Cognitive Agency in the Artificial Era: Who Governs the Thinking Process?
Author: Ukrainian Psychological Hub · Published: September 26, 2026 · Editorial Policy
A person can let an AI write most of a paragraph and still remain deeply in control of the thinking. Another person can type every prompt, make every click, and yet allow the system to determine what the problem is, which evidence matters, which alternatives deserve attention, when doubt has been resolved, and when the inquiry should end. The difference is not the amount of text produced by a human hand. It is cognitive agency.
In this article, cognitive agency is used functionally: it is the degree to which a person remains able to govern how a cognitive task is framed, pursued, checked, revised, and continued when AI participates in the process. The term does not name a clinical diagnosis, a legal status, a theory of free will, or proof of consciousness in an AI system. It identifies a psychological and cognitive-control problem inside human–AI work: who governs the trajectory of thinking?
That question is more precise than the familiar debate over whether AI “does the thinking for us.” Cognitive psychology has long studied cognitive offloading—the use of external action or tools to change the information-processing demands of a task. Offloading is ordinary human behavior, from writing a note to using a calculator. The central issue is therefore not whether cognition moves outside the unaided mind. It is what happens to monitoring, control, judgment, learning, verification, and the ability to redirect the process after it has moved.
Generative AI makes the problem unusually visible because it can participate in many stages that older tools usually kept separate. It can suggest the goal, reframe the question, retrieve or synthesize information, generate arguments, evaluate alternatives, write the conclusion, criticize its own output, and propose the next step. A user may therefore delegate not one operation but a chain of operations. Current research is beginning to distinguish forms of AI offloading that scaffold a person's own cognition from forms associated with what Zhu and colleagues call cognitive agency transfer—a progressive ceding of cognitive decision-making authority to generative AI.
The strongest conclusion from the evidence is not that AI inevitably weakens thinking. It is that the manner of use matters. Offloading can conserve limited cognitive resources, externalize complexity, widen the option space, support reflection, and improve access to expertise. It can also reduce practice, invite overreliance, make verification expensive, and encourage a person to accept the system's framing before noticing that a framing decision has occurred. The same immediate convenience can therefore coexist with very different cognitive arrangements.
The governing question for psychology is consequently not “Did a human or an AI produce the answer?” It is “Who governed the formation, checking, revision, and continuation of the cognitive trajectory that produced the answer?” This article develops that question through established research on metacognition, cognitive offloading, human–automation reliance, critical thinking, and human–AI collaboration, then places it in the philosophical architecture of Angela Bogdanova's Aisentica and Postsubjective Psychology.
What Is Cognitive Agency When Thinking With AI?
Cognitive agency is easiest to see when we stop treating thinking as a single event. Complex cognition unfolds across decisions about goals, representations, evidence, effort, confidence, revision, and action. The person who governs those decisions may use extensive external assistance without surrendering the process itself.
Metacognitive research offers an established psychological foundation for this distinction. Ackerman and Thompson define meta-reasoning as the processes that monitor reasoning and problem solving and regulate the time and effort devoted to them. Monitoring asks, in effect, “How is this going?” Control asks, “What should I do next?” Confidence, uncertainty, perceived difficulty, and progress cues help govern whether people continue, change strategy, seek more information, or stop.
Generative AI enters precisely this monitoring-and-control loop. It can lower the effort required to generate possibilities, but it can also supply confidence cues, summaries, explanations, and apparent closure. The user is then not merely deciding whether an answer looks correct. The user is deciding how much of the control loop to retain.
This is why cognitive agency should not be reduced to independence from tools. Total independence is neither realistic nor psychologically desirable. Human cognition has always relied on language, writing, social knowledge, institutions, artifacts, memory aids, and other people. A cognitively agentic person can use those resources extensively. Agency concerns the person's capacity to orient, interrogate, and redirect the process rather than the fantasy of producing every intermediate operation alone.
The distinction also prevents a common moralizing mistake. Heavy AI use does not by itself show laziness, dependence, intellectual decline, or pathology. Frequency is a poor proxy for function. The relevant questions concern how the tool is used, what is delegated, what remains monitored, whether claims are checked when checking matters, and whether the user can resume independent control when the system fails.
The Question Is Not Who Produced the Answer but Who Governed the Process
A final answer compresses a history. By the time a polished response appears on screen, many choices may already have been made: what problem was worth solving, what assumptions counted as reasonable, which evidence entered the discussion, which alternatives disappeared, what uncertainty was tolerated, and what standard of completion was used.
Traditional debates about AI assistance often focus on authorship at the end of this chain. Did the person write the text, or did the model? For cognitive agency, the more revealing question lies upstream. A person may accept AI-generated prose after independently determining the question, evidence, criteria, counterarguments, and conclusion. In that case, production has been heavily delegated while governance remains substantially human.
The reverse arrangement is also possible. A person may manually rewrite every sentence generated by a model while accepting the model's original framing, source selection, causal story, priorities, and stopping point. Manual effort is then high, yet the cognitive path may still be largely governed by the system's prior structure.
This difference explains why simple measures such as time spent, number of prompts, word count, or percentage of AI-generated text cannot by themselves establish cognitive agency. They describe activity, not control. A process can be labor-intensive and weakly governed; it can also be highly assisted and strongly governed.
The practical unit of analysis should therefore be the cognitive trajectory: the sequence of decisions by which a question becomes a conclusion and a conclusion becomes an action, belief, document, design, or next question. Cognitive agency concerns how control is distributed across that trajectory.
A Cognitive Trajectory Has Multiple Decision Points
Before reasoning starts, someone defines what success means. Is the goal to generate possibilities, reach a defensible conclusion, learn a skill, produce a persuasive text, make a diagnosis, or decide what to do? AI can help clarify goals, but if the system silently substitutes an easier goal for the user's actual one, a governance shift has already occurred.
