Marxism Research Network
Unofficial English Translation

Yang Yunxia: The Generation and Elimination of Hallucinations in AI-enabled Ideological and Political Education

Xi Jinping Thought (New Era)

Artificial intelligence hallucination, also known as "large model hallucination," generally refers to content generated by a model that is unfaithful to its source information or inconsistent with real-world facts. This includes hallucinations in text generation, image generation, and voice interaction. Hallucinations typically arise when a model, lacking relevant information, generates content through probabilistic selection rather than based on a real-world knowledge base or logical reasoning. This renders its output not only untrustworthy but potentially misleading to users. A study by the British Broadcasting Corporation (BBC) revealed that among responses generated by mainstream AI tools—including Microsoft’s Copilot, OpenAI’s ChatGPT, Google’s Gemini, and Perplexity—more than half contained "significant problems," with approximately one-fifth of the answers introducing obvious factual errors. In January 2025, the "Global Risks Report 2025" released by the World Economic Forum indicated that false and misleading information had ranked as the top short-term risk for two consecutive years, posing a major threat to social cohesion and governance. These risks erode public trust, exacerbate domestic and international divisions, and are being amplified by generative AI. In the field of ideological and political education, AI has a wide range of application scenarios, including personalized learning, intelligent assessment, virtual teaching assistants, interactive pedagogy, big data analysis, emotional recognition and intervention, and interdisciplinary integrated education. Accurately identifying, deeply analyzing, and effectively eliminating these hallucinations is of great significance for fully leveraging the positive role of AI in ideological and political education and promoting its modernization.

I. Categories of AI Hallucinations in Ideological and Political Education

There are numerous categories of AI hallucinations. According to different classification standards, they can be divided into intrinsic and extrinsic hallucinations; closed-domain and open-domain hallucinations; input-conflicting, context-conflicting, and fact-conflicting hallucinations; or factual and faithfulness hallucinations. Considering the particularities of the ideological and political education field and the differences in content, this article categorizes the technical hallucination phenomena produced by large language models and other technologies into three types: factual and epistemic deviations, logical misalignments, and axiological shifts.

(1) Factual and Epistemic Deviations Factual and epistemic deviations refer to instances where AI systems output content that seriously contradicts objective facts while processing information related to ideological and political education. A test conducted on ChatGPT found that among 100 induced false narratives, ChatGPT had a deception rate as high as 80%, resulting in "subtle yet false and misleading statements on major topics (such as the COVID-19 pandemic, the Ukraine crisis, and school shootings)." The types of knowledge covered by ideological and political education are diverse, including facts about current affairs and policies, political theory, morality and ethics, legal literacy, and knowledge of the "Four Histories" [1]; here, the accuracy of information is paramount. For example, when recounting major events in modern and contemporary Chinese history, AI might incorrectly state the time, location, or key figures involved. Such erroneous presentations of historical facts would lead students to construct flawed knowledge systems, seriously undermining the knowledge-imparting function of ideological and political education. This historical misalignment, caused by fragmented training data or erroneous information sources, is essentially a "symbolic simulation" of the cognitive framework of historical materialism by the technical system. Another form of epistemic deviation manifests as improper expression of fixed concepts by certain AI models, such as referring to the "Socialist Core Values" as the "Value System of Socialism with Chinese Characteristics."

(2) Logical Misalignment One of the goals of ideological and political education is to cultivate students' logical thinking abilities, guiding them toward correct values through rational argumentation and reasoning. Logical misalignment in AI output refers to structural logical defects, such as confusion or contradictions in the reasoning of the generated content. This manifests specifically in three ways. First, the breaking of causal chains. For instance, when explaining the Socialist Core Values, AI might mechanically splice conceptual elements together, leading to a "suspended conclusion" phenomenon and a fractured expression of the core value system. This reductionist tendency stems from the cognitive limitations of algorithms regarding the causal relationships within complex social systems. Second, the confusion of axiological coordinates. When processing pluralistic value judgments, algorithms are prone to disordered logical hierarchies, such as placing "individual development" and "collectivism" in opposition. Third, the absence of dialectical thinking. When dealing with contradictory propositions, AI often presents an "either-or" linear logic or exhibits characteristics of "pseudo-dialectics," failing to grasp the relationship of dialectical unity emphasized in Marxist theory. Particularly in value-judgment scenarios, although the semantic coherence index of the model-generated text may be high, the logical self-consistency index is low. This reflects the limitations of algorithmic formal imitation of dialectical logic. The roots of this logical misalignment lie in the limitations of the technical system, including the stripping of historical context from training data, the reductionist treatment of dialectics by formal logic, and the obscuring of the process of ideological evolution by the "algorithmic black box." This hallucination could deconstruct the scientific system of Marxist theory, causing learners to form fragmented and superficial understandings of values during their cognitive construction.

