Ten Deviations in AI General Education Courses

The previous article discussed the distinctive value of the AI general education course, and why it needs independent standing within the curriculum. This article discusses ten common deviations in the current practice of AI general education.

It should be said at the outset that all of our discussion consists of judgments made from the standpoint of AI general education. These judgments are not evaluations of the worth of the practices concerned; they are meant to clarify where those practices diverge from AI general education in goal and direction.

1. The Tendency toward Tool Training

Tool training is the deviation that arises most easily in AI general education: the use of AI tools becomes the main content of teaching, such as how to converse with a large model, how to generate slides, images, videos and official documents with one click, how to do vibe coding with WorkBuddy, how to build workflows, and so on.

Such classes are highly practical. What is learned can be used at once and the results are tangible, which is why many AI courses are designed this way.

However, an AI course of this kind cannot be called an "AI general education course".

Tool use is of course worth teaching; it helps students learn, express themselves and solve practical problems. Yet however skilled, tool use cannot produce a systematic understanding and basic knowledge of AI, still less train the capacity for cognitive autonomy. Worse, if the classroom trains only tool use and does not exercise thinking, learners are trained to be more willing to accept the system's answers unconditionally. This does not exercise human cognitive ability; it damages it.

At the practical level, AI is developing rapidly, tools are replaced quickly, and what has been learned soon becomes obsolete. Learners who do not understand the basic principles behind the tools can only passively accept each upgrade. The likely result is that they are worn out by keeping up, waste a great deal of time and cost, and lose confidence the more they learn.

The approach offered by AI general education is this: convey to learners, within the course, the basic principles behind the tools, and show students how to find and assess relevant resources so that they can quickly learn to use new tools.

Behind this model, in which "systematic learning of principles" drives "self-directed learning of tool use", is the idea of teaching people to fish rather than giving them fish.

In future, as tools become ever more intelligent, they will also become ever more convenient to use, and will not need to be deliberately studied. Courses whose main content is tool use not only cannot serve as AI general education courses; their own value will also steadily decline.

2. The Tendency toward Technical Training

The opposite kind of deviation from tool training is excessive technical training: for example, taking the core content of specialized AI courses, such as programming, algorithms and model training, simplifying it, and spreading it as a general education course.

AI knowledge and technology are necessary content for helping learners understand the internal mechanisms of AI systems. Learning Python, classifiers, neural networks, data annotation and model tuning, and carrying out machine learning projects, are reasonable in themselves.

However, technical learning alone cannot achieve the goal of AI general education.

Students may train an image classifier without understanding how training data reflects social bias; may implement a recommendation algorithm without understanding how the recommendation objective affects human attention; may know that a neural network is made up of layers and parameters yet be unable to judge whether a piece of AI medical advice needs to be reviewed by a doctor; may be able to write code yet not know how to explain the boundaries of responsibility in high-risk settings. Technical knowledge is an important foundation of AI general education, but technical knowledge does not automatically generate social judgment and cognitive autonomy.

AI general education needs to reorganize knowledge. Basic content such as data, representation, learning, reasoning, models and feedback should be retained, and should be developed around understanding the mechanisms of intelligence, recognizing the limits of capability and forming independent judgment. Programming and model training can serve as means of cultivation, but their depth and demands should be subordinate to the goal of general education.

For example, in studying the same classification model, a general education course can vary the samples and the test conditions so that students understand the sources and limits of experience, and reflect on whether their own judgments are likewise shaped by limited experience. Such a transformation of a course involves redesigning the learning content, tasks and assessment, and goes much deeper than "teaching the specialized course in a simpler way".

3. The Tendency toward Hollowing Out Knowledge

Opposite to the tendency toward technical training is the tendency toward hollowing out knowledge. Those who take this view hold that general education is about cultivating thinking and literacy, and need not go deeply into AI knowledge. The course therefore mainly covers the history of AI, stories of scientists, the names of frontier technologies and application cases, and the social impact of technology; students are organized to read materials, exchange impressions and discuss ethical issues, while the principles and methods of AI itself are passed over lightly.

Such a course is easy to understand and convenient to discuss, and it lowers the demands on teachers' specialized knowledge and teaching preparation. However, without systematic learning of AI knowledge, the teaching goal of "cognitive autonomy" is hard to reach.

