What Is Cognitive Autonomy?

In the previous commentary we discussed the formation of the intelligent society. As social structures and processes grow ever more complex, humans keep relying on tools and organizations, handing some cognitive activities over to external resources. AI has given part of these activities a computational realization and, through large-scale copying and institutional embedding, takes a continuing part in society's knowledge production, judgment formation and real-world action.

This change poses a fundamental question for education: as people rely more and more on external intelligence to understand the world, solve problems and take action, how should education respond so that people keep what is distinctive about being human and their capacity for autonomy?

The Redbook's answer is cognitive autonomy. So what exactly is cognitive autonomy? This commentary sets out the meaning and scope of the concept in detail.

1. Defining Cognitive Autonomy

The Redbook defines the concept as follows:

Cognitive autonomy is the process by which individuals, in complex cognitive environments, actively regulate their own belief formation, use of evidence, allocation of trust, boundaries of judgment and responsibility for action^[1].

This definition places cognitive autonomy within a continuous process of knowing and acting. Why a person accepts a claim, on what basis they make a judgment, which tasks they are willing to hand over to a system, under what conditions they need to verify again, and how a conclusion will affect real-world action all belong to this process.

For example, when a school uses a system to generate evaluations of students' learning, teachers need to know what data the evaluation relies on, what the scores can show, which aspects of learning may be missed, and what teaching decisions the results are fit to support. Students should also have the chance to understand the evaluation, add material and raise objections. Once an evaluation affects learning opportunities, corresponding review and accountability arrangements are needed. The whole process requires people to be clear-sighted about the data, the system, the consequences of their actions and their own responsibilities, and to manage this cognitive process actively. This process is what we call "cognitive autonomy".

2. Understanding Cognitive Autonomy

1. Cognitive autonomy is a process of actively regulating cognitive activity.

Cognitive autonomy runs through the whole process of obtaining information, understanding knowledge, solving problems, forming judgments and choosing actions. Throughout, individuals need to keep identifying sources of information, assessing the quality of evidence, adjusting their degree of trust, keeping judgments within bounds, and taking responsibility for the actions they take on the basis of those judgments. Cognitive autonomy is continuous and dynamic: it unfolds and adjusts as task goals, the state of the evidence, the external environment and potential consequences change.

2. Cognitive autonomy addresses the complex cognitive environments that pervade modern society.

A complex cognitive environment is one in which information comes from many sources, evidence varies in quality, knowledge is constantly updated, and different judgments compete, and in which the generation, filtering, dissemination and amplification of information are increasingly mediated by algorithms, platforms, organizations and institutions. In such an environment, the content an individual encounters has usually already been selected and processed several times over, and more information does not by itself lead to more reliable understanding.

AI raises the complexity of this environment further. It can rapidly generate content, organize knowledge, offer suggestions and take part in decisions, but it can also produce errors, omissions and unsupported conclusions. Individuals no longer face only the question of whether a given piece of information is true; they also need to judge how the information was produced, what conclusions the evidence can support, in what situations a system applies, and how much trust it deserves. Cognitive autonomy is precisely the foundational capacity needed to face this complex cognitive environment.

3. The core of cognitive autonomy is active regulation.

The activeness that cognitive autonomy emphasizes means that individuals can recognize which information, tools and social mechanisms are influencing their understanding, and actively adjust their cognitive activity according to task goals, the state of the evidence and possible consequences. Individuals can accept reliable information, draw on expert knowledge and use AI, and can also re-verify, revise judgments and adjust actions when conditions change, evidence is insufficient or risks rise.

Cognitive autonomy therefore includes an awareness of risk, but does not stop at warnings about risk. What it pursues is a positive process in which individuals keep understanding, choosing, acting and correcting in complex environments; the ability to carry out this process consistently is what constitutes an individual's capacity for cognitive autonomy.