(3) Axiological Shifts Ideological and political education bears the heavy responsibility of fostering students' correct worldview, outlook on life, and values; the correctness of its value orientation is its core. Axiological shifts refer to instances where the content generated by AI, when dealing with value judgments and guidance related to ideological and political education, violates or deviates from the mainstream values advocated by society. This is mainly reflected in: First, the distortion of values. AI generates viewpoints or content that run counter to mainstream social values, such as concluding that liberalism or individualism is supreme when analyzing social phenomena, while ignoring important values like collectivism and the spirit of dedication. Second, deviations in moral judgment. In assessing moral issues, because AI cannot accurately grasp the essence and connotation of morality, it provides explanations and solutions for complex moral dilemmas that do not conform to moral common sense or norms. Third, errors in value guidance. For instance, it may overemphasize the pursuit of material interests while neglecting spiritual enrichment and elevation, leading students to focus excessively on external material enjoyment while ignoring the cultivation of their inner character. Fourth, misunderstandings of cultural values. For example, it may provide one-sided interpretations of certain values in traditional Chinese culture, or fail to correctly convey the core values of different cultures during cross-cultural exchange. The phenomenon of blurred value positions in AI, caused by training on cross-cultural corpora, is particularly prominent.

II. The Technical Roots of AI Hallucinations in Ideological and Political Education

In the context of ideological and political education, the technical roots of AI hallucinations exhibit complex, multidimensional, and interwoven characteristics. From data input to model operation, and from algorithmic design to scenario application, the mechanisms of factual identification and logical reasoning jointly constitute a deep-seated hallucinatory mechanism. These defects in various technical dimensions overlap and intertwine, ultimately giving rise to AI hallucinations in ideological and political education that are both concealed and deceptive.

(1) Bias in Data Quality

  1. Fitting risks from insufficient training data In the field of ideological and political education, the scarcity of training data for AI models triggers multiple fitting risks. First, data sparsity leads to a symbolic cognition of ideological concepts by the model. When training corpora for core concepts such as the Socialist Core Values or the basic principles of Marxism are insufficient, models tend to construct shallow semantic associations through word frequency statistics and co-occurrence relationships, leading to "conceptual hollowing." For example, the vector representation of "a community with a shared future for humanity" might only capture geospatial features while ignoring its political-philosophical connotations. Second, the fracturing of knowledge graphs exacerbates logical fitting biases. The unique dialectical thinking and theoretical system of ideological and political education require support from complete knowledge lineages. If training data is fragmented, the model's causal chain completeness index may decline when constructing complex propositions like the "evolution of the principal contradiction in society," resulting in a breakage of the cognitive framework of historical materialism. Third, the latent risk of alienated value orientations. Incomplete data coverage forces models to rely on probability distributions to fill knowledge gaps. For instance, when involving sensitive issues such as the "comparison between Chinese and Western values," the model may generate content with latent axiological shifts based on statistical advantages. When the proportion of the Theory of Socialism with Chinese Characteristics in the training data is too low, the consistency coefficient of the value position in the model's output will drop below the critical threshold, creating an ideological infiltration risk of "unconscious technical bias." This data-driven fitting defect essentially reflects the technical system's struggle with the formalized simulation of the laws of ideological and political education.

  2. Data pollution from incorrect labeling Data pollution caused by incorrect labeling has become a key risk source threatening the effectiveness of technical applications. This data alienation manifests primarily through three types of structural contradictions. First, conceptual confusion at the ontological level. When texts related to the basic principles of Marxism are incorrectly labeled as general philosophical theories, the model undergoes a conceptual vector shift in its latent semantic analysis, leading to a distortion of semantic representation. Second, value-oriented label inversion. If discourse on Western constitutional democracy is mistakenly labeled as content belonging to socialist political civilization, it triggers a topological distortion of the knowledge graph. Third, pollution through the breaking of contextual associations. The unique historical narrative logic of ideological and political education requires strict chronological labeling of events; incorrect labeling of policy texts causes diachronic misalignment in generated content, leading to a sharp drop in timeline accuracy. This labeling bias essentially reconstructs the technical system's cognitive framework, causing the attention mechanism within the model architecture to suffer from disordered value-weight distribution in the ideological dimension. More seriously, the incorrect parameters solidified through model training on polluted data produce an "intergenerational transmission of cognitive bias." During the data dissemination stage, this may lead to "viral" spreading of polluted data. Because data dissemination is characterized by massive volume, high speed, multi-channel reach, and high penetration, it significantly enhances the dissemination efficiency and social harm of polluted data.

  3. Accuracy loss from data bias Ideological and political education data may have missing values for various reasons, such as historical materials being difficult to obtain in full due to their age, or data omissions caused by technical, human, or institutional factors during collection. Simultaneously, bias may exist in the collection process, with excessive focus on certain regions, groups, or themes while ignoring the diversity of others. In AI-driven analysis systems for ideological and political education, the loss of accuracy caused by data bias has evolved into a technical-ethical problem of the nature of a "paradigm crisis." This cognitive deviation manifests primarily through two deconstructive effects: First, conceptual collapse at the ontological level. When the data coverage of the theory of the Sinicization of Marxism is insufficient, the AI model's theoretical interpretation of the "Two Combinations" [2] will exhibit a breakage of core propositions, forming a mechanical deconstruction of theoretical integrity. Second, the latent subversion of value hierarchies. When the ratio of collectivist to individualist cases in the training data is imbalanced, the model's generated conclusions on "the relationship between the individual and society" will show an implicit utilitarian tendency. Such a shift in the axiological baseline would break through the threshold of ideological security. From the perspective of Marxist epistemology, data is not only the means of production for a technical system but also the material carrier for the dissemination of ideology. The hallucination phenomena exposed in the current process of intelligentizing ideological and political education essentially reflect the deep contradictions in data production relations—meaning that when technical systems process education data with clear value orientations, traditional data governance paradigms struggle to adapt to the special requirements of ideological education.