The reason is simple: cognitive autonomy has to rest on a solid foundation of AI knowledge. Without understanding the basis of a model's judgment, it is hard to analyse its errors; without understanding the relation between training data and the setting of application, it is hard to judge its scope of applicability. Lacking this knowledge, so-called independent judgment easily becomes subjective opinion, and so-called critical thinking easily becomes blind doubt.

The knowledge meant here is not terms and concepts, but an understanding of the basic methods and underlying logic of AI. Some courses introduce concepts such as data, algorithms, models and neural networks, but do not explain the relations among them or analyse how a system operates, so students still do not understand why AI can accomplish a task, or why it makes mistakes. Learning of this kind is equally inadequate.

Therefore, the key content that ought to be studied in depth must be studied in depth. How data and knowledge are represented, how rules and models take part in reasoning, how machines learn from samples, how a system's capabilities and limitations take shape: these form the basis for understanding AI, and cannot be deleted or made shallow in the name of "general education".

For different stages of schooling, a course can adjust its mathematical demands, technical detail and task difficulty, but it must retain the key relations among principles and the necessary process of learning. Accessibility should be reflected in the way things are explained, the choice of cases and the steps of teaching; it must not come at the cost of depth of knowledge.

In short, AI general education must maintain sufficient depth of knowledge, so that learners, through solid study of the discipline, gradually form well-grounded understanding and judgment. The AI general education course must not be taught as a "history class", a "language class" or an "ethics and law class" that merely takes AI as its topic.

4. The Tendency toward Activities

In many schools, AI education first appears in the form of popular science lectures, science festivals, summer camps and project showcases. This path is reasonable in practice. When a new course enters a school, it often begins with pilot interest activities, building up teachers' experience and student cases. However, such activities are not the normal form of AI general education.

First, these activities are short. They are lively at the time and forgotten afterwards, and cannot produce a systematic and comprehensive understanding of AI.

Second, such activities can usually reach only a small number of students, and are hard to extend into general education that covers everyone.

Third, an activity-based approach also tends to bring "success-case bias": projects with good results receive more attention, while the trial and error that really tempers thinking goes unvalued.

What AI general education needs is not scattered activities, but a scientific and complete system of knowledge. With such a system in place, corresponding activities can be designed. Activities of that kind have vitality, and are also more likely to yield results.

5. The Tendency toward Competitions

Some schools adopt the approach of "promoting learning through competition": they select students who are interested in AI, form clubs and competition teams, concentrate teachers, equipment and extracurricular time on training, and then use competition rankings and award-winning works to display the school's achievements in AI education.

Competitions can motivate students to learn in depth, and help to discover and cultivate outstanding talent. Awards are clear-cut and works are easy to see, so they readily become the evidence by which schools display their achievements. Competitions can serve as a forerunner to the full roll-out of AI general education, and also as a supplement to it.

However, this approach of "promoting learning through competition" cannot be equated with AI general education.

First, competition is aimed only at certain students; most students are onlookers. A school may win many awards, but these have nothing to do with the majority of its students.

Second, many current competitions are concerned with programming and modelling, and have nothing to do with the core mission that AI general education pursues, "cultivating the capacity for cognitive autonomy".

Third, if a school's funding, teaching staff and class hours are all arranged around competition teams, this may lead to a distorted view of achievement in which "awards come first", crowding out the educational resources of general education and trading the public interest for the accomplishments of a few students.

What we advocate is that AI general education should first secure the common learning of all students, and then offer those with interest and ability opportunities for deeper development such as clubs, projects and competitions. The two kinds of education can support each other, but should be assessed separately: competition results reflect the level of exploration of some students, while general education must examine the basic knowledge and cognitive abilities acquired by all students.

In short, a school must first ensure that every student receives the cultivation that is due; the competition results of a few cannot stand in for the general education of the many.

6. The Tendency toward Frontier-Chasing and Fragmentation

AI is developing rapidly, and new terms such as large models, agents, multimodality and embodied intelligence keep emerging. Some courses make "keeping up with the frontier" their main feature, introducing new concepts, new results and new applications.

Frontier content can attract students, and keeps a course connected with real technological development. For a field developing as quickly as AI, it is plainly inappropriate to leave content unrevised for long.

However, if a course chases hot topics from beginning to end, basic knowledge is easily neglected. Students have heard the names of many models and products, yet have not understood the relations among data, models, training and generation. When the next batch of new technologies appears, they still need to hear an introduction from the beginning, and find it hard to understand on their own by relying on what they already know. This not only fails to help build basic abilities, but also leaves learners more confused because of fragmentation.