4. Cognitive autonomy means keeping human cognitive leadership amid open dependence.

Human cognitive activity has always rested on open dependence. Language, knowledge, education, experts, institutions and technical tools all take part in how people come to know things. AI widens the scope of this dependence further, so that information retrieval, organizing materials, pattern recognition, computation and deduction, generating options and expressing content can all be supported by external capabilities. Cognitive autonomy allows people to call upon and organize external cognitive resources selectively as a task requires, without the individual having to complete every cognitive task alone.

Such cognitive dependence is also the basic condition of modern knowledge societies. Hardwig's discussion of epistemic dependence points out that individuals in modern knowledge communities often have to form beliefs on the basis of others' testimony and expert labour^[2]. Hutchins's research on distributed cognition further shows that cognitive activity in complex tasks is often distributed across individuals, tools, symbol systems and organizational environments^[3]. AI makes this dependence denser, faster and less visible, and makes it all the more important for people to organize their relationships of dependence actively.

Cognitive autonomy emphasizes human cognitive leadership within this open dependence, and that leadership shows mainly in keeping control over initiating, reviewing and overseeing the whole cognitive process. Two roles in particular need to stay on the human side: posing questions and checking results. What question is posed determines the goal toward which cognitive activity is directed, and embodies people's understanding of real needs, values and why a problem matters; checking results requires people to judge whether a conclusion is adequately supported, whether it applies to the present situation, whether important factors have been missed, and whether it can be acted upon. Together, these two steps set the direction, boundaries and ultimate validity of cognitive activity.

Extending human capability with AI and maintaining cognitive autonomy can therefore be achieved at the same time. Individuals can make full use of AI's memory, computation, generation and analysis, while continuing to understand how the task is unfolding, which steps AI has taken on, what the results rest on, and which conclusions still need further verification. What cognitive autonomy safeguards is human leadership over the overall direction and validity of cognitive activity, so that AI's capabilities always remain part of a cognitive process that people initiate, people check and people answer for.

3. The Six Dimensions of Cognitive Autonomy

The Redbook unfolds cognitive autonomy into the following six dimensions^[1].

Source vigilance. Individuals need to ask who, or what system, produced a piece of information. Sources may be teachers, experts, papers, media, institutions, platforms, models or commercial systems, or mixed sources produced jointly by humans and machines. Source vigilance requires individuals to avoid accepting or rejecting information simply on the basis of who it comes from, and to go further in analysing how it was produced, the expertise behind it, the incentive structure, accountability and the path by which it spread. Sperber and colleagues proposed the concept of epistemic vigilance, stressing that in communicating with others we must assess the reliability of both the communicator and what is communicated^[4]. The AI environment complicates this, because model output takes the linguistic form of testimony yet lacks the relationships of responsibility found in ordinary human communication.

Evidence sensitivity. Individuals need to distinguish between the form in which a conclusion is expressed and the evidence that supports it. A conclusion may come from personal experience, statistical data, experimental research, expert judgment, historical material, model generation, benchmark tests or commercial promotion. Different kinds of evidence differ in strength and scope. Research on epistemic cognition by Chinn and colleagues points out that learners need to understand how knowledge claims are supported by evidence and how different disciplines establish justification^[5]. In an AI environment, evidence sensitivity means that students do not equate fluent language, realistic images or precise numbers with reliability.

Boundary awareness. AI systems learn from data, and data comes from particular populations, settings, languages, cultures and times. A model that works on one distribution is not guaranteed to be equally reliable in a new setting. Boundary awareness turns the problem of generalization in machine learning into an everyday habit of judgment: does the present case lie beyond the bounds of the evidence? Is the sample representative of the whole? Can experience be transferred? Boundary problems of this kind are extremely common in AI. For example, Recht and colleagues' re-evaluation of the ImageNet and CIFAR-10 test sets showed that models can suffer marked drops in performance on new test sets^[6]. Research of this kind reminds educators that claims about performance need to be understood in connection with data sources and settings of use.