(2) Defects in Model Understanding Capabilities

  1. The dimension-reduction dilemma of semantic understanding

Language in ideological and political education possesses rich connotations and specific contexts, posing a massive challenge to the semantic understanding of AI models. In the practice of AI-empowered ideological and political education, the simplified processing of complex semantics by algorithms is creating a deep-seated cognitive predicament. When technical systems mechanically disassemble political concepts rich in theoretical depth—such as "a community with a shared future for humanity" or the "Four Confidences" [3]—into word associations and statistical models, they are essentially performing a "flattening" reconstruction of ideological discourse. For example, AI might simplify "whole-process people's democracy" into quantitative indicators like "election frequency" or "number of participants," yet it remains incapable of interpreting the political-philosophical logic of the people’s role as masters of the country [4]. This dimension-reducing understanding leads to "symbolic idling" in educational content, where learners receive only the conceptual shell while the dialectical and historical core of the theory is filtered out by the technology. A more hidden crisis lies in the "decontextualized" processing of semantics by algorithms, which distorts value transmission: when AI breaks down "Socialist Core Values" into isolated words for matching, the system might one-sidedly associate "justice" with economic distribution while ignoring the political-ethical dimension, or limit "rule of law" to legal application while diluting its essential nature of serving the people. This predicament of dimension-reduction in semantic understanding reflects a conflict between technical tools and the laws of ideological education, calling for the establishment of an intelligent educational paradigm with greater humanistic depth.

2. The Risk of Disembedding through Contextual Reconstruction The content of ideological and political education is often closely linked to specific historical, social, and cultural contexts. If AI models fail to fully grasp complex contextual information when processing texts, it impairs their accurate interpretation of educational content. In the process of AI reshaping ideological and political education, the technological deconstruction and reorganization of context are triggering deep-seated risks of cultural rupture. When algorithms transform educational content carrying specific historical memories—such as the "Yan'an Rectification Movement" [5] or the "Great Discussion on the Criterion of Truth" [6]—into decontextualized knowledge modules, the original spatio-temporal coordinates and spiritual core may be quietly stripped away. This narrative reconstruction of spatio-temporal dislocation causes cracks in the chain of transmission for the revolutionary spirit. Especially when dealing with ideologically sensitive topics, the deliberate avoidance of sensitive words by models leads to a break in contextual coherence. Even more alarming is the sluggish response of technical systems to dynamic social contexts; for instance, when AI uses fixed models to interpret the "Principal Contradiction Facing Chinese Society in the New Era" [7], it often ignores the complexity of the real-world context, causing theoretical explanations to hover above social practice. This technical "disembedding" (tuō yù) [8] not only severs the historical depth of theory but also weakens the vitality of values education in generational transmission. The core emotional resonance and value saturation of ideological and political education may thus be quietly dissolved by algorithms.

(III) Algorithmic Deviation from Mainstream Values

1. Erosion of Core Values by Algorithmic Bias The value-embedding mechanism of AI systems is gradually becoming a frontier issue in the philosophy of technology. Research indicates that algorithmic bias has transcended the scope of traditional technical flaws, evolving into a new type of values crisis in the age of digital civilization. This process of value erosion is characterized by a triple-dynamics: value inertia based on historical data, value alienation within the "technical black box," and the value diffusion effect in system applications. Together, these constitute a deep deconstruction of the ethical order of human society. Examined from an ontological dimension, when algorithmic systems transform social reality into objects of calculation through the process of datafication, they inevitably carry specific value presuppositions. Numerous cases confirm that algorithmic systems are not neutral tools; rather, they achieve a "material solidification" of value systems through technical stages such as index selection and feature weighting. The crisis at the epistemological level manifests as a shift in the benchmarks of value judgment triggered by algorithmic bias. Systematic deviations shown by Natural Language Processing (NLP) models in correlation tests such as "rule of law vs. politics" confirm the extension of Foucault’s knowledge-power theory into the digital age—machine learning systems, through the reproduction of power relations in corpora, are reshaping the value coordinate system of social cognition. This hidden process of value shaping causes the public to unconsciously accept an algorithmically constructed "regime of digital truth" during daily technological encounters with search engines and recommendation systems.

2. Algorithmic "Black Boxes" Triggering Ideological Security Risks If AI models over-rely on overseas open-source corpora, they are highly likely to generate numerous cases containing tendencies toward historical nihilism [9]. This typical scenario reveals the risk of value-orientation deviation faced by AI-empowered ideological and political education—when technical logic encounters the complexity of ideological dissemination, algorithmic hallucinations are triggering ideological security crises. From the perspective of the philosophy of technology, the value coding of AI systems possesses a dual nature of alienation. On one hand, the value positions absorbed by pre-trained models through massive data often have implicit conflicts with Socialist Core Values; on the other hand, the uncontrollability of value output caused by the algorithmic black box makes it possible for technical systems to become "Trojan Horses" for Western value penetration. This mechanism of value deviation confirms a core thesis of technocracy: instrumental rationality is eroding the foundation of value rationality in the educational field. Research in the sociology of knowledge reveals a more hidden crisis of cognitive reconstruction: namely, that AI educational systems—through "information cocoons" constructed by knowledge graphs and discourse frameworks set by chatbots—are essentially reshaping the value judgment paradigms of learners, leading to a trend of fragmented value cognition. This algorithmically mediated ideological dissemination results in the hidden transfer of what Foucault called "discursive power" from the educational subject to the technical system. The hallucination phenomena currently exposed in AI systems for ideological and political education are essentially external manifestations of the deep conflict between the algorithmic black box and the laws of education.