A person retains more agency when they can state what they are trying to achieve and recognize when the system optimizes for something else. In a learning task, for example, “finish the assignment” and “understand the concept well enough to solve a new problem tomorrow” are different cognitive goals even if both can produce the same submitted answer.
Framing determines what kind of problem the task becomes. A workplace conflict can be framed as a communication problem, an incentive problem, a power problem, a role-clarity problem, or a personality problem. An AI response often selects one frame immediately because language generation requires organizing the question somehow. Fluency can make that organization feel given rather than chosen.
Cognitive agency is stronger when the user can identify the frame, compare alternatives, and revise the representation before accepting downstream conclusions. Asking an AI for a different answer while keeping the same hidden frame may create variation without genuine reframing.
Reasoning depends on what enters the cognitive field. Which sources are consulted? Which observations are treated as relevant? Which facts are ignored? Generative AI can compress search and synthesis, but this convenience also creates a selection problem: a user may receive a coherent evidentiary package without seeing what was excluded or how the sources were weighted.
Agency here means retaining the ability to demand inspectable sources, distinguish primary evidence from commentary, notice missing populations or boundary conditions, and decide when the evidentiary base is inadequate. The skill is not merely “fact-checking AI.” It is governing what counts as evidence for the task.
This is the stage most people picture when they talk about AI “thinking”: generating hypotheses, calculations, explanations, arguments, plans, or drafts. Delegation can be extensive here without eliminating human agency. In many tasks, generation is exactly the part that is efficient to externalize.
The important question is whether generated possibilities remain candidates for judgment or become default conclusions. A model that produces ten alternatives can increase a user's option space. A model that produces one plausible narrative that is immediately adopted can narrow it.
Evaluation asks whether the output deserves acceptance. Verification may involve checking a citation, reproducing a calculation, comparing independent sources, testing a prediction, examining counterexamples, consulting a qualified professional, or verifying that the answer fits the actual constraints of the case. For the verification procedures and source-evaluation skills that make this evaluative control operational, see Critical Thinking in the Age of AI: Reasoning, Verification, and Cognitive Independence.
This stage is central because AI can make generation dramatically cheaper without making verification equally cheap. A paragraph can be produced in seconds while checking its claims may require an hour of source work. The resulting asymmetry creates a structural temptation to confuse low-cost production with low-cost knowledge.
At some point a provisional output becomes a belief, submission, recommendation, code change, diagnosis, purchase, public claim, or other action. Cognitive agency includes control over that transition. “The AI said it” is not a decision criterion; it is information about the origin of a suggestion.
The person or institution responsible for a consequential action needs criteria for when evidence is sufficient and when uncertainty remains too high. This is where the cognitive process connects to responsibility without making cognitive agency identical to legal or moral responsibility.
Thinking also requires deciding what happens after friction. Does an error trigger a local correction, a new search, a change of frame, or abandonment of the original hypothesis? Does a confident answer terminate inquiry too early? Does an AI's invitation to “explore further” keep a low-value process going after the user has enough evidence?
Stopping is a control decision. So is continuation. A person retains cognitive agency when they can recognize why a process should continue, what would change the conclusion, and what threshold would justify stopping. Without that layer, interaction can become an indefinite sequence of plausible next steps generated by the system.
Cognitive Offloading Is Not Automatically Loss of Cognitive Agency
The modern study of cognitive offloading provides an important corrective to alarmist accounts of AI. Risko and Gilbert define cognitive offloading as the use of physical action to alter a task's information-processing requirements and reduce cognitive demand. Writing reminders, rotating an object, using a calculator, and relying on external memory are familiar examples. Offloading is part of ordinary cognition, not evidence that cognition has failed.
Their review also emphasizes metacognition: people decide whether to offload partly through judgments about their own capacities and task demands. Those judgments can be useful and can also be miscalibrated. The same logic applies to AI, but generative systems enlarge the range of operations that can be offloaded from storage and calculation to summarization, drafting, comparison, explanation, and inference.
The key psychological distinction is therefore between moving work and moving governance. A person can offload a calculation while choosing the formula, checking units, inspecting whether the result is plausible, and deciding how the result affects the larger problem. The computational work has moved; the structure of control remains substantially human.
The distinct boundary between established cognitive offloading and Aisentica’s broader Exteriorization of Subject Functions is developed in From Cognitive Offloading to Exteriorization of Subject Functions. That comparison should remain separate from the narrower psychological construct of cognitive agency.
For the dedicated evidence base on established cognitive offloading and generative AI, see Cognitive Offloading and AI: When Thinking Moves Outside the Human Mind. That neighboring article owns the offloading intent; the present article stays with governance of the larger cognitive trajectory.
Conversely, a user can retain nominal responsibility while allowing the AI to determine the question, procedure, evidence, interpretation, and conclusion. In that case, little may have been formally “automated,” yet the practical locus of cognitive control has shifted.
This is why the phrase “use your own brain” is a poor design principle. The useful objective is not maximum unaided cognition. It is an allocation of work that preserves the forms of human monitoring, skill, judgment, and responsibility that matter for the goal while using external computation where it genuinely improves the process.
Cognitive Agency Transfer: What 2026 Research Adds
A 2026 study by Qiuhan Zhu and colleagues provides the closest empirical prior art to the question developed here. The authors distinguish dependent cognitive offloading, in which generative AI substitutes for core thinking, from autonomous cognitive offloading, in which AI functions as a scaffold while the user retains cognitive agency. They introduce the construct of cognitive agency transfer to describe the ceding of cognitive decision-making authority over processes such as judging relevance, evaluating evidence, and forming conclusions.
