Agency Is Not a Voice: It Is a Feedback Loop

What if agency does not begin when an individual makes a clear decision, but when a system learns that its actions can change what happens next?

A three month old infant does not need language, intention, or a theory of mind to begin discovering this. A kick moves a mobile. The mobile moves again. After enough repetitions, the infant's movements cease to be mere biological activity and become a rudimentary experiment: When I move this way, does the world answer?

Something strikingly similar is happening in scientific work as artificial intelligence enters peer review. An AI system that reads a paper, identifies weaknesses, locates relevant literature, and suggests revisions is not simply automating proofreading. It is inserting a new participant into the feedback loop through which researchers decide what their work means and what to do next.

These developments appear unrelated: one concerns the emergence of agency in infancy, the other the use of artificial intelligence to improve research papers. But together they illuminate a deeper question:

How does a system become capable of purposeful action when its intentions are not yet fully formed?

The answer is not through isolated intelligence. It is through structured feedback, meaningful resistance, and the ability to alternate between movement and stillness. This offers a useful way to think about both human learning and AI assisted scholarship. The goal is not to eliminate uncertainty or replace judgment. The goal is to build environments in which better judgment can emerge.

Agency begins with a world that answers back

We often imagine agency as an internal possession. A person has a goal, forms an intention, and then acts. This picture is intuitive, but it is incomplete. Before an infant can represent a goal in words, the infant is already participating in a dynamic relationship with the surrounding world.

Consider a baby lying beneath a mobile. At first, the baby's arm movements may be random. The mobile turns because of a draft, an accidental contact, or a movement in the crib. Yet over time, the interaction can become organized. The infant's movement and the mobile's movement form a coupled system. The infant's body changes the environment, and the environment supplies information that changes the infant's next movement.

The important unit is not the infant alone. It is the infant plus the mobile plus the timing between them.

This is why movement and stillness both matter. Movement tests the environment. Stillness allows the system to register what happened. Without movement, there is no experiment. Without stillness, there is no interpretation. Agency emerges in the rhythm between perturbation and observation.

A simple mental model is the agency loop:

1. The system acts.
2. The environment responds.
3. The system detects a pattern.
4. The system changes its next action.

This loop can be primitive or sophisticated. A baby may discover that a certain kick makes a mobile sway. A scientist may discover that a claim depends on an assumption never tested. In both cases, agency is not just the capacity to produce an output. It is the capacity to become more selective because of feedback.

That distinction matters in an age of instant answers. A fast response is not necessarily useful feedback. Feedback becomes useful when it changes the quality of the next action.

The reviewer as a designed source of resistance

Scientific writing often fails long before formal peer review. A definition is vague. A comparison is unfair. A result is overinterpreted. A relevant paper is missing. The author may not see these weaknesses because writing creates a powerful illusion: once an argument is coherent in the writer's head, it can feel coherent on the page.

An AI review system can interrupt that illusion. By reading a draft and returning critiques, questions, possible references, and venue specific suggestions, it creates a rapid response from the environment. The paper is no longer a private object carried from the author's mind to an editor's inbox. It becomes an object that can be tested repeatedly before publication.

This is valuable not because the machine possesses final authority, but because it supplies productive resistance. A draft that encounters no resistance tends to preserve its hidden assumptions. A draft that encounters too much resistance becomes impossible to revise. The right feedback creates a navigable tension: enough friction to expose weaknesses, not so much noise that the author loses the thread.

The infant and the researcher therefore share a problem. Both need an environment that responds in ways they can use.

For the infant, the mobile is useful because its motion is coupled to the infant's own behavior. For the researcher, a review is useful when it is connected to the actual claims, evidence, methods, and intended audience of the paper. Generic praise and generic criticism are like a mobile that moves randomly. They create activity without teaching the learner which actions matter.

This suggests a criterion for judging AI feedback that is more important than eloquence:

Good feedback does not merely describe the current state of a system. It makes the next experiment more informative.

If a reviewer says, "This section is unclear," the author has received a label but not a usable perturbation. If it says, "Your definition of agency changes between the introduction and the analysis, so compare the two sentences and state whether agency means control, intention, or sensitivity to consequences," the feedback creates a specific next action. The researcher can revise, test, and inspect the result.

The difference is the difference between commentary and coordination.

Why stillness is part of intelligence

The promise of AI assisted review is speed. A researcher can receive comments in minutes rather than waiting weeks or months. That changes the economics of iteration, especially in fields where literature evolves quickly and deadlines are unforgiving. But speed introduces a danger: it can turn revision into an endless sequence of reactions.

The author uploads a paper, receives suggestions, accepts some, rejects others, submits a new version, and repeats the cycle. The document improves locally while the argument loses its center. Sentence by sentence, the paper becomes more polished and less owned.

This is where stillness becomes essential. Stillness is not passivity. It is the interval in which a system determines which signals deserve action. The infant's pause after a movement is not wasted time. It is part of learning whether the world responded contingently. In research, the pause is the moment when an author asks: What is the central claim of this paper, and which criticism actually bears on it?

A useful distinction is between reaction speed and learning speed. Reaction speed measures how quickly a system generates another response. Learning speed measures how quickly the system improves its model of the problem. These are not the same.

A researcher who immediately incorporates every suggestion may react quickly while learning slowly. A researcher who studies a review, groups its comments by underlying issue, runs a new analysis, and then rewrites the argument may appear slower while learning much faster.