(IV) Deviations in Interaction and Application Scenarios

1. Vague Instructions Triggering Inaccurate Outputs Due to the limited expressive abilities of users or a lack of deep understanding of the problem, they may issue vague instructions to AI or harbor other implicit intentions. In the practice of AI-empowered ideological and political education, output deviations caused by vague instructions are forming deep educational risks. When a teacher inputs a broad instruction such as "explain the basic principles of Marxism," AI systems often fall into the trap of "conceptual collage," simplifying "practice is the sole criterion for testing truth" into a linear arrangement of knowledge points, severing its organic connection to the historical context of Reform and Opening-up, thereby resulting in generated lesson plans with fractured theoretical logic. A more hidden inaccuracy is seen in the implicit shift of value stance: when AI is asked to "compare Chinese and Western democratic systems" without a specified methodology, the system may mechanically list institutional features while ignoring the essential attributes of socialist democratic politics, or even allow the penetration of Western discourse such as "the primacy of procedural democracy." Furthermore, this output deviation has transcended the category of technical flaws and may evolve into a trigger for a "crisis of educational subjectivity." Teachers who rely long-term on AI to generate content for vague instructions will gradually lose their originality in theoretical exposition, and the value-guiding power of their classroom teaching will decline accordingly.

2. Incompatibility Challenges Caused by Changes in Application Scenarios In the process of AI-empowered ideological and political education, the dynamic migration of application scenarios is giving rise to a new type of technical adaptation predicament. When AI shifts from standardized classroom knowledge transmission to immersive teaching scenarios that fuse the virtual and the real, the lag in the technical system’s understanding of educational laws is drastically magnified. For example, in a "Red" VR [10] educational scenario, a certain AI platform misread the resolute body language of a revolutionary as a "high emotional volatility index," leading some participants to form cognitive biases regarding the historical figure's character. This adaptation crisis is essentially a conflict between technical instrumentality and educational humanism: algorithms attempt to use fixed parameters to capture the dynamically developing needs of values education, yet they ignore the differences in cognitive construction across different scenarios. In the application of online discussion communities, AI misjudgments of metaphorical expressions and youth subculture symbols frequently occur, often simply classifying critical discussions on terms like "lying flat" (tǎng píng) [11] as negative emotions, thereby blocking the possibility of dialogue for ideological guidance. An even more severe challenge is the distortion of value transmission in cross-scenario migration; for instance, using AI to write a rural field investigation report might transform the issue of education for "left-behind children" [12] into a technical proposition of "insufficient frequency of home-school interaction," obscuring the essence of the structural contradictions between urban and rural areas. This failure in scenario adaptation warns that technological empowerment must be built upon a profound grasp of the political attributes of educational scenarios.

(V) Shortcomings in Logical Reasoning and Abstraction

1. Broken Logical Chains Inducing Absurd Conclusions AI performs reasoning by constructing logical chains. If logical inconsistencies appear—such as a lack of reasonable correlation between premises and conclusions, or jumps and contradictions between reasoning steps—the results will be inconsistent with the actual situation, producing hallucinations. For example, when analyzing the causal relationship of events, the AI might mistakenly treat a chronological sequence as a causal one, leading to unreasonable conclusions. The breaking of logical chains is giving rise to an alarming phenomenon of "conclusion alienation." When a technical system takes a complete proposition like "the leadership of the Communist Party of China is the most defining feature of socialism with Chinese characteristics" and disassembles it into isolated "leadership indices" or "institutional efficiency parameters" for quantitative analysis, the internal logic of the theoretical exposition is quietly dismembered by the algorithm. This logical collapse may evolve into an absurd collage of knowledge in educational practice, leading to a value dislocation across time and space. A more hidden risk lies in "implicit logical substitution": when algorithms replace "causal proof" with "data correlation," the "superiority of socialism" may be transformed into a single-dimensional explanation of economic growth curves, obscuring the complex logical foundation of institutional advantages. This logical deconstruction not only dissolves the theoretical depth of ideological and political education but also, under the guise of technical rationality, quietly shakes the cognitive foundation of values education.

2. Cognitive Disorder Caused by Conflicting Multi-source Information In the process of AI integrating ideological and political education information, data from different sources will inevitably contain differences or contradictions. If the AI cannot effectively fuse and coordinate this information, confusion will arise during reasoning, leading to generated content that deviates from facts—a hallucination phenomenon. In particular, value conflicts in multi-source information may create systematic disorder in the cognitive ecosystem. When an AI system mechanically aggregates heterogeneous information—such as mainstream ideological theory, internet subculture discourse, and Western value concepts—a crisis of misaligned cognitive coordinates seethes beneath the surface of "technical neutrality." This creates the potential to confuse the essential differences between the Socialist Core Values and the value system of "universal values" [13]. When algorithms are trained on a mixture of the "bitter and glorious" history of the Party’s century-long journey and internet deconstructionist narratives, the generated content will show a tendency toward the "entertainmentization of history," perhaps interpreting the Long March spirit as a "reality show of extreme survival challenges," thereby completely dissolving the sublimity of the revolutionary narrative. This cognitive disorder has transcended the scope of information overload and evolved into a technologized path toward the loss of ideological territory, allowing technical tools to alienate into invisible hands that dissolve the authority of mainstream values.