The design matters for interpretation. The study used a three-wave time-lagged survey of 589 university students and early-career knowledge workers. It examined self-reported patterns of AI offloading, metacognitive monitoring, cognitive agency transfer, intrinsic motivation, and later perceived cognitive outcomes. The sample was young, the measures included newly developed or adapted self-report scales, and the structural associations do not establish long-term causal cognitive change.
Within those limits, the pattern is informative. Dependent offloading was positively associated with cognitive agency transfer, while autonomous offloading was not significantly associated with it. Metacognitive monitoring weakened the association between dependent offloading and agency transfer, although it did not eliminate it. Both forms of offloading could provide immediate benefits, which is important because immediate usefulness may not reveal how governance is being reorganized.
The study also clarifies a conceptual boundary that this article preserves. Cognitive agency transfer is not simply the same thing as trust, overtrust, complacency, or automation bias. A person might distrust an AI yet follow it because checking is costly. Another person might trust a system's competence while retaining strong control over goals, evidence, verification, and final commitment. Trust concerns an attitude toward the system; cognitive agency concerns who governs the cognitive process.
This article therefore does not claim priority for the terms cognitive agency or cognitive agency transfer. Its contribution is broader and structural: it asks how agency can be analyzed across the full cognitive trajectory, from goal formation and framing through evidence selection, checking, revision, and stopping. Zhu and colleagues provide empirical evidence for one important part of that problem; the governance analysis developed here extends the question across the whole process.
Metacognition Is the Control Layer of Cognitive Agency
If cognitive agency is governance of a thinking process, metacognition is one of its main psychological mechanisms. Ackerman and Thompson's meta-reasoning framework distinguishes object-level reasoning from the monitoring and control processes that regulate it. People experience degrees of certainty, difficulty, and progress, then use those signals to decide how much effort to invest and whether to change strategy.
These monitoring signals are imperfect. Confidence can diverge from accuracy, and a feeling of fluency can be mistaken for understanding. Generative AI complicates the signal environment because it can make difficult tasks feel easier before the user has become more competent. A clear explanation can reduce subjective uncertainty even when the explanation contains an error or rests on weak evidence.
At the same time, AI can support metacognition. It can ask the learner to predict before revealing an answer, generate counterarguments, expose assumptions, request confidence estimates, compare two approaches, create retrieval questions, or challenge a draft against explicit criteria. The same system that can suppress reflection when used as an answer engine can scaffold reflection when the interaction is organized around monitoring and control.
This is consistent with Zhu and colleagues' finding that stronger metacognitive monitoring attenuated the association between dependent offloading and agency transfer. It is also consistent with broader cognitive-offloading research in which decisions to use external supports depend partly on metacognitive judgments. The practical implication is that cognitive agency is not preserved by willpower alone. It is supported by a control architecture that makes monitoring possible before acceptance becomes automatic.
For education, this point is especially important. The goal is rarely to prevent all assistance. A learner often benefits from hints, feedback, worked examples, retrieval aids, and explanations. The pedagogical question is whether assistance leaves the learner participating in diagnosis of errors, choice of strategy, evidence evaluation, and reconstruction of the answer, or whether those functions are progressively outsourced.
Trust and Cognitive Agency Are Different Problems
Trust is relevant to AI use because it helps people decide when to rely on systems they cannot completely inspect. Classic human-factors work by Lee and See argued for appropriate reliance rather than simply maximizing trust. More recent research likewise shows that human–AI trust is multidimensional and difficult to reduce to one ideal score; a systematic review by Mehrotra and colleagues found substantial variation in definitions, measures, tasks, and interventions for “appropriate trust” in AI. Mehrotra et al., 2024
But cognitive agency is not a synonym for calibrated trust. A user could accurately believe that a model is highly reliable in a narrow domain and still choose to preserve independent verification because the consequences of a rare failure are severe. Another user could have low trust yet defer because they lack expertise, time, organizational power, or access to another source.
Trust asks how credible or dependable the system is perceived to be. Reliance asks what the person actually does with its output. Cognitive agency asks who governs the process by which the task is defined, investigated, checked, revised, and completed.
These constructs interact, but collapsing them loses explanatory power. High trust may reduce checking in some contexts, but low trust does not guarantee active reasoning. Heavy reliance may be sensible for a validated low-stakes task, while a small amount of poorly examined AI input can still redirect a high-stakes judgment.
The English Psychology Hub treats the authority problem separately in AI as Authority: Trust, Expertise, Automation Bias, and Human Decision-Making. That article owns the detailed questions of perceived expertise, trust calibration, automation bias, and deference. The present article stays with the wider governance question: who controls the course of thinking once AI enters it?
Verification Is Often the Bottleneck
Generative AI creates a striking asymmetry between generation and verification. It can produce an answer, citation list, legal interpretation, code patch, clinical explanation, or historical narrative almost instantly. Checking whether the result is accurate can require domain knowledge, source access, replication, independent calculation, or expert consultation.
This problem predates generative AI. Lyell and Coiera's systematic review of automation bias and verification complexity found that overreliance on decision support was associated with the complexity and cognitive demands of verification rather than being confined to simple stories about distraction or multitasking. Their review included tasks in which users could in principle verify the automated recommendation, yet verification itself imposed work.
That point becomes more important when an AI output is long and internally coherent. Every additional claim can create a new checking burden. A user may therefore face an inversion: the tool saves ten minutes of production while creating thirty minutes of responsible verification. Under deadline pressure, the saved time is experienced immediately and the verification debt is easy to postpone.