Frontier demonstrations also tend to lead to over-interpretation and produce cognitive bias. AI can write poems, paint, generate videos and control robots. These results are very appealing, but the data, settings and constraints on which a demonstration depends are often hard to see. Without basic principles and analysis of failures, students easily mistake amazement at a demonstration for understanding of the technology, and overestimate a system's capability in real settings, which in turn brings cognitive bias.

AI general education should of course introduce frontier advances, but the precondition is a stable structure of knowledge, so that frontier results can be classified, analysed and understood. For example, when introducing agents, one can connect perception, task, action and feedback; when introducing generative models, one can draw on the basic mechanisms of data, learning and generation. Frontier content thus becomes material with which students apply knowledge and make sense of new phenomena, rather than a spectacle glanced at in passing.

At the same time, the AI general education course should also keep updating its structure of knowledge. Today's frontier content may become basic technology after a while. This was true of deep learning, and it is true of large models. Continually organizing and refining the system of knowledge, through analysis and synthesis of how technology has developed, is a long-term task of AI general education.

7. The Tendency toward Ethics as an Add-On

AI ethics has entered most frameworks for AI education. Issues such as privacy protection, algorithmic bias, fairness, transparency, deepfakes, copyright, academic integrity and high-risk applications appear frequently in textbooks and courses.

Pointing out risks and discussing ethical issues are a proper part of AI general education. However, many courses merely arrange one session on ethics and safety at the end, and remind students to use AI responsibly.

Such an arrangement is easy to carry out, and can cover the common requirements of ethics education. That students know AI carries risks is of course better than their having no awareness of risk at all.

However, knowing about a risk is not the same as being able to deal with it; knowing that one should "protect privacy" does not mean being able to judge what is reasonable practice when data is collected; understanding that "models may be biased" does not mean understanding where the bias comes from and how to improve matters.

The larger problem is that ethical reminders may convey a self-contradictory view: earlier the course shows only how powerful AI is, and later it warns of how dangerous AI is. Students have understood neither the principles by which capability takes shape nor the mechanisms by which risk arises, and easily swing back and forth between blind trust and vague fear.

AI general education should place ethics and responsibility within the concrete process of learning. This does not mean that ethical reminders should pop up at every moment of learning; it means discussing the causes and effects of ethical problems from basic principles, and carrying out targeted training of thought and hands-on practice.

For example, when introducing a speech recognition system, one can compare the amounts of data for different accents and analyse the relation between sample size and recognition rate; discuss, in connection with the model's scalability, the performance gaps for older people or children; and use credibility to judge the impact of such gaps.

In this way, students can explain risk from principles, and make judgments and choices according to concrete conditions, so that ethical discussion is well grounded and does not become empty talk.

8. The Tendency toward Shallow Assessment

Another prominent problem of AI general education is that an assessment system matched to its goals has not yet taken shape. Many courses assess students by concept tests, tool operation, project works, competition results and classroom presentations. Each of these has its function: concept tests can check basic knowledge, tool operation can check practical ability, project works can show integrated application, and competition results can reflect the creativity of some students.

However, measured against the goals of general education, their shortcomings appear at once: these methods cannot assess whether students can explain why an AI output may be wrong, whether they can transfer a framework of judgment to a new setting, whether they can recognize insufficient evidence and boundary conditions, or whether they can reflect on their own dependence on AI.

Specifically, shallow assessment shows itself in the following ways:

The first is a focus on terminology. Students can recite concepts such as supervised learning, training set, neural network, generative AI and algorithmic bias, but may not be able to use them to analyse real cases. Memorizing terms is the starting point of learning, but cannot be the end point of AI literacy. The educational value of AI concepts lies in helping students explain phenomena and make judgments; a list of terms can only provide the first threads of knowledge.

The second is a focus on works. After students complete an AI project, assessment often concentrates on whether the interface is attractive, whether the functions run, and whether the presentation is smooth. The quality of a work easily masks the process of learning. In a generative AI environment in particular, tool assistance can markedly raise the completeness of a work, and from the work alone it is hard to judge the student's own understanding and gains. If assessment looks only at results, students may receive high marks while lacking problem definition, verification of evidence, error analysis and judgment of responsibility.

The third is a focus on competitions. Competitions can motivate outstanding students, and can also drive schools to invest resources. But competition results are not suited to representing the general education development of all students. General education emphasizes a common foundation for all learners, and assessment should cover different levels of ability, different stages of schooling and different ways of learning. If assessment relies too heavily on competitions, AI general education will concentrate on a few students, and fairness and universality will suffer.