III. Subjective Cognitive Misconceptions Amplifying AI Hallucinations in Ideological and Political Education

At the intersection of technological and educational revolution, the cognitive misconceptions held by educational subjects (teachers and students) are becoming key variables that amplify technical hallucinations. This cognitive alienation stems not only from the shielding effect of the "technical black box" but also, at a deeper level, reflects the loss of value-orientation by educational subjects during the process of technical socialization. When we shift the focus of scrutiny from technical flaws to subject cognition, we find three intertwined cognitive traps reshaping the technical practice logic of ideological and political education.

(I) The Over-Deification of AI as "Omniscient and Omnipotent" by Educational Subjects Enveloped by the trend of technological utopianism, educational subjects are falling into a state of cognitive delirium regarding AI. The formation of this "technological fetishism" is essentially a dimensional reduction in the understanding of the nature of technology—alienating an AI system, which possesses instrumental attributes, into the ultimate arbiter of value and truth. Some educators believe that "around 2050, AI will comprehensively surpass human intelligence"; simultaneously, some learners believe that AI possesses a knowledge reserve superior to humans, while others directly equate AI-generated content with verified truth. In the practice of ideological and political education, this cognitive misalignment is highly likely to evolve into a dangerous surrender of authority: for instance, directly using AI-generated theoretical expositions as standard teaching answers, leading to ideological deviations in the interpretation of core concepts.

The cognitive mechanism of technological deification stems from a triple misjudgment: First, equating data storage with the depth of knowledge. When educators neglect the value-construction process inherent to ideological and political education and delegate the theoretical explanation of "whole-process people's democracy" to AI, the system may output tens of thousands of words of seemingly rigorous discourse; however, its internal logic severs the dialectical relationship between Party leadership, institutional advantages, and the mass line [14]. Second, confusing algorithmic efficiency with educational efficacy. Using AI to instantaneously generate classroom Q&A significantly increases the frequency of teacher-student interaction, yet the depth-of-inquiry index may actually decline. Third, mistaking technological iteration for cognitive evolution. When educators blindly revere large language models like ChatGPT, they often overlook their inherent flaw of "value neutrality." This cognitive alienation forms a vicious cycle in intergenerational transmission: a vast number of young students view AI as a "mobile library of Marxist theory," leading to a marked decrease in time spent independently studying classic works.

The ultimate danger of technological deification lies in the erosion of the educational subject’s capacity for value judgment. For instance, when AI explains the "socialist market economy" as an "efficiency-oriented mixed economic model," the educated may fail to promptly detect the neoliberal tendencies within. This cognitive crisis warns us: when educators surrender the right of theoretical interpretation to algorithms, ideological and political education faces the risk of becoming a technological appendage.

(2) Technological dependence caused by the expansion of technical rationality in the subjects of ideological and political education

Under the dominance of instrumental rationality, educational subjects are undergoing a cognitive metamorphosis from using technology to depending on it. This alienation process follows an evolutionary path of "the temptation of efficiency—competency degradation—subject dissolution," eventually reducing the educator to a mere operating terminal for technical systems. Teachers who over-rely on AI for lesson preparation gradually lose the ability to independently design pedagogical schemes, and their capacity for classroom innovation inevitably declines.

The pathological characteristics of this "dependency syndrome" are manifested in three dimensions: First, technological dependence reconstructs the spatial-temporal structure of educational cognition. When the real-time generation of teaching resources by AI becomes the norm, a teacher's knowledge reserve shifts from diachronic accumulation to instantaneous acquisition. This ultimately causes the integrity of the Marxist theoretical system mastered by the educator to gradually decay and their ability to grasp the Party's innovative theoretical system [15] to decline. Second, technological tools are alienated into the end goal of education. For example, technological dependence leads teachers to equate the optimization of technical parameters with the improvement of teaching quality, trapping them in digital myths such as "click-count worship" or "interaction frequency competitions." Finally, technological dependence leads to a dual dissolution of educational subjectivity. Educators are reduced to data entry clerks, while the educated are relegated to information receivers; the intersubjective dialogue unique to ideological and political education is simplified into a technical process of human-computer interaction.

The typical symptoms of this cognitive alienation lead to a triple result: First, theoretical interpretation loses its depth. For instance, in AI-generated teaching cases involving "common prosperity," discussions remain largely at the level of income distribution and fail to touch upon the essential attributes of socialism. Second, value guidance loses its "warmth." For example, AI heart-to-heart counseling systems based on affective computing process students' ideological confusion into several emotional tags, yet they remain incapable of understanding the value anxiety of the "post-00s" youth. Third, educational practice loses its innovation. In virtual simulation laboratories, the Long March spirit [16] is deconstructed into several task "levels"; while students achieve a 100% pass rate, the index of emotional resonance may actually decrease compared to traditional practical teaching.

(3) Educational subjects becoming intoxicated by the "perfect fusion" of human-machine synergy

Amidst the romantic imagination of human-machine synergy, educational subjects are constructing a dangerous cognitive schema of "frictionless technology." This cognitive utopia views technological intervention as an inevitable choice for educational evolution while deliberately ignoring the essential differences and inherent tensions between human and machine systems. For instance, many educators believe that "AI can perfectly integrate into the entire process of ideological and political education," and some schools have even listed "human-machine synergy" as a core metric for evaluating teaching quality.