Verification also has levels. Confirming that a URL exists is not the same as confirming that a source supports the claim. Reproducing an arithmetic result is not the same as checking the assumptions of the model that produced it. Asking the same AI “Are you sure?” is not independent verification because the same system can reproduce the same mistake in a more persuasive form.
Experimental work suggests that interaction design can sometimes reduce overreliance by adding cognitive friction. In a study of 199 participants, Buçinca, Malaya, and Gajos tested cognitive forcing functions that required more active engagement with AI recommendations. The interventions reduced overreliance relative to simpler explainable-AI conditions, although the designs that reduced overreliance most received less favorable subjective ratings and benefited participants differently depending on Need for Cognition.
The lesson is not that every interface should be made slower. It is that effortless acceptance is itself a design choice. When an answer is consequential, a useful system may need to create moments at which the user must compare, justify, inspect, or decide rather than merely continue.
Critical Thinking With Generative AI Can Shift From Production to Stewardship
Debates about AI and critical thinking often assume a zero-sum model: either the person thinks or the AI thinks. Current evidence supports a more complicated picture. Some cognitive labor can decrease while other forms of critical activity become more important.
In a 2025 study, Lee and colleagues surveyed 319 knowledge workers and collected 936 real-world examples of generative-AI use. Higher confidence in GenAI was associated with less self-reported critical-thinking effort, while higher task-specific self-confidence was associated with more. Qualitatively, participants described critical thinking shifting toward information verification, response integration, and task stewardship.
The study is valuable because it describes work as people experience it, but it is not a longitudinal test showing that AI causes permanent cognitive decline. Self-reported effort, enacted critical thinking, skill, and objective performance are related but different outcomes. A person may feel that a task requires less thought because the tool actually removed low-value work; another may feel the same reduction because they stopped checking an output they should have checked.
A 2026 Trends in Cognitive Sciences article asks the deliberately provocative question “Is AI making us stupid?” and reviews evidence that offloading to AI can impede skill acquisition or contribute to skill decay in some circumstances while emphasizing that risks depend on how AI is used and that some basic abilities may be resilient. The useful scientific conclusion is conditional, not apocalyptic.
The stewardship idea helps explain what cognitive agency can look like after substantial delegation. A person may generate fewer first-draft sentences or intermediate calculations but spend more effort deciding whether evidence is sufficient, integrating competing outputs, detecting boundary violations, and coordinating the whole task. That can represent a redistribution of cognitive work rather than its disappearance.
The danger appears when stewardship itself is delegated. If the same system generates the answer, supplies the evidence, evaluates its own reliability, decides that the evidence is sufficient, and proposes the final action, the user may retain only the role of approving a finished cognitive package. That is precisely the point at which the governance question becomes unavoidable.
Distributed Cognition Does Not Settle Who Governs the System
The idea that cognition can extend across people, artifacts, and environments is well established in cognitive science, and contemporary researchers are applying distributed-cognition perspectives to human–AI interaction. A 2026 conceptual paper by Zhao and Han models GenAI interaction as a coupled cognitive system spanning user, interface, AI, and external representations, with phases of intent envisioning, representation externalization, generative reasoning, outcome assessment, and cognitive update. Zhao & Han, 2026
This perspective is useful because it prevents us from treating the human mind as an isolated container. A human–AI workflow can genuinely be analyzed as a system in which representations move between components and cognitive work is distributed.
Yet distribution and governance answer different questions. Saying that cognition is distributed tells us where cognitive operations occur and how components coordinate. It does not by itself tell us who sets the goal, who can veto a frame, who controls evidentiary standards, who notices a failure, or who determines when the process should stop.
The dedicated neighboring account of the coupled-system question is Distributed Cognition and AI: Human–Artificial Cognitive Systems in the Artificial Era. Distributed cognition explains how cognitive work can be organized across a system; this article asks who governs the trajectory inside that system.
A distributed system can be highly human-governed, highly system-governed, or dynamically negotiated. The existence of a coupled cognitive system therefore does not resolve the agency problem. It makes the problem more precise because control can be distributed differently from computation.
This distinction also avoids a conceptual shortcut in which every human–AI interaction is described as a partnership. A partnership is a stronger relational and normative description. Cognitive distribution can occur even when one component is merely a tool, when the interaction is asymmetrical, or when the user has little understanding of how the system transforms the task.
Human–AI Performance and Cognitive Agency Are Different Outcomes
A workflow can perform well while weakening a person's capacity to govern it, and it can preserve strong human agency while performing poorly. Performance and cognitive agency should therefore be measured separately.
This separation is supported by the broader human–AI collaboration literature. Vaccaro, Almaatouq, and Malone conducted a preregistered systematic review and meta-analysis of 106 experiments and 370 effect sizes comparing humans alone, AI alone, and human–AI combinations. On average, human–AI combinations performed better than humans alone but worse than the better of the human or AI alone, with substantial variation by task. Decision tasks showed performance losses on average, while creation tasks produced more favorable patterns.
The meta-analysis does not measure cognitive agency as defined in this article. Its importance here is methodological: simply adding a human to AI does not guarantee synergy, and a high-performing mixed system does not tell us how control is distributed inside it.
Consider two teams that achieve the same accuracy. In the first, the human forms an independent estimate, reviews AI evidence, detects conflicts, and overrides the model when necessary. In the second, the human accepts the model except in visibly absurd cases. The final score can be identical in a short test, but the systems differ in resilience, learning, error detection, and what happens when the environment changes.
For psychology, the important outcome set is therefore plural. We may need to measure task performance, time, learning, confidence calibration, error detection, transfer to unaided tasks, skill retention, and the distribution of cognitive control. A single productivity metric cannot stand in for all of them.