Trust calibration. Individuals need to assign appropriate confidence to their beliefs and choices of action according to the strength of the evidence, the quality of the source, the risk of the task and the consequences of error. Research in psychology and neuroscience shows that people's confidence in their own judgments does not always match their accuracy. Fleming and Dolan's review of the neural basis of metacognition notes that metacognitive accuracy can be dissociated from task performance^[7]. In AI research, Guo and colleagues found that the predicted probabilities of modern neural networks can also be poorly calibrated, so that accuracy and confidence do not naturally agree^[8]. Both people and models can be overconfident, and education needs to train learners to regulate the strength of their confidence.

Cognitive agency. Building on an understanding of AI's basic principles, the limits of its capabilities and its social impact, individuals can regard AI with an open and confident attitude, treat it as a cognitive resource that can be understood, explored and called upon, and think actively about how it can take part in learning, creating, judging and solving problems. Cognitive agency is chiefly a cognitive disposition: a willingness to understand and to try when facing new technologies, and openness to new ways of knowing. In concrete tasks, it also shows in the ability to start from one's own questions and goals, spot the steps where AI might play a role, and envisage possible ways of using external capabilities to extend one's own cognitive reach. It provides the inner drive for subsequent choice of tools, organization of tasks and human–machine collaboration, and puts the regulation of sources, evidence, boundaries, trust and responsibility in the service of active learning, creation and action.

Responsibility judgment. Cognitive autonomy must ultimately enter into action. If a conclusion is used for brainstorming, the standard of evidence required is low; if it is used in medicine, law, finance, educational evaluation or public governance, the standard is much higher. Lee and See's research on trust in automation points out that trust between people and automated systems must match the system's capability, the transparency of its process and the purpose of its use^[9]. The NIST AI Risk Management Framework likewise stresses that AI risks should be managed with reference to specific contexts and the people affected^[10]. Responsibility judgment requires learners to understand that different settings carry different standards of verification, requirements for human review, and arrangements for accountability.

4. The Social Significance of Cognitive Autonomy

Cognitive autonomy has clear social significance. In an intelligent society, judgments seldom affect only the individual. Algorithmic recommendation shapes public discussion, AI-generated content shapes the spread of knowledge, automated evaluation shapes educational opportunity, risk models shape medical resources, and credit scoring shapes financial services; at the same time, AI is lowering the barriers to acquiring knowledge, creating content, developing technology and collaborating professionally. Whether individuals can use AI actively and prudently is closely bound up with public institutions, organizational responsibility, social trust, and the chance to take part in new forms of learning and production.

First, cognitive autonomy bears on the quality of public discussion. Research on fake news by Lazer and colleagues stresses that the spread of misinformation results from the combined effects of technology, cognition, social motives and platform structures^[11]. Vosoughi and colleagues' study of diffusion on Twitter showed that false news may spread faster and further than true news^[12]. As generative AI lowers the cost of producing content, public discussion needs participants with source vigilance and evidence sensitivity all the more. If citizens cannot judge how content is produced and what evidence it rests on, public reason will be weakened by fluent narratives, emotional contagion and platform amplification.

Second, cognitive autonomy bears on the quality of organizational decision-making. Once automated systems enter recruitment, lending, healthcare, educational evaluation and public services, members of organizations need to understand the limits within which model outputs apply. Parasuraman and Riley's research on the use of automation points out that automation can give rise to misuse, abuse and disuse^[13]. Excessive reliance on systems within organizations may lead to the outsourcing of responsibility; rejecting systems altogether may mean missing gains in efficiency and quality. Cognitive autonomy requires professionals to retain the capacity to explain, review and hold to account within organizational processes.

Third, cognitive autonomy bears on individuals' capacity to take part in learning, production and creation. AI can help people enter unfamiliar fields quickly, compare multiple options, cross some professional barriers of expression, and turn preliminary ideas into objects that can be further tested and improved. Cognitive agency makes individuals willing to discover and explore these possibilities, while source vigilance, evidence sensitivity, boundary awareness, trust calibration and responsibility judgment ground the extension of capability in understanding and careful regulation. Cognitive autonomy thus both safeguards people's position as agents and widens the range of practices in which they can take part.