In practice, this cognitive bias creates three illusions. First, the illusion of role complementarity. Educators envision AI handling "low-level" tasks like knowledge transmission, while they focus on "high-level" tasks like value leadership. In fact, however, among groups of teachers who use AI for knowledge explanation over the long term, their capacity for value guidance actually decreases—due to the lack of emotional interaction and dialectical interpretation during the knowledge transfer process, the efficiency of students' value internalization also declines. Second, the illusion of efficacy superposition. Although AI significantly increases the volume of classroom information, the integrity of students’ theoretical construction shows a trend of decreasing rather than increasing, characterized by the coexistence of knowledge fragmentation and value suspension. Third, the illusion of evolutionary symbiosis. Some educators expect to achieve co-evolution through continuous human-computer interaction; in reality, teachers who over-rely on technical systems experience a regression in innovative capacity. The essence of this cognitive error is the reduction of educational complexity to a problem of technological adaptability. This also warns us: human-machine synergy is by no means a simple functional complement, but involves a value contestation concerning the very essence of education.

IV. Paths for Eliminating AI Hallucinations in Ideological and Political Education

The deep integration of AI and ideological and political education must transcend the paradigm of technological instrumentalism and construct a value-oriented development path under the guidance of the Marxist view of technology. By establishing a four-dimensional governance system of "data model governance—algorithmic auditing—cognitive revolution—institutional guarantees," we can achieve the dialectical unity of technological empowerment and value leadership, providing a new paradigm for the innovative development of ideological and political education in the New Era.

(1) The Technical Level: Data Model Governance

  1. Strengthening Data Quality Control Building a dam against technological hallucinations at the data source is the core of the data governance system. In the context of ideological and political education, data quality control possesses a dual attribute: it is both a guarantee mechanism for technical precision and a protective shield for ideological security.

Faced with accuracy issues such as the fragmentation of core theoretical data, the homogenization of practical case data, and the blurring of value-orientation data encountered by AI platform enterprises, establishing a precision data filtering mechanism has become an urgent priority. Technically, a five-layer filtering model can be constructed to achieve data purification: the first layer filters out non-authoritative sources, excluding low-quality data such as web forum data, explicitly labeled AI-generated data, and clearly contaminated data; the second layer performs ideological compliance testing to identify potential value biases; the third layer conducts theoretical integrity verification to ensure the systematicity of the basic principles of Marxism; the fourth layer performs contextual restoration to preserve the spatial-temporal characteristics of educational scenes; and the fifth layer completes educational adaptability assessments to match the cognitive patterns of different school stages.

The "quantitative hallucination" caused by insufficient data diversity can be resolved through data expansion. In the temporal dimension, a diachronic database of the developmental history of Marxist theory should be established; in the spatial dimension, unique practical cases from various regions across the country should be collected; and in the morphological dimension, various educational media such as text, video, and physical artifacts should be integrated. By establishing a multi-dimensional data ecosystem, the AI model will have reliable data to draw upon.

  1. Enhancing Model Understanding Capabilities Breaking through the algorithm's formalistic imitation of the laws of ideological and political education requires the establishment of a value-oriented cognitive enhancement framework. Traditional AI models often fall into a "semantic dimensionality reduction" trap when processing the complex concepts of Marxist theory—reducing rich theoretical connotations to mere word co-occurrence relationships. To this end, knowledge networks and mind maps containing core Marxist concepts and the Marxist theoretical system can be constructed to guide the model in establishing logical associations within political discourse.

Strengthening contextual understanding is key to eliminating technological hallucinations. Through the three-layer architecture of the AI model, the model's sense of "historical orientation" can be enhanced: the first layer embeds "era-characteristic" tags to automatically identify the historical context of a text; the second layer constructs event-causal chains to restore the practical logic of theoretical development; and the third layer implants a value-analysis dimension to ensure the value consistency of interpretations.

The cultivation of value-judgment capabilities requires breaking through the cognitive myth of "algorithmic neutrality." A "value alignment" training framework can be utilized to refine Socialist Core Values into several evaluation dimensions. Through reinforcement learning, the model can be guided to establish value priorities, thereby enhancing its value-judgment capacity.

  1. Developing Multimodal AI Models The limitations of single-modality technology are particularly prominent in ideological and political education. Taking the promotion of the Yan'an Spirit [17] as an example: when a model only processes text data, its understanding of the Yan'an Spirit lacks the concrete support of visual symbols, and its interpretation of the lyrics of the Yellow River Cantata [18] lacks the emotional transmission of auditory impact. By constructing a "Red Multimodal Large Model" with a three-layer fusion architecture, these limitations can be overcome: the bottom layer establishes a cross-modal alignment space to achieve semantic mapping between images, text, audio, and video; the middle layer constructs a situational understanding module to restore the multi-dimensional context of historical events; and the top layer develops a value-reinforcement network to ensure the value consistency of multimodal expressions.