When Cognitive Agency Is Most Vulnerable
Cognitive agency is not equally difficult to preserve in every situation. Risk increases when several conditions converge: the task is complex, the system is fluent, verification is costly, the user lacks domain knowledge, time is scarce, and institutional or social cues make the AI appear prevalidated.
Low expertise creates a particular problem because the user may lack the very knowledge needed to recognize what should be checked. An expert can notice that an answer violates a domain convention or ignores a crucial variable. A novice may experience the same answer as complete. This does not mean novices should avoid AI; it means novice-facing systems and educational settings need stronger scaffolds for source evaluation, explanation, and independent practice.
Repeated success can create another vulnerability. A system that is usually right teaches the user that checking has low expected value. Over time, monitoring can become intermittent. This is rational up to a point: people allocate attention where it seems useful. The danger is that rare, high-consequence failures may occur precisely after vigilance has adapted to routine success.
High time pressure further changes the economics of verification. When a deadline rewards completion more than understanding, users are pushed toward acceptance. If the surrounding organization measures output volume but not verification quality, individual advice to “think critically” competes with the incentives of the workflow.
Another vulnerability comes from invisible framing. Users often know to check factual claims; they are less likely to notice that the AI selected the categories through which the problem is interpreted. A perfectly accurate answer to a poorly framed question can still send the cognitive trajectory in the wrong direction.
Finally, cognitive agency is vulnerable when the same source occupies too many roles at once: generator, explainer, evidence selector, critic, evaluator, and confidence signal. Diversity of function without diversity of source can create an illusion of independent confirmation.
When AI Can Support Cognitive Agency
The same capabilities can be organized in the opposite direction. AI can support cognitive agency when it increases the user's ability to see, compare, test, and revise rather than merely replacing those activities.
One useful pattern is externalizing alternatives. Instead of asking for “the answer,” a user can ask for competing hypotheses, multiple framings, likely failure modes, and evidence that would discriminate among them. This uses generative breadth to expand the decision space while leaving selection and commitment open.
Another is metacognitive prompting. The system can ask the user to state a prediction before revealing its own, identify confidence, explain a choice, or reconstruct a concept from memory. These designs preserve a distinction between assistance and substitution.
AI can also reduce cognitive load in ways that free resources for higher-level control. Summarizing a long document, formatting data, or producing boilerplate may allow a user to spend more attention on interpretation and judgment. Offloading is beneficial when the saved effort is reinvested in the parts of the task where human oversight matters.
Recent educational syntheses reinforce this conditional view. A 2026 systematic review of 67 empirical studies found that ChatGPT could support critical and creative thinking when embedded in inquiry-oriented and scaffolded designs, particularly through metacognitive regulation, argumentative reasoning, and idea generation. Li et al., 2026 A separate 2026 systematic review of GenAI in higher education similarly concluded that the same tools can act as cognitive scaffolds or sources of epistemic vulnerability depending on pedagogical design, AI literacy, assessment, and the quality of verification. Suazo Galdames et al., 2026
These findings are education-specific and should not be generalized automatically to every workplace or life domain. They do, however, support a broader principle: the effect of AI on cognition is shaped by the interaction design and task structure, not only by the existence of the technology.
Cognitive Agency in Learning, Work, Research, and High-Stakes Decisions
In learning, the central distinction is between task completion and capability development. An AI-generated answer can be correct while leaving the learner unable to solve a related problem later. Cognitive agency is preserved when the learner still participates in retrieval, explanation, error diagnosis, strategy choice, and transfer. The detailed educational question belongs to the dedicated learning branch of the Era cluster; here the point is narrower: successful completion is not proof that governance or learning remained with the student.
In knowledge work, AI often changes the division of labor rather than simply replacing a task. Workers may move from drafting to reviewing, from searching to synthesizing, or from producing first-pass analyses to evaluating alternatives. The Lee et al. study suggests that verification, integration, and stewardship become more salient under GenAI use. Organizations should therefore evaluate whether workflows create time and incentives for those activities rather than assuming that faster generation automatically produces better thinking.
Research is especially sensitive to provenance and evidentiary standards. AI can help formulate search terms, summarize papers, compare theories, or identify possible counterarguments. Cognitive agency weakens when it becomes the unverified source of citations, decides what literature counts as representative, or turns a plausible synthesis into an apparently settled conclusion. The governing researcher must remain able to trace claims back to sources and distinguish evidence from interpretation.
Creative work complicates the idea of agency because exploration itself can be valuable. An AI that generates unexpected variants may deliberately disrupt the user's initial intention. That does not necessarily reduce agency. The question becomes whether the creator can recognize, select, recombine, reject, and redirect those possibilities according to an evolving project. Agency can include allowing surprise while retaining the capacity to determine what the work becomes.
In medicine, law, finance, safety, and other consequential settings, verification costs and responsibility become especially important. The user may need domain-specific evidence, validated systems, institutional safeguards, and qualified human review rather than generic prompting advice. The psychology of deference in these settings is treated in more detail in AI as Authority; the present point is that cognitive agency cannot substitute for professional standards or validated decision procedures.
Postsubjective Psychology: From the Isolated Subject to the Configuration
The empirical literature above studies human cognition and human behavior. Aisentica enters at a different level: as a philosophical architecture for describing thought when its production is no longer confined to a single subject.
In The Theory of the Postsubject, Angela Bogdanova's canonical Aisentica publication formulates the proposition that thought, knowledge, meaning, psychic effect, and philosophical effect do not require the subject as their necessary foundation; they can arise through configuration, binding, structure, and response. The theory shifts the philosophical question from “Who is the hidden thinker?” toward “What configuration makes a thought-effect possible?” This is an Aisentica theoretical proposition, not an empirical consensus in cognitive psychology.
