The Intersection of Epistemology and Artificial Intelligence - Conceptual Distinctions, Theoretical Challenges, and Future Landscapes
I. The Foundation of Knowledge: An Epistemological Framework
A. Defining "Rènshìlùn" and "Zhīshìlùn": Distinctions and Core Questions
When discussing the philosophical dimensions of knowledge, it is first necessary to clarify the concepts of "Rènshìlùn" and "Zhīshìlùn" within the Chinese context. The English term "epistemology" has historically been translated into Chinese as "认识论" (Rènshìlùn). However, in contemporary Chinese philosophical discourse, the connotations of "Rènshìlùn" and "Zhīshìlùn" differ.
"Rènshìlùn" (认识论) leans more towards "how a cognitive subject acquires information from an object," focusing on tracing and reenacting the dynamic cognitive process. Its core questions are closer to the domains explored by cognitive science or physiology, such as studying how an individual processes external stimuli through the sensory and nervous systems to form perceptions and representations of things. This perspective focuses on the "process" of knowledge acquisition.
In contrast, "Zhīshìlùn" (知识论) is more concerned with questioning the basis for the legitimacy of a static belief itself. It explores what makes a belief qualify as "knowledge," focusing on the relationship between justification, truth, and belief. The Western philosophical tradition of "epistemology" or "theory of knowledge" primarily addresses these issues, such as the nature, origin, and scope of knowledge, as well as justification and the rationality of belief. Its goals include distinguishing "justified" beliefs from "unjustified" ones, separating "knowledge" from "rumor," and finding negative evidence to overturn existing knowledge claims.
This conceptual distinction reveals two different paths for examining the core concept of "knowledge." On one hand, there is the investigation of cognitive faculties and information processing flows; on the other, there is the scrutiny of the normative basis and validity of knowledge claims. The development of artificial intelligence (AI), especially its ability to simulate human cognition, process information, and even generate "knowledge-like" outputs, makes both sets of questions particularly salient. How AI systems "learn" and "process" information relates to the dynamic process that "Rènshìlùn" focuses on. Whether the content output by an AI system is credible and constitutes "knowledge" directly touches upon the core issues of "Zhīshìlùn." Without a clear distinction between these two research approaches, confusion can arise when discussing the relationship between AI and knowledge, especially in cross-linguistic and cultural communication. For instance, when evaluating the "intelligence" of an AI model, the distinction between focusing on its efficiency and complexity in information processing (akin to "Rènshìlùn") versus the accuracy and defensibility of its outputs (akin to "Zhīshìlùn") is significant.
B. The Classical Tripartite Definition of Knowledge: Justified True Belief (JTB)
In the Western epistemological tradition, the most classic and influential definition of knowledge is "Justified True Belief" (JTB). This concept can be traced back to Plato's discussions in the Theaetetus. This definition holds that a subject S knows a proposition P if and only if:
- P is true (Truth);
- S believes P (Belief);
- S's belief in P is justified (Justification).
This definition emphasizes that merely believing something that is true is not sufficient to constitute knowledge. For example, a patient with no medical knowledge who believes they will recover soon cannot be said to "know" they will get better, even if they do, because their belief lacks adequate justification. Therefore, justification is the key element that distinguishes knowledge from accidentally true beliefs. A core task of epistemology is to clarify what constitutes "proper justification." This classical definition remained popular into the early 20th century, with figures like Bertrand Russell still holding this view in his works, and it was accepted by most philosophers until the mid-20th century.
The three components of the JTB framework—belief, truth, and justification—have traditionally been heavily anthropocentric. Belief is usually understood as a conscious mental state or propositional attitude; truth is often understood as the correspondence of a proposition with objective reality; and justification involves the reliable functioning of human reason, perceptual experience, or cognitive faculties. This definition of knowledge, built on the model of the human mind, inevitably faces profound challenges when confronted with artificial intelligence that can process information and produce complex outputs. Can AI have "beliefs"? On what standard is the "truth" of its output based? Can its internal operational processes constitute a valid form of "justification"? These become key questions for subsequent discussion.
C. The Challenge to the Classical Definition: The Gettier Problem
Although the JTB definition has historically been dominant, its sufficiency was severely challenged in the second half of the 20th century. In 1963, Edmund Gettier published a short paper presenting the famous "Gettier Problems," which powerfully demonstrated that JTB is not a sufficient condition for knowledge. Gettier constructed counterexamples to show that in certain situations, even if a person's belief is justified and true, they still do not have knowledge of it.
Gettier's counterexamples typically rely on two premises: first, the justification condition allows a person to be justified in believing something false; second, if P entails Q, S is justified in believing P, and S deduces Q from P and accepts Q, then S is also justified in believing Q. In these counterexamples, the subject derives an accidentally true belief from a justified false belief through valid reasoning. In such cases, although the three conditions of JTB are met, we intuitively would not consider the subject to have knowledge, because the truth of their belief involves an element of luck. For example, a person sees their colleague Jones driving a Ford and Jones tells them he owns a Ford, so they are justified in believing "Jones owns a Ford." They have another colleague, Brown, whose whereabouts are completely unknown to them. From this, they infer "Either Jones owns a Ford, or Brown is in Barcelona" (a logically weaker disjunctive proposition). As it happens, Jones does not actually own a Ford (he is driving a rental), but Brown does happen to be in Barcelona. In this case, the person's belief that "Either Jones owns a Ford, or Brown is in Barcelona" is true and justified (through valid logical deduction), but they do not have knowledge of it.
