learning actions should change its own future learning affordances. That is already partially present in DE11, where structural actions consume energy and complexity budget while predictive success replenishes the system. It is also present in Rudolph’s more semantic version, where operator drift reshapes the future transformation rules themselves. A genuinely teleodynamic architecture would make this loop explicit enough that one could inspect how current structure narrows or expands the next set of viable structural moves.

Emergent stabilization is another decisive requirement. Teleodynamic Learning formalizes this through a local objective plus a no-op action and insists that structural exploration should stop when more structure is not worth its cost. Rudolph expresses a similar design pressure with the claim that perfect rigidity leads to freezing while unbounded openness leads to incoherence; sustainable semantic systems need bounded openness. Across both papers, the shared engineering idea is that good learning systems should neither keep growing forever nor collapse into static closure too early.

Interpretability by construction is the strongest near-term promise. Teleodynamic Learning offers explicit logical rules; active-inference xAI work argues for explicit hierarchical generative models that are auditable by humans; and the interpretable-ML literature keeps pressing the field toward models whose internal structures are themselves intelligible. So the most compelling teleodynamic architecture is not one that merely learns forever, but one that can say, in inspectable internal terms, what structures it created, what resource pressures shaped them, and why additional growth stopped.

A reasonable engineering inference is that the best near-term testbeds for teleodynamic AI are unlikely to be static leaderboard tasks alone. The more natural arenas are open-world and embodied settings where novelty, distribution shift, representation growth, and resource trade-offs actually matter: open-world learning, developmental robotics, symbol emergence, tool-using agents, and continual adaptation problems. Nature Machine Intelligence’s open-world learning perspective is especially relevant here, because it argues that future systems must detect, characterize, and adapt to structurally unexpected environmental changes, and that evaluation itself becomes conceptually hard when the unexpected matters.

Promise, risks, and research verdict

The credible promise of teleodynamic AI is not mystical agency; it is adaptive interpretability under constraint. If a system can make its own representational distinctions, track the costs of doing so, and settle into a stable organization that remains inspectable, then it could satisfy a strong version of “interpretability by design” that today’s post-hoc XAI often does not. Teleodynamic Learning’s rule-based results, active inference’s explicit generative-model story, and the broader inherently interpretable ML agenda all point in this direction.

The main risk is conceptual overreach. Teleodynamics is attractive precisely because it talks about purpose, self-maintenance, and organization in a way that feels deeper than “minimize loss.” But that also makes it easy to hide unresolved engineering questions inside impressive language. The evidence so far justifies that caution. Teleodynamic Learning is still small-scale, explicitly non-state-of-the-art, partly schedule-scaffolded, and limited in scalability. Rudolph’s manuscript is mathematically ambitious but still architectural and programmatic, ending with the acknowledgement that implementation remains an empirical and engineering problem. Open-world-learning researchers also warn that when systems must handle the unexpected, evaluation itself becomes unusually difficult.

So the best research verdict is this: teleodynamic AI is a serious early hypothesis about how to fuse structure learning, continual adaptation, resource awareness, and interpretability into one dynamical view of intelligence. It is already more than a slogan, because it now has explicit definitions, mathematical commitments, and at least one working implementation. But it is not yet a standard, a benchmarked subfield, or a demonstrated path beyond current LLM/RL/NAS systems. For now, it is best treated as a promising synthesis agenda—one that will only become compelling if it can beat strong baselines on open-world or embodied tasks while making its internally learned organization genuinely auditable.

A compact formulation that is faithful to the current evidence would be: teleodynamic AI is a program for building learning systems in which internal structures, parameter dynamics, and resource variables mutually constrain one another so that viable organization emerges, stabilizes, and remains inspectable, rather than being fully fixed in advance by a static loss or reward function.