Of course, the deep integration of multimodal technology also faces challenges. First, the cross-modal semantic gap requires the development of alignment algorithms with ideological sensitivity. Second, the difficulty of dynamic adaptation to educational scenes requires the model to sense the modal needs of the teaching environment in real time. Third, computational ethical risks necessitate the establishment of a value-review mechanism for multimodal outputs. However, as long as we maintain the unity of instrumental and value rationality, and always take the educational task of "striving to cultivate more 'young people of the era' who put the Party's mind at ease, are patriotic and dedicated, and shoulder the heavy responsibility of national rejuvenation" as the value coordinate for technological innovation, AI can truly become an empowerer rather than a deconstructor of ideological and political education in the New Era.

(2) The Management Level: Algorithmic Auditing

  1. Implementing Manual Review and Intervention In ideologically sensitive fields, manual review is not a technical auxiliary but a necessary defense line for maintaining educational sovereignty. Purely relying on algorithmic auditing leads to the leakage of content with ideological bias. Therefore, a dual-track review system is extremely necessary: primary review is completed by AI through keyword screening and semantic alerts, while advanced review is conducted by institutions composed of experts to perform value judgments. In implementing manual intervention, the grasp of timeliness is particularly crucial. A three-stage response mechanism can be used to move intervention nodes forward: first-stage intervention occurs during the algorithm training phase, where ideological and political education experts can be invited to participate directly in engineering design; second-stage intervention occurs during the model deployment phase, establishing a dynamic monitoring system for teaching scenarios; and third-stage intervention occurs during the content generation phase, setting a "Red Emergency Stop Button" to achieve real-time blocking of abnormal outputs.

  2. Introducing Official Data and Expert Knowledge The injection of authoritative knowledge is a key path to resolving the "value suspension" of algorithms. By implementing the "Red Knowledge Embedding Project," the state can lead the construction of an official data pool containing three major knowledge modules: the classic Marxist literature database, the Party's innovative theory database, and the database of practical cases with Chinese characteristics. Furthermore, through expert knowledge distillation technology, the interpretation processes of authoritative theoretical experts can be systematically recorded and organized, extracting several value-judgment rules to be injected into the algorithm to successfully correct the cognitive biases common to models in the field of Marxist theory. Given the practical difficulties of knowledge transformation barriers, it is extremely necessary to establish an expert collaboration mechanism. Expert knowledge can be deconstructed into four computable dimensions: conceptual networks, logical paradigms, value coordinates, and contextual rules. Among these, priority rules for Socialist Core Values should be explicitly set to effectively improve value consistency.

  3. Regularly Updating AI Models The dynamic evolution of the educational context requires the establishment of a "biological clock mechanism" for model updates. A real-time update can be achieved through a four-dimensional update system: setting quarterly mandatory update nodes in the temporal dimension; continuously collecting new practical samples from various regions in the spatial dimension; tracking the development of the Party's innovative theories in the theoretical dimension; and integrating the latest algorithmic breakthroughs in the technical dimension. Through real-time updates, the timeliness of AI model interpretations can be comprehensively improved. At the same time, update quality must be guaranteed through verification mechanisms. For example, "Red-Blue Team Testing" can be established to simulate ideological contestation scenarios: the "Red Team" inputs the Party's innovative theories, while the "Blue Team" designs test cases for Western value penetration. This allows for the successful identification of new rhetorical variants in Western ideological penetration, driving the upgrade of model defense strategies and effectively improving the ability to defend against ideological risks.

  4. Increasing the Transparency and Explainability of AI Models Cracking the "algorithmic black box" is a fundamental project for rebuilding educational trust. A three-layer explanatory framework for a decision-traceability system can be established to achieve algorithmic transparency: the first layer displays the knowledge sources of the generated content, accurate even to document paragraphs; the second layer displays the visualized value-judgment path, presenting the distribution of value weights step-by-step; and the third layer reveals potential cognitive biases, marking major risk warning signs one by one.

The deepening of explainable technology requires overcoming the difficult problem of visualizing value-based reasoning. This can be achieved by developing "Red Decision Tree" technology, which transforms the generative processes of artificial intelligence into explanatory paths consistent with Marxist methodology. For instance, by mapping nodes from problem definition, contradiction analysis, and historical positioning to theoretical interpretation and value judgment—with each node linked to theoretical foundations and practical cases—one can successfully trace deviant paths in models and promote algorithmic optimization.

(III) Educational Dimension: Cognitive Revolution

1. Correctly Understanding the Essence of AI Technology

The profound transformation of ideological and political education empowered by artificial intelligence is, in essence, a digital transition of the educational cognitive paradigm. As technical tools reshape the basic logic of knowledge production and value transmission, educational subjects [19] urgently need to reconstruct the "coordinates of human cultivation" in the technological era through a cognitive revolution. Naturally, dispelling the fog of technological cognition is the ideological prerequisite for preventing educational alienation. The "technology neutrality theory" and "algorithm omnipotence theory" currently prevalent in the educational field are, in fact, the metaphysicalization of technological cognition. The root of this cognitive deviation lies in the neglect of the social nature of technology—when educators ignore the capital logic, cultural hegemony, and technological politics hidden within AI training data, they essentially allow technological power to penetrate ideological frontiers.