At the same time, cognitive autonomy bears on educational equity. If education in an intelligent society focuses only on operating advanced tools, well-resourced schools may gain platforms and practical opportunities more quickly, while under-resourced schools are left out. With cognitive autonomy as a central aim of general education, low-cost tools, open resources, paper-and-pencil simulations, analysis of everyday situations and human–machine collaborative tasks can all play important roles. Students do not necessarily need the most advanced equipment; they can still learn how to pose questions, choose tools and organize tasks, while understanding data bias, model limits, verification of generated content, recommendation mechanisms and responsibility judgment. Such a positioning helps shift education away from differences in equipment toward building foundational capacities in which every learner can take part.

Finally, cognitive autonomy bears on human dignity and agency. The AI4People principles proposed by Floridi and colleagues emphasize beneficence, non-maleficence, autonomy, justice and explicability, and discuss AI ethics in the context of human well-being and social governance^[14]. From a sociotechnical-systems perspective, Selbst and colleagues criticize abstracted notions of fairness, pointing out that algorithmic systems are always embedded in institutional and social contexts^[15]. Together these studies suggest that general education for an intelligent society must help students understand the relationships between technology and people, institutions and society. Cognitive autonomy enables individuals to keep their position as agents within a technological environment, while understanding that individual judgment must enter structures of public responsibility.

Overall, safeguarding citizens' capacity for cognitive autonomy bears on the foundations on which modern society runs. Modern society rests on citizens who can judge independently, take part in public discussion and take responsibility for their actions. Only when citizens can form beliefs autonomously, examine evidence, regulate trust and recognize the boundaries of judgment do public deliberation, collective decision-making and the functioning of institutions have a genuine foundation in their participants. If citizens' understanding and judgment are increasingly governed by algorithmic filtering, generated content and automated systems, and their capacity for cognitive autonomy is steadily weakened, the civic foundation that supports society will gradually be shaken. Safeguarding cognitive autonomy is therefore both a matter of safeguarding individuals' leadership over their own understanding and action, and a matter of safeguarding the basic order on which public life and the normal functioning of society depend^[1].

5. How Cognitive Autonomy Becomes the Modern Task of General Education

As AI helps people complete more and more tasks, humanity faces a twofold crisis: first, the informational complexity of society is rising rapidly, creating ever more complex information environments; second, as more cognitive tasks are handed off to AI, the risk of cognitive offloading grows.

Education therefore needs to explore a new model for the age of AI. In such a model, what knowledge and skills are learned certainly matters, but what matters more is that learners, through training, improve their cognitive capacities and gain the ability to actively regulate their own belief formation, use of evidence, allocation of trust, boundaries of judgment and responsibility for action. This is precisely the capacity for cognitive autonomy.

The historical mission of general education, to maintain a common cognitive foundation and develop free personhood, thereby acquires a concrete task in the intelligent society: to enable members of society to understand the new conditions of cognition, organize external resources actively, and regulate their understanding and action on sound grounds. Cognitive autonomy is established first as an educational aim, and AI general education organizes knowledge, practice and evaluation under this aim.

The Redbook further organizes the development of these capacities around induction, generalization, judgment, and self-awareness: forming understanding from materials, testing boundaries in new situations, making choices that weigh evidence and values, and continually reflecting on one's own understanding, trust and responsibility. The four capacities provide an educational framework for curriculum and assessment, while the six dimensions described above are used to analyse the concrete process of cognitive autonomy.

6. Conclusion

Cognitive autonomy is a capacity: to keep managing one's own cognitive process effectively in complex environments, to make good use of tools without being swayed by them, and to embrace technology actively while keeping the lead and control over how it is applied.

Cognitive autonomy is an attitude: to keep one's independence in complex environments, to safeguard the human ability to think, and to uphold the dignity of being human.