For AI general education, assessment should centre on the fundamental goal of the course, namely whether the capacity for cognitive autonomy has really been formed.

The Redbook divides this capacity into induction, generalization, judgment, and self-awareness. This provides workable dimensions for assessing learning outcomes. Induction examines whether students can sum up regularities and explain phenomena; generalization examines whether students can apply what they have learned to new situations; judgment examines whether students can weigh evidence, risk and responsibility; self-awareness examines whether students can reflect on their own process of judgment.

Such assessment includes both standardized tests and project-based tasks. For the latter, assessment should be process-oriented. Students need to explain the background and process by which a work came into being, make clear the division of labour between person and AI and the share each contributed, and answer challenges from the teacher and from other students.

Assessment of this kind takes more time and demands more of teachers' professional ability, but it fits the goals of AI general education better. If AI general education is to develop learners' capacity for cognitive autonomy, assessment cannot look only at results; it must reach into the process of thinking.

9. The Tendency to Substitute "Empowerment" for "Education"

Some schools report "AI-empowered education" as "AI general education", to show that AI-related work has already been carried out. At some conferences and forums on the theme of "AI education", a large share of the content is "AI-empowered education".

It should first be made clear that AI education (AI Education) and AI-empowered education (AI for Education) are two entirely different things.

AI education is a design for teaching: it teaches AI knowledge and trains AI skills. AI general education is AI education reshaped with "cognitive autonomy" as its goal.

AI-empowered education is a direction of technology: it uses AI tools and systems to improve the quality and efficiency of teaching, covering every aspect related to the organization and processes of education, including course preparation, the conduct of teaching, the collection of learning data, personalized tutoring and school management.

For education, both AI education and AI-empowered education are important, but empowered education cannot be used in place of education itself; to do so is a conceptual error.

At the same time, the two are not unrelated.

In a sense, AI education, and AI general education in particular, is the foundation of AI-empowered education. When AI general education proceeds well, it can raise the overall AI literacy of teachers and students, including their ability to use AI tools. With this foundation, teachers and students become more adept when they use AI tools for teaching and learning, with more confidence, a clearer sense of direction and more command, and the results are more evident.

Conversely, using AI tools to assist teaching and learning is itself an important practical component of AI general education. How to judge the credibility of AI output, how to review content and decide what to keep, and how to use AI in a compliant and reasonable way that accords with ethical norms: all of these train the capacity for cognitive autonomy, and all confirm and feed back into AI general education.

10. The Tendency toward Superficial Teacher Training

The quality of AI general education depends to a large extent on teachers, but much current teacher training remains superficial. A great deal of it takes as its main content lesson preparation with AI, generating courseware, setting and marking exercises, and making office work more efficient. Such training does meet certain practical needs in teachers' work, and can help teachers become familiar with the new technological environment. But it cannot give teachers the grounding and ability needed to teach an AI general education course.

The AI general education course first requires teachers to have command of the basic principles, basic ways of thinking and basic knowledge framework of AI, and to have a fairly deep understanding of AI applications and the AI frontier. Without these as a foundation, one cannot step up to the podium.

In particular, once the goal of this course is calibrated to "cognitive autonomy", the demands on teachers rise sharply. Teachers need a thorough understanding of AI, the ability to design cases for cognitive training, and, further, the skill to organize inquiry-based classes.

AI general education therefore needs a dedicated system for preparing teachers. Training should help teachers build basic concepts and a structure of knowledge, learn to design tasks that train analysis, judgment and reflection, and learn the corresponding methods and means of assessment.

Conclusion

It should be stressed once more that all of the practices described above have their own value, and all are important explorations in AI education.

What we have discussed is that, seen from the standpoint of "AI general education (as distinct from AI education)", these practices cannot fulfil general education's mission for our time of "cultivating the capacity for cognitive autonomy"; there is a misalignment and deviation in their goals.

We also stress that these deviations are not a "departure"; on the contrary, a great deal of this work has already laid a solid foundation for cultivating cognitive autonomy.

For most earlier practice, all that is needed is a calibration and an alignment: to fix the goal of cultivation clearly on "cultivating the capacity for cognitive autonomy", and to address the fundamental question of "what makes a human being human" in the intelligent society.

Only in this way can the course be worthy of general education's mission for our time, and only in this way is it qualified to become a general education course that everyone can study and everyone should study.