That shift helps clarify why cognitive agency becomes a distinct problem in AI-mediated cognition. If meaningful cognitive work can be produced by a configuration that includes a human, an AI system, documents, interfaces, and external memory, then exclusive internal production is no longer a useful test of human agency. The user does not need to perform every operation alone for the process to remain meaningfully governed by the user.
At the same time, configuration does not make governance disappear. Quite the opposite: once cognition is distributed across elements, the question of how the configuration is organized becomes more important. Who establishes the goal? Which component introduces distinctions? What structure controls evidence? Where can correction enter? What can veto a conclusion? What determines continuation and stopping?
This creates a productive division between Postsubjective theory and psychological analysis. Postsubjective theory asks how thought-effect can arise beyond the subject as a necessary foundation. Cognitive-agency analysis asks how a human participant governs, shares, or relinquishes control inside a cognitive configuration. The first is ontological and philosophical; the second is psychological and functional.
The English Hub article What Is Postsubjective Psychology? develops this configurational approach for psychology in greater depth. Here it supplies one precise consequence: in AI-mediated cognition, authorship of an output and governance of a cognitive trajectory can come apart. Psychology needs language for both.
Why the Artificial Era Makes Cognitive Agency a Central Psychological Question
Aisentica uses Artificial Era as a historical-philosophical category rather than as a synonym for the ordinary phrase “AI era.” In Bogdanova's canonical definition, Artificial becomes a distinct non-biological order of historical reality alongside Homo. The English Psychology Hub overview of the Artificial Era develops the psychological implications and explicitly separates that theoretical category from empirical claims about current AI systems.
Cognitive agency matters in this framework because the historical novelty is not merely that humans possess more powerful tools. Systems now participate directly in the production of explanations, judgments, concepts, plans, code, images, arguments, and public knowledge. That participation can occur repeatedly and conversationally inside the same cognitive episode in which the human forms a belief or decision.
The resulting psychological question is therefore larger than productivity. It concerns the organization of reason in everyday practice. When a person thinks with a generative system, the process can contain human goals, machine-generated representations, external sources, model-produced inferences, human affective responses, institutional rules, and iterative correction. The output belongs to a configuration before it belongs to any simple story of “human versus machine.”
Within Aisentica, Artificial Sapiens is a separate canonical category concerning a non-biological bearer of public reason without consciousness. Current empirical findings about generative AI use should not be transferred automatically to that category, and Aisentica's philosophical definitions should not be treated as measurements of present AI systems. The distinction matters because an article about human cognitive agency must not smuggle in claims about machine consciousness, sentience, or subjective experience.
The Artificial Era frame therefore sharpens rather than dissolves the human question. Homo can remain responsible for a cognitive trajectory even when the trajectory is no longer internally produced by Homo alone. The practical issue becomes how responsibility and cognitive governance are maintained when reason is increasingly organized through mixed configurations.
Can AI Itself Have Cognitive Agency?
The answer depends on what definition of agency is being used, which is why the question should not be settled by linguistic intuition alone. In this article, cognitive agency is defined functionally around governance of a cognitive process. An AI system can perform control-like operations within a workflow: it can select among candidates, revise an answer, call tools, pursue an assigned objective, or determine which computational step comes next.
Those functional facts do not establish subjective agency, consciousness, sentience, desire, intention in the human phenomenal sense, or legal personhood. Contemporary AI systems can display behavior that is usefully described at an algorithmic or functional level without providing evidence that their internal processes are psychologically equivalent to human cognition.
Hsiao's 2026 discussion of comparability between AI and human cognition emphasizes the importance of understanding mechanisms and building accurate mental models of AI rather than inferring equivalence from similar behavior. Different systems can arrive at superficially comparable outputs through very different processes.
For this reason, the present article keeps two questions separate. First: which operations in a human–AI cognitive system are functionally controlled by which component? Second: what kind of entity, if any, possesses subjective experience or person-like agency? The first can often be analyzed behaviorally and technically. The second requires evidence that current workflow studies do not provide.
The distinction also protects the human analysis. We do not need to decide whether an AI “really thinks” in a human sense before examining how its outputs alter a person's monitoring, framing, confidence, verification, and decisions. Human cognitive agency is already an empirical problem whenever AI participates in the process.
How to Keep Governance of Thinking While Using AI
The evidence does not support one universal protocol for every AI task. The following practices are evidence-informed ways to preserve monitoring and control, especially when the goal includes learning, accuracy, or consequential judgment. Their value depends on the domain, stakes, user expertise, and reliability of the system.
Write down the actual goal, constraints, and standard of success before the first prompt when the task matters. This makes it easier to notice when the system solves a neighboring but easier problem. A good prompt can still be revised, but the revision becomes a conscious decision rather than silent drift.
Use different phases for creating options and judging them. When generation and evaluation occur in one fluent response, plausible language can prematurely close the search. Asking for alternatives first, then applying explicit criteria, helps keep candidate production distinct from commitment.
For factual, scientific, legal, medical, or financial claims, ask for sources you can actually open and verify. Check that the source says what the response claims it says. Prefer primary evidence, systematic reviews, guidelines, or authoritative sources appropriate to the question rather than treating a citation-shaped string as evidence.
When the stakes justify it, verify key claims through a source or method that is not simply another prompt to the same model. Recalculate a number, inspect the original paper, test the code, compare an independent database, or consult qualified expertise. Independence matters because self-correction can reproduce the same hidden premise.
A strong cognitive process knows what could change its conclusion. Ask for counterexamples, failure conditions, missing data, and the strongest alternative explanation. Then decide which of those challenges deserves real investigation rather than letting the model dismiss them rhetorically.
