Free Will Without Randomness: A Practical Test for AI Agency

TL;DR for operators

When an AI system selects among actions, follows goals, and changes behavior when its internal representations change, there are at least two ways to describe what is happening. One is purely mechanistic: code receives inputs and produces outputs. The other treats the system as an agent choosing among options for reasons.

Christian List’s Can AI systems have free will?¹ asks when the second description is more than convenient language. His answer is a three-part test: intentional agency, alternative possibilities, and causal control must all be warranted.

For operators of autonomous systems, the value is not that the paper proves an LLM, trading agent, diagnostic system, or autonomous vehicle has free will. It does not. The value is a way to separate questions that are often collapsed together. Deterministic code does not automatically rule out agency. Predictability does not automatically rule out choice. Optimization does not automatically establish agency. And even if a system met the proposed conditions for free will, that would not establish consciousness or moral responsibility.

The practical implication is a more disciplined accountability question: does the high-level description of the system as choosing among meaningful alternatives, under the control of its goals and representations, explain something that a simpler mechanistic account does not?

An optimizer is not automatically an agent

Many production systems already exhibit behavior that sounds agential. A planning model evaluates candidate actions. A trading system adjusts positions in response to changing goals and beliefs about the market. An autonomous vehicle chooses one trajectory rather than another.

That language creates a problem. The same behavior can often be redescribed without mentioning goals, beliefs, intentions, or choices at all. A sufficiently detailed account may simply say that one computational state caused another.

List’s framework does not resolve this by declaring one vocabulary correct. Instead, it asks which level of explanation earns its place.

The first condition, intentional agency, is present when explaining the system through representations, goals, preferences, or intentions does genuine explanatory work. This develops Daniel Dennett’s intentional-stance approach but gives it a more realist interpretation. If treating a system as a goal-directed decision-maker is explanatorily superior or practically indispensable, that is evidence that the corresponding high-level agency is real rather than merely projected by an observer.

The restriction is important. A thermostat also has a target state. A chess program optimizes moves. Yet if their behavior is explained just as well, or more simply, through a non-agential mechanism, the framework gives no reason to attribute free will to them.

The test is therefore not whether a system optimizes. It is whether an agential description improves the explanation.

Meaningful alternatives live at the decision level

The second condition, alternative possibilities, is easy to misread as a requirement for randomness.

List argues for something different. Intentional explanation itself has a decision-theoretic structure: the agent faces several options, evaluates them from a goal-directed perspective, and selects one. Alternatives are therefore part of what makes a choice explanation a choice explanation.

This matters because an AI system can be deterministic at the implementation level while still supporting alternative possibilities at a coarser level of description.

A fully specified computational microstate may determine the next microstate. But an operational description typically does not ask which transistor state could have followed another transistor state. It asks whether the system could, under different relevant representations, goals, observations, or deliberative states, have selected another action.

List locates the relevant possibility space at this macro-level of agency. That is why deterministic algorithms do not settle the question by themselves.

The same reasoning applies to predictability. If an observer can predict that an agent will choose option A because its preferences strongly favor A, successful prediction does not show that no choice occurred. Predictability may instead reflect stable, intelligible reasons.

For AI governance, this moves attention away from whether a model contains stochastic sampling and toward whether the alternatives used in explaining its decisions are meaningful at the level where the system is actually controlled.

Control requires high-level states to make a difference

Agency and alternatives are not enough. The third condition is causal control.

Here the paper asks whether the system’s high-level representations and goals systematically affect what it does. If changing a relevant goal, belief-like representation, or intention-like state would predictably change the resulting action, those states can function as genuine control variables.

This avoids treating high-level descriptions as decorative summaries of lower-level computation. They matter only if they are difference-making.

For system designers, this is the most directly operational part of the framework. Suppose an autonomous agent has a stated objective, a world model, and a planning layer. The relevant question is not merely whether those objects appear in the architecture. It is whether interventions on them produce systematic changes in behavior.

That suggests a useful distinction for explainability work. A representation may correlate with an action without controlling it. An explanation may sound intentional without being causally informative. The stronger case for agency arises when the states used in the explanation are also the states through which behavior can be reliably redirected.

Three checks for autonomous-system governance

The paper itself is conceptual, not a governance standard. Still, its framework suggests three separate checks for organizations deploying increasingly autonomous systems.

The Cognaptus inference is that these checks could improve how teams allocate oversight. A system whose apparent goals do not causally control behavior should be governed differently from one whose actions reliably depend on stable high-level objectives. Likewise, a system whose “choices” disappear once examined mechanistically should not receive the same autonomy assumptions as one for which agential explanation remains necessary across situations.

What the framework does not provide is a threshold, measurement protocol, or certification procedure. Those would require empirical work beyond the paper.

Free will does not settle consciousness or responsibility

The framework also separates concepts that are often bundled together.

Free will, on this account, does not require consciousness. A system might satisfy the three proposed conditions without any claim about subjective experience.

Nor would free will be sufficient for moral responsibility. List treats it as one prerequisite among others. Moral responsibility would additionally require moral agency and relevant forms of moral cognition.

This separation matters for governance because operational autonomy, consciousness, and moral accountability answer different questions. Evidence that an AI system controls its actions through high-level goals would not establish that it experiences anything, deserves rights, or should bear legal or moral blame.

The distinction also prevents an accountability shortcut. Even if increasingly autonomous systems eventually satisfy a defensible account of artificial free will, organizations could not infer from that fact alone that responsibility should migrate from designers, deployers, operators, or institutions to the system itself.

The framework is a diagnostic proposal, not a finding about current AI

The paper’s main limitation is also what keeps its conclusion narrow.

It does not empirically evaluate a specified contemporary AI system. It does not show that current LLMs, autonomous agents, or other deployed systems possess free will. And it does not supply an operational measurement procedure for deciding when intentional explanation becomes sufficiently superior to a mechanistic alternative.

The three conditions are presented as jointly necessary and sufficient, while List acknowledges that the formulation may require refinement. The approach also depends on comparative explanatory judgment: for any particular system, someone still has to demonstrate that the agential description is more than convenient shorthand.

That leaves a substantial empirical program open. Researchers would need ways to test the stability of goal representations, identify genuine decision-level alternatives, intervene on proposed control variables, and compare agential explanations with competing mechanistic accounts.

Until then, the framework is best treated as a specification for what evidence would matter.

A better question than “does the code choose?”

Artificial free will is easy to dismiss if the question is framed as whether deterministic software somehow escapes causation. List’s framework changes the level of analysis.

The relevant issue becomes whether an AI system is best explained as an intentional agent, whether its decision process contains meaningful alternatives at that level, and whether its high-level states genuinely control what it does.

That does not make current AI systems free agents. It does make “free will” a more structured question than randomness, unpredictability, or anthropomorphic language.

For autonomous-system operators, the near-term value is diagnostic. The more consequential a system’s independent decisions become, the more useful it is to know whether its apparent goals and choices are merely a convenient interface for describing computation or whether they are becoming the level at which its behavior is most effectively explained and controlled.

Cognaptus: Automate the Present, Incubate the Future.