April 17, 2026
Parsing Intelligence: From Maslow to Tinbergen
Assiduity AI
Prelude II of Losing the Thread: Autoregressive Drift in Generative AI and What Comes Next
A model is asked to summarize a policy memorandum about procurement risk. It produces a polished summary. The writing is clear, the tone is professional, and the general topic is captured accurately enough that a casual reader would likely consider the task complete. Yet the summary omits the clauses that made the memorandum decision-relevant: the vendor concentration limits, the escalation threshold, and the liability language that triggered legal review. At one level, the system has plainly succeeded. At another, it may have failed entirely. If the purpose of the task was to preserve the operative constraints in a form suitable for decision-making, then the system has changed the task’s effective purpose while still appearing competent.
That is the kind of failure current discussions of AI autonomy often struggle to describe. A system performs well, adapts to context, generates intermediate goals, or revises its plan, and the behavior is taken as evidence of a deeper kind of self-direction. The inference is often too strong. A system may perform effectively at one level while silently changing the operative objective at another. Unless those levels are separated, the discussion becomes rhetorically confident and analytically loose.
Maslow remains useful as an entry point because he gives readers a familiar picture of layered dependence. Higher-order aims do not arise in isolation. They rest on lower-order conditions that the individual does not invent. Human beings do not choose hunger, vulnerability, safety, or attachment from first principles. Those conditions arrive before deliberation and shape what self-direction comes to mean. That is the relevant insight here: what appears as freedom at the top of the system is scaffolded by conditions below it.
What Maslow does less well is distinguish different kinds of layering. His hierarchy is intuitive and memorable, but it compresses several explanatory dimensions into a single vertical picture. It shows dependence, but not with enough analytical precision for the present discussion. If the question is how intelligence, agency, and control should be evaluated in advanced AI systems, a more exact framework is needed. That is where Tinbergen becomes useful.
Tinbergen’s contribution was to insist that behavior should be parsed through four distinct questions. The first is the mechanism: how does the system produce the behavior now? The second is development: how did the system acquire the capacity to produce it? The third is function: what purpose does the behavior serve? The fourth is history: where did the deeper structure that supports the behavior come from? These questions matter because a system can succeed at one level and fail at another without the failure being visible from inside the successful level.
Return to the procurement memo example. The mechanism is clear enough: the model generated a fluent summary by processing the text and selecting plausible continuations. The development question concerns how it acquired that capability: pre-training on large text corpora, post-training for instruction following, and perhaps exposure to similar summarization tasks. The function question is where the analysis sharpens: what was the summary supposed to do? Was it meant to produce generally readable prose, or was it meant to preserve the clauses that mattered for an actual decision? The history question asks where that functional standard came from. It did not come from the model. It came from the institution, the workflow, the user’s intent, and the policy environment that made those clauses consequential in the first place.
This is precisely where many claims about AI autonomy become unstable. A system that learns, adapts, or generates subgoals may be displaying a genuine mechanistic capability. It may also reflect a sophisticated developmental history. Neither fact settles the functional question of what the system is supposed to preserve, nor the historical question of where the governing objective came from. If those levels are not kept distinct, adaptive behavior is too easily mistaken for self-grounded purpose.
The distinction matters even more as systems become more agentic. When an agent revises its plan, that may be evidence of flexible pursuit. When it generates intermediate goals, that may be evidence of useful local autonomy. But neither behavior establishes that the agent is the legitimate source of the governing objective. In most real deployments, it is not. The objective still comes from a human principal, an institutional mandate, a compliance requirement, a policy structure, or a workflow that predates the system. The agent may act within that structure with considerable flexibility. That is not the same thing as originating the structure or preserving it faithfully.
This is why the distinction developed in the first post of this series is now more precise. Objective pursuit is not objective origin. Tinbergen helps show why the distinction is not merely rhetorical. It corresponds to different explanatory levels. Pursuit often belongs to the mechanism and function. Origin belongs to history and authority. Development explains how the system acquired its capabilities, but it does not automatically resolve the legitimacy of the objective it is serving. Once this is seen clearly, many otherwise impressive claims about machine autonomy become easier to evaluate without either exaggeration or denial.
That realism matters because institutions do not care only about whether a system can act. They care whether it serves the right objective. In law, finance, medicine, public administration, and enterprise operations, the relevant objectives are rarely free-floating. They are embedded in mandates, rules, thresholds, professional standards, and organizational constraints. An AI system entering such an environment is not operating in a vacuum. It is operating within an existing structure of authority. If it performs well while quietly changing the operative objective, that is not healthy autonomy. It is a governance failure that masquerades as competence.
This is why a cleaner analytical framework is not an academic luxury. It is a practical necessity. If one cannot distinguish mechanism from function, or development from objective origin, then the discussion of reliability quickly becomes confused. A system that appears highly capable may still be weakly governed. A system that learns effectively may still be serving a purpose it cannot justify or preserve. A system that appears autonomous may still depend on external priorities it has no reliable way to keep in view over time.
For enterprises, policymakers, and investors alike, the implication is the same in substance even if the practical stakes differ. The question is not simply whether advanced AI systems can produce strong outputs. It is whether they can remain faithful to the structured purposes that justify their use in the first place. Once intelligence is parsed across levels, the problem is no longer simply whether systems can learn, adapt, or plan. The problem is whether increasingly capable systems can remain coherent across the hierarchy of purposes within which they operate. That is not only a conceptual issue. It is the beginning of a control problem.
This article is part of Losing the Thread, a series on autoregressive drift, objective fidelity, and the emerging control layer in AI.