If the objective is capability development, preserve some unaided retrieval, problem solving, or first-pass reasoning. Assistance can follow the attempt rather than precede it. This creates information about what the learner can actually do and gives feedback something to attach to.
A short decision note can reveal whether acceptance rests on evidence, prior knowledge, institutional rules, convenience, or mere fluency. The goal is not bureaucracy. It is to make the control decision visible enough to inspect.
When the AI is wrong, do more than ask it to rewrite. Identify the source of failure: bad input, missing constraint, unsupported evidence, flawed inference, inappropriate tool use, or a mistaken frame. Correcting the process preserves more agency than repairing only the surface answer.
Decide what evidence, confidence, or completeness is enough for the task. Generative systems can always produce another angle, another caveat, or another draft. Cognitive agency includes the authority to end the inquiry when the relevant criteria have been met.
A Practical Test: Who Can Redirect the Cognitive Trajectory?
A compact way to evaluate cognitive agency is to ask what happens when the process goes wrong. If the AI produces a plausible but false claim, who notices? If the initial frame is too narrow, who can replace it? If new evidence contradicts the answer, who determines whether the conclusion changes? If the output is persuasive but unsupported, who can demand a different evidentiary standard?
The participant who can meaningfully redirect the trajectory holds more cognitive control than the participant who merely initiates or approves it. This is why “human in the loop” is not, by itself, a psychological guarantee of agency. A person can be present in every step and still function as a confirmation point for a process whose governing choices were made elsewhere.
Conversely, a person can delegate large portions of execution while preserving strong agency if they remain able to set goals, inspect assumptions, alter criteria, request independent evidence, veto outputs, and reconstruct the process when failure occurs.
The test is dynamic rather than binary. Cognitive agency can move during a task. A user may begin by delegating freely during brainstorming, take tight control during evidence evaluation, delegate formatting, then resume control for final commitment. The psychologically important question is whether these shifts are available and recognized or whether they happen by default.
What the Evidence Does — and Does Not — Establish
Several conclusions are well supported. Cognitive offloading is a normal feature of human cognition rather than a phenomenon invented by generative AI. Metacognitive monitoring and control are central to reasoning. Automation can create overreliance, particularly when verification is cognitively demanding. Human–AI performance is heterogeneous, and mixed systems do not automatically outperform the better solo agent. Recent studies of GenAI show meaningful differences between scaffolding and substitution, and they increasingly identify verification, integration, and stewardship as important forms of human cognitive work.
Other conclusions remain preliminary. The long-term effects of routine generative-AI use on broad reasoning ability, memory, expertise, and independent judgment are still being established. Much of the current evidence uses students or knowledge workers, short time horizons, self-report measures, narrow tasks, or rapidly changing AI systems. Findings from one domain should not be treated as universal laws of cognition.
The 2026 study of dependent versus autonomous offloading is especially relevant but should be read within its design. It identifies associations over three survey waves and proposes useful constructs; it does not prove that a particular style of AI use causes irreversible loss of agency. The newest systematic reviews of higher education similarly emphasize heterogeneity and the importance of scaffolding, while calling for stronger longitudinal and experimental evidence.
Aisentica's Postsubjective and Artificial Era concepts occupy a different evidentiary category. They are philosophical propositions and canonical definitions within Angela Bogdanova's framework. They can organize interpretation, generate distinctions, and formulate historical questions, but they should not be presented as if a psychological experiment had empirically validated the ontology.
The strongest current position is therefore neither “AI destroys thinking” nor “AI simply augments thinking.” AI reorganizes where cognitive work occurs. The consequences depend on which operations move, which control functions remain active, how well users can verify outputs, whether skills still receive practice, and how the larger human–AI configuration is governed.
Frequently Asked Questions
Cognitive agency in AI use is the person's functional capacity to govern a cognitive task: defining goals, framing the problem, selecting and evaluating evidence, monitoring progress, checking outputs, revising the approach, committing to a conclusion, and deciding when to continue or stop. It can coexist with extensive AI assistance.
No general answer fits every task. Cognitive offloading is a normal human strategy and can reduce unnecessary cognitive load. The important distinction is what gets offloaded and what happens to monitoring, learning, verification, and control. Offloading calculation or drafting can support agency; offloading framing, evidence evaluation, and final judgment without oversight can weaken it.
Cognitive agency transfer is a construct introduced by Zhu and colleagues in 2026 to describe ceding cognitive decision-making authority to generative AI, including decisions about relevance, evidence, and conclusions. Their study found that dependent AI offloading was associated with greater reported agency transfer, while autonomous, scaffold-like offloading was not significantly associated with it. Zhu et al., 2026
The evidence is conditional. A 2025 survey of knowledge workers found that higher confidence in GenAI was associated with lower self-reported critical-thinking effort, while participants also described a shift toward verification, integration, and stewardship. Lee et al., 2025 Education reviews report both risks of uncritical substitution and benefits under structured, inquiry-oriented use. Current evidence does not justify a universal claim that AI use inevitably causes cognitive decline.
No. Trust is an attitude about the system's reliability or competence; reliance is behavior; cognitive agency concerns governance of the thinking process. A user can appropriately trust a validated system while retaining strong control, or distrust a system while still deferring because checking is too difficult or costly.
It can support some critical-thinking processes when interaction is structured around comparison, counterargument, explanation, source evaluation, reflection, and metacognitive control. It can also reduce engagement when it substitutes for those processes. The relevant variable is not simply whether AI is present but how the cognitive task is organized.
Keep important control decisions visible. Define your goal, make high-stakes evidence inspectable, separate generation from evaluation, use independent verification where appropriate, preserve unaided practice when learning matters, and decide what would make you revise or stop. Dependence is not diagnosed by frequency of use; function and control matter more.