The practical lesson is to introduce deliberate pauses into AI assisted revision. After receiving a review, do not begin by editing. First sort the feedback into four categories:

1. Validity: Does the criticism identify a genuine error or unsupported claim?
2. Importance: If true, does it affect the paper's main contribution?
3. Actionability: What specific experiment, explanation, or structural change would address it?
4. Ownership: Would accepting the suggestion clarify the paper, or merely make it sound more conventional?

This process converts a flood of comments into a map of decisions. It also protects the author's agency. The machine can reveal possibilities, but the researcher must decide which possibilities belong to the work.

The danger of confusing coordination with authority

There is a tempting but mistaken way to interpret AI review: the machine finds the flaws, the human fixes them, and the paper becomes objectively better. That model treats the reviewer as an external judge and the author as a repair technician.

Yet feedback is never neutral. A review system is shaped by its training data, its prompts, its retrieval sources, its assumptions about rigor, and its statistical preference for familiar forms. It may identify a missing citation while overlooking a foundational conceptual error. It may recommend a prestigious venue because the paper resembles past work from that venue. It may favor conventional framing over an unconventional but valuable idea.

In other words, the system can coordinate with a researcher without understanding the researcher's purpose. It can improve local alignment while weakening global originality.

The infant analogy helps clarify the issue. A mobile can teach an infant that movement has consequences, but it cannot determine what the infant ought to value. Its response supports the emergence of agency; it does not supply a complete purpose. Likewise, an AI reviewer can help a researcher see how a paper behaves under scrutiny, but it cannot decide what question is worth pursuing.

This is the difference between instrumental feedback and normative judgment. Instrumental feedback asks: Does this method support the claim? Is the comparison fair? Is the explanation clear? Normative judgment asks: Is this the claim we should make? Is this problem important? What kind of knowledge deserves to be produced?

AI is increasingly useful for the first class of questions. The second remains deeply human, not because humans are magically free of bias, but because research is an activity of choosing ends as well as optimizing means.

The best workflow therefore assigns the system a role with carefully defined boundaries. Let it act as a skeptical reader, literature scout, consistency checker, methodological questioner, and simulation of possible reviewers. Do not let it become the silent author of the paper's intellectual priorities.

A new model of revision: perturb, pause, reorient

The intersection of infant coordination and AI assisted review suggests a broader model for learning and creation. Call it perturb, pause, reorient.

Perturb means introducing a meaningful challenge. Ask an AI reviewer to search for alternative explanations, identify untested assumptions, reconstruct the strongest opposing argument, or compare the paper's claims with current literature. The point is not to collect more comments. It is to expose the paper to conditions under which its weak points become visible.

Pause means suspending automatic response. Read the critique without editing. Mark which comments alter the paper's core reasoning and which concern presentation. Look for clusters. Five comments about definitions may represent one conceptual problem. Three suggestions for additional citations may reveal that the contribution has not been situated clearly.

Reorient means changing the next action based on the pattern that emerged. Sometimes the right response is a new experiment. Sometimes it is a narrower claim. Sometimes it is a paragraph explaining a limitation. Sometimes it is rejecting the feedback because it pushes the work toward a question the researcher does not intend to answer.

This model also explains why the earliest feedback is often the most valuable. Before a paper becomes highly polished, its conceptual structure is still flexible. A critique at that stage can change the trajectory of the work. Late feedback tends to produce cosmetic repairs because the cost of changing the underlying design has become too high.

An author can make this process concrete by running three distinct review sessions:

• The coherence review: What is the paper claiming, and does every major section serve that claim?
• The adversarial review: What would a careful skeptic challenge, and what evidence would change the skeptic's mind?
• The consequence review: If the findings are correct, what should readers believe, do, or investigate next?

These sessions should not be collapsed into one request for a general review. Different questions create different feedback dynamics. A system asked to comment on everything often produces a long inventory. A system asked to test one dimension can produce a more informative disturbance.

The same principle applies beyond research. Students can use feedback to test their understanding, designers can test prototypes with real users, and leaders can use disagreement to detect blind spots. In each case, agency grows when feedback is specific enough to guide a next move and open enough to preserve choice.

Key Takeaways

• Treat feedback as an experiment, not a verdict. Ask what the critique enables you to test or clarify next.
• Build pauses into fast workflows. Do not revise immediately after receiving AI generated comments. First identify patterns, priorities, and conflicts.
• Separate craft from purpose. Use AI to improve clarity, consistency, evidence, and methodological rigor. Keep the choice of question, contribution, and values under human control.
• Prefer targeted reviews to general reactions. Request separate analyses of coherence, alternative explanations, evidence, and implications.
• Measure learning by changed decisions. A review is successful when it alters the quality of your reasoning, not merely when it produces more polished prose.

The deepest promise of AI in research is therefore not that it can imitate a peer reviewer. It is that it can make the environment of thought more responsive. A researcher can encounter informed resistance earlier, test more possibilities, and discover weaknesses while they are still affordable to fix.

But responsiveness alone is not agency. Agency requires a system that can distinguish signal from noise, connect action to consequence, and pause long enough to decide what matters. The infant beneath the mobile is not learning simply because the mobile moves. The infant is learning because movement, consequence, and attention begin to form a loop.

The same is true of a paper under review. Its quality does not emerge from criticism alone. It emerges when criticism becomes an occasion for deliberate reorientation.

Intelligence is not the ability to produce endless movement. It is the ability to learn which movements change the world, which responses matter, and when to remain still.

That may be the most useful way to think about AI assisted scholarship. The machine does not need to become the author, the judge, or the source of purpose. It needs to become a well designed surface against which thought can push and from which thought can learn. The future of better research may depend less on getting answers faster than on creating better conditions for discovering which questions, actions, and silences are truly our own.