Under the guidance of the Marxist view of technology, the cognitive framework of the subject needs to achieve breakthroughs in four dimensions. First, at the level of technological ontology, it must reveal the social-relational attributes of AI systems. By analyzing the capital control mechanisms of recommendation algorithms and the Western discourse hegemony in natural language processing, educational subjects can recognize the essence that "algorithms are politics." Second, at the level of technological axiology, it must critique the myth of "value neutrality." Through training in technological critical thinking, educational subjects can enhance their ability to identify value deviations in AI content. Third, at the level of technological methodology, it must establish the educational standpoint that "technology serves humanity." By training educators on the dialectical relationship between "technological instrumentality" and "educational subjectivity," the educational subject can truly establish the conscious awareness of "using technology rather than being used by it." Fourth, at the level of technological praxis, it must improve the technical reflexivity of the educational subject. This involves deconstructing technological myths by analyzing the limited game space behind AlphaGo’s victory over human players; dissecting how social algorithms shape youth values to reveal technological power; and discussing effective paths for human-machine collaboration to explore technological transcendence.

2. Enhancing the AI Literacy of Educators and Students

The structural defects in current AI literacy education are primarily manifested in the disconnection between skills training and value education, between knowledge transmission and critical thinking, and between theoretical learning and practical innovation. Therefore, cultivation should not stop at operational skill training but should construct a three-dimensional literacy system of "technological cognition–value judgment–innovative application."

Constructing an AI literacy education system with Chinese characteristics requires strengthening four pillars. First, a tiered education system: the basic education stage focuses on technical cognition and information security; the higher education stage emphasizes value criticism and innovative application; and the teacher development level carries out the cultivation of AI educational leadership. Second, curriculum integration and innovation: by offering AI ethics courses, technological criticism can be embedded into professional teaching. For instance, the capital logic of recommendation algorithms can be analyzed in the "Basic Principles of Marxism" course, and the ethical boundaries of affective computing can be discussed in the "Ideological and Moral Character and the Rule of Law" course. Third, practice-empowered platforms: reshaping literacy education through practical scenarios, such as deconstructing the value-judgment mechanisms of ChatGPT models, developing AI assistants for ideological and political education, and conducting human-machine "value games." Fourth, the driving force of evaluation reform: by establishing AI literacy development archives, the traditional quantitative evaluation model can be transcended, implementing process-based evaluation across three dimensions: depth of technical understanding, reliability of value judgment, and effectiveness of innovative application.

(IV) Ethical and Value Dimensions: Institutional Guarantees

1. Embedding Ethical Rules into Algorithms

Traditional governance of technological ethics often relies on post-hoc reviews and passive correction; however, in the scenario of ideological and political education, ethical rules must be transformed into the underlying logic of algorithmic operation. This can be achieved by establishing a multi-level governance model. At the "Value Core Layer," a Marxist ethical knowledge graph should be constructed to translate abstract ethical principles into several computable parameters. At the "Decision Logic Layer," an ethical decision tree based on dialectical materialism should be developed, requiring the algorithm to pass through logical checkpoints such as "historical legitimacy review," "realistic rationality verification," and "value righteousness testing" [20] before generating content. At the "Output Control Layer," a dynamic ethical review gateway should be established to monitor generated content in real-time via neural networks. When high-risk expressions are detected, the system should automatically trigger a "Red Circuit Breaker Mechanism" to block the dissemination of content and mark it for review.

2. Establishing Institutional Guarantees Targeted at Mainstream Values in Algorithms

The ideological security struggle in the intelligent era has extended to the algorithmic dimension. In the digital space of pluralistic value contests, artificial intelligence must become a firm disseminator of mainstream values rather than a passive carrier. Only by deeply embedding Socialist Core Values into the value-encoding layer of technical systems can we ensure that AI truly becomes a strategic tool for consolidating the mainstream ideology. To defend the mainstream values of society and eliminate the uncontrollable risks of generated content, a "Five-Sphere Integrated" institutional guarantee system should be constructed.

First, establish an algorithm filing system for value embedding. China’s "Interim Measures for the Management of Generative Artificial Intelligence Services" clearly requires "adhering to Socialist Core Values." Subsequently, all generative AI models could be required to have a built-in Socialist Core Values database during the training phase, eliminating algorithmic bias through adversarial training and filing core value codes with regulatory departments for review. Second, construct a multi-party collaborative three-dimensional regulatory network. Establish a governance matrix composed of technical ethics committees, third-party certification agencies, and user supervision platforms to monitor the entire lifecycle of AI output. By adopting a risk-graded regulatory model, a dual mechanism of "real-time content auditing + user warnings" can be implemented for social chatbots. Third, perfect the legal framework of the digital space. Legislation should clearly define the value dissemination responsibilities of AI service providers and establish a traceability mechanism for AI-generated content; in AI dialogues, non-human identity must be declared to provide a legal basis for value correction. Of course, while injecting "value rationality" into AI through a legal framework, institutional innovation must also consider the inclusiveness of technological development. As General Secretary Xi Jinping pointed out: "Artificial intelligence is a strategic technology leading this round of scientific and technological revolution and industrial transformation, possessing a strong 'bellwether' effect of spillover-driven growth," and "accelerating the development of the new generation of artificial intelligence is a strategic issue concerning whether China can seize the opportunities of the new round of scientific and technological revolution and industrial transformation." Only by adhering to the dialectical unity of instrumental rationality and value rationality can artificial intelligence truly become the "light of civilization" benefiting humanity.

(Author’s affiliation: School of Marxism, Northwestern Polytechnical University) Source: Marxist Theoretical Teaching and Research, 2025, No. 3 Web Editor: Jing Mu