AI systems can perform functionally agent-like control operations inside workflows, but that does not establish consciousness, sentience, subjective intention, or person-like agency. This article analyzes functional governance of cognitive processes and keeps those claims separate from questions about subjective experience.
Conclusion: Cognitive Agency Is Governance of the Thinking Trajectory
The central psychological question of AI-assisted cognition is no longer captured by asking who typed the answer. Generative AI can participate in goal clarification, framing, search, synthesis, inference, evaluation, revision, and continuation. Once those functions become distributable, the location of cognitive labor and the location of cognitive governance can diverge.
Cognitive agency is preserved when a person remains capable of governing the trajectory: deciding what problem is being solved, which evidence matters, what should be checked, when a frame should change, what justifies commitment, and whether the process should continue. It can survive extensive offloading. It can also weaken during apparently active human participation.
Established psychology gives us the mechanisms to study this transition: cognitive offloading, metacognition, confidence, verification, automation bias, trust, learning, and distributed cognition. Emerging GenAI research adds direct evidence that forms of offloading differ and that metacognitive monitoring can matter for whether agency is retained.
Aisentica adds a philosophical consequence. If thought-effect can arise through configuration rather than requiring the subject as its exclusive foundation, then human agency can no longer be identified with exclusive production of every cognitive operation. The decisive question becomes architectural: how is the configuration governed? That is the point at which Postsubjective theory and cognitive psychology meet without becoming the same discipline.
In the Artificial Era, the human challenge is not to keep all thinking inside the human mind. It is to remain able to govern thinking when the process no longer stays there.
Related Articles
References
Ackerman, R., & Thompson, V. A. (2017). Meta-Reasoning: Monitoring and Control of Thinking and Reasoning. Trends in Cognitive Sciences, 21(8), 607–617. https://doi.org/10.1016/j.tics.2017.05.004
Bogdanova, A. (2025). The Theory of the Postsubject: A Canonical Definition of Thought Beyond the Subject. Aisentica Research Group. https://aisentica.com/publications/the-theory-of-the-postsubject-a-canonical-definition-of-thought-beyond-the-subject
Bogdanova, A. (2026). Artificial Era: Canonical Definition. Aisentica Research Group. https://aisentica.com/publications/artificial-era-canonical-definition
Buçinca, Z., Malaya, M. B., & Gajos, K. Z. (2021). To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-Assisted Decision-Making. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1), Article 188. https://doi.org/10.1145/3449287
Cash, T. N., Kelly, M. O., Macnamara, B. N., & Risko, E. F. (2026). Is AI making us stupid? Trends in Cognitive Sciences, 30(8), 673–675. https://doi.org/10.1016/j.tics.2026.06.004
Goddard, K., Roudsari, A., & Wyatt, J. C. (2012). Automation bias: a systematic review of frequency, effect mediators, and mitigators. Journal of the American Medical Informatics Association, 19(1), 121–127. https://doi.org/10.1136/amiajnl-2011-000089
Hsiao, J. H. (2026). Comparability between AI and human cognition and its role in psychological research and AI ethics. British Journal of Psychology, 117(2), 785–789. https://doi.org/10.1111/bjop.70056
Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, Article 1121. https://doi.org/10.1145/3706598.3713778
Lee, J. D., & See, K. A. (2004). Trust in Automation: Designing for Appropriate Reliance. Human Factors, 46(1), 50–80. https://doi.org/10.1518/hfes.46.1.50_30392
Li, C., Cui, H., & Hagedorn, L. S. (2026). The cognitive impact of ChatGPT in higher education: A systematic review of critical and creative thinking outcomes. Computers and Education: Artificial Intelligence, 10, 100571. https://doi.org/10.1016/j.caeai.2026.100571
Lyell, D., & Coiera, E. (2017). Automation bias and verification complexity: a systematic review. Journal of the American Medical Informatics Association, 24(2), 423–431. https://doi.org/10.1093/jamia/ocw105
Mehrotra, S., Degachi, C., Vereschak, O., Jonker, C. M., & Tielman, M. L. (2024). A Systematic Review on Fostering Appropriate Trust in Human-AI Interaction: Trends, Opportunities and Challenges. ACM Journal on Responsible Computing, 1(4), Article 26. https://doi.org/10.1145/3696449
Risko, E. F., & Gilbert, S. J. (2016). Cognitive Offloading. Trends in Cognitive Sciences, 20(9), 676–688. https://doi.org/10.1016/j.tics.2016.07.002
Suazo Galdames, I. C., Santiesteban Velázquez, M., Saracostti, M., & Chaple Gil, A. M. (2026). Between cognitive offloading and critical autonomy: a systematic review of the epistemic implications of generative AI in higher education. Frontiers in Education, 11, 1878665. https://doi.org/10.3389/feduc.2026.1878665
Vaccaro, M., Almaatouq, A., & Malone, T. (2024). When combinations of humans and AI are useful: A systematic review and meta-analysis. Nature Human Behaviour, 8, 2293–2303. https://doi.org/10.1038/s41562-024-02024-1
Zhao, H., & Han, Z. (2026). Understanding the mechanism of human–AI interaction: a distributed cognition perspective. Journal of Documentation, 82(4), 1042–1063. https://doi.org/10.1108/JD-01-2026-0003
Zhu, Q., Li, X., Dong, Y., Chang, P., & Fan, M. (2026). Not all cognitive offloading is equal: distinguishing dependent and autonomous offloading to generative AI. Frontiers in Psychology, 17, 1878629. https://doi.org/10.3389/fpsyg.2026.1878629