Cognitive autonomy is a goal, one so important that no voice should be allowed to drown it out: to enable each person to have a complete personhood with which to live in society, and to enable everyone to have such personhood so that human society can continue. This is the first principle of education in an intelligent society.

References

1. [1] Wang D, Ma S, eds. Artificial Intelligence General Education: Cognitive Autonomy in an Intelligent Society[R/OL]. v2.3.3 (draft for comments). Center for Artificial Intelligence General Education (AIGE), Department of Computer Science and Technology, Tsinghua University, 2026-08 [2026-09-26]. https://aige.cs.tsinghua.edu.cn/redbook/
2. [2] Hardwig J. Epistemic Dependence[J]. The Journal of Philosophy, 1985, 82(7): 335–349. https://doi.org/10.2307/2026523
3. [3] Hutchins E. Cognition in the Wild[M]. Cambridge, MA: MIT Press, 1995. https://mitpress.mit.edu/9780262581462/cognition-in-the-wild/
4. [4] Sperber D, Clément F, Heintz C, et al. Epistemic Vigilance[J]. Mind & Language, 2010, 25(4): 359–393. https://doi.org/10.1111/j.1468-0017.2010.01394.x
5. [5] Chinn C A, Buckland L A, Samarapungavan A. Expanding the Dimensions of Epistemic Cognition: Arguments from Philosophy and Psychology[J]. Educational Psychologist, 2011, 46(3): 141–167. https://doi.org/10.1080/00461520.2011.587722
6. [6] Recht B, Roelofs R, Schmidt L, et al. Do ImageNet Classifiers Generalize to ImageNet?[C]//Proceedings of the 36th International Conference on Machine Learning. PMLR, 2019, 97: 5389–5400. https://proceedings.mlr.press/v97/recht19a.html
7. [7] Fleming S M, Dolan R J. The Neural Basis of Metacognitive Ability[J]. Philosophical Transactions of the Royal Society B: Biological Sciences, 2012, 367(1594): 1338–1349. https://doi.org/10.1098/rstb.2011.0417
8. [8] Guo C, Pleiss G, Sun Y, et al. On Calibration of Modern Neural Networks[C]//Proceedings of the 34th International Conference on Machine Learning. PMLR, 2017, 70: 1321–1330. https://proceedings.mlr.press/v70/guo17a.html
9. [9] Lee J D, See K A. Trust in Automation: Designing for Appropriate Reliance[J]. Human Factors, 2004, 46(1): 50–80. https://doi.org/10.1518/hfes.46.1.50_30392
10. [10] Tabassi E. Artificial Intelligence Risk Management Framework (AI RMF 1.0): NIST AI 100-1[R]. Gaithersburg, MD: National Institute of Standards and Technology, 2023. https://doi.org/10.6028/NIST.AI.100-1
11. [11] Lazer D M J, Baum M A, Benkler Y, et al. The Science of Fake News[J]. Science, 2018, 359(6380): 1094–1096. https://doi.org/10.1126/science.aao2998
12. [12] Vosoughi S, Roy D, Aral S. The Spread of True and False News Online[J]. Science, 2018, 359(6380): 1146–1151. https://doi.org/10.1126/science.aap9559
13. [13] Parasuraman R, Riley V. Humans and Automation: Use, Misuse, Disuse, Abuse[J]. Human Factors, 1997, 39(2): 230–253. https://doi.org/10.1518/001872097778543886
14. [14] Floridi L, Cowls J, Beltrametti M, et al. AI4People—An Ethical Framework for a Good AI Society: Opportunities, Risks, Principles, and Recommendations[J]. Minds and Machines, 2018, 28: 689–707. https://doi.org/10.1007/s11023-018-9482-5
15. [15] Selbst A D, boyd d, Friedler S A, et al. Fairness and Abstraction in Sociotechnical Systems[C]//Proceedings of the Conference on Fairness, Accountability, and Transparency. New York: ACM, 2019: 59–68. https://doi.org/10.1145/3287560.3287598