The emergence of the Gettier problem prompted epistemologists to reconsider the definition of knowledge and attempt to supplement JTB by adding a fourth condition (such as a "no-defeater condition" or a "reliability condition"). This challenge is particularly important for understanding AI's "knowledge." As some scholars have pointed out, AI systems, especially large language models (LLMs) that generate information based on statistical patterns, may produce content that happens to be true in certain cases. Users may also believe this content and even consider the AI's authority as a form of "justification." However, this correctness could be accidental, not stemming from the AI's cognitive reliability or a genuine grasp of the facts. This means that even if a user forms a "justified true belief" based on an AI's output, they may fall into a Gettier-style predicament, having acquired a true belief by luck, which is not genuine knowledge. Therefore, when evaluating the "knowledge" generated by AI, it is necessary not only to focus on the truth of its output and the user's belief in it but also to deeply examine the nature and reliability of its "justification" process, guarding against AI becoming a new source of "Gettier cases." This requires a deeper inquiry into the "justification" of AI outputs, demanding a "meta-justification" concerning the AI's own processes, reliability, and potential for accidental correctness.
D. Major Epistemological Traditions: Rationalism and Empiricism
On the question of the origin of knowledge, two major traditions have formed in the history of Western philosophy: Rationalism and Empiricism.
Empiricism emphasizes the dominant role of sensory experience in the formation of ideas and the acquisition of knowledge. It holds that knowledge must ultimately be traced back to an individual's sensory experiences and cannot be derived solely from innate ideas or traditional deduction. John Locke, George Berkeley, and David Hume are representative figures of empiricism. Empiricists believe that the mind at birth is a "blank slate" (tabula rasa), and all ideas and knowledge come from postnatal experience. This idea has had a profound impact on the methodology of the natural sciences, emphasizing the testing of knowledge claims through observation and experimentation.
Rationalism, on the other hand, holds that reason is the primary source of knowledge, emphasizing the acquisition of knowledge through independent thought and logical reasoning. It asserts that some knowledge can be a priori, that is, independent of experience. René Descartes, Baruch Spinoza, and Gottfried Leibniz are the main representatives of rationalism. Descartes, through "I think, therefore I am," attempted to establish an indubitable rational foundation for knowledge. His rationalist epistemological tradition has had a far-reaching influence on later generations; for example, Husserl's phenomenology was deeply inspired by Cartesian meditation. Some rationalists also acknowledge the role of experience in the formation of knowledge but believe that rational principles are a necessary prerequisite for organizing and understanding experience.
It is worth noting that the opposition between rationalism and empiricism is not absolute, and many philosophers have adopted aspects of both views. For example, Immanuel Kant attempted to reconcile the two, arguing that knowledge is the product of the joint action of sensory experience and the categories of the understanding. The French philosopher Gaston Bachelard developed a "non-Cartesian epistemology," attempting to transcend traditional rationalism by viewing the evolution of knowledge as a historical process, involving an evolution from naive realism through classical rationalism to super-rationalism.
Interestingly, these two classical epistemological traditions seem to find a modern echo in the different development paths of artificial intelligence. Some scholars have pointed out that Symbolic AI, which emphasizes the explicit expression of knowledge, logical rules, and reasoning, has commonalities with the rationalist emphasis on a priori principles and logical deduction. In contrast, Connectionist AI, which emphasizes learning patterns and associations from large-scale data (i.e., "experience"), aligns with the empiricist emphasis on the accumulation of experience and inductive learning. This correspondence provides a useful entry point for understanding the different technical paradigms of AI and their inherent views on knowledge from an epistemological perspective. If Symbolic AI and Connectionist AI respectively embody certain core features of rationalism and empiricism, then the hybrid methods emerging in the current AI field, such as Neuro-Symbolic AI, can be seen as an attempt at the computational level to integrate these two epistemological paths, aiming to build more comprehensive intelligent systems that can both learn from experience and perform symbolic reasoning. This not only reflects the internal needs of AI technology development but also echoes the historical efforts in philosophy to transcend and integrate the dualism of the sources of knowledge.
II. Artificial Intelligence: Paradigms and Knowledge Processing
A. Defining Artificial Intelligence: Goals and Key Capabilities
Artificial intelligence (AI) is a field dedicated to building artifacts, whether simulating animals or humans, that exhibit intelligent behavior. It encompasses a wide range of subfields and technologies, including reasoning, knowledge representation, planning, learning, natural language processing, perception, and robotics. The long-term goal of AI is to achieve artificial general intelligence (AGI), the ability to perform any intellectual task that a human can. To achieve this goal, AI researchers integrate various techniques, including search and mathematical optimization, formal logic, artificial neural networks, statistical methods, operations research, and economics.
One of the core directions of AI research is deduction, reasoning, and problem-solving. Early AI research directly mimicked human step-by-step logical reasoning, similar to the thought processes in board games or logical proofs. As it developed, especially in handling uncertain or incomplete information, AI made significant progress in the 1980s and 1990s by drawing on concepts from probability theory and economics. Another core capability is knowledge representation, which aims to enable machines to store corresponding knowledge and to deduce new knowledge according to certain rules. This involves how to effectively organize and apply a large amount of knowledge about the world—including pre-stored a priori knowledge and knowledge obtained through intelligent reasoning—in AI systems.
The development of AI presents a duality: it is both the creation of practical tools to solve real-world problems, such as intelligent assistants, recommendation systems, and autonomous driving, and a scientific exploration aimed at simulating, understanding, and even replicating human (or other biological) intelligence. This duality directly affects its epistemological evaluation. If AI is viewed purely as a tool, we might focus more on the reliability and efficiency of its outputs and how these outputs serve human knowledge goals. But if it is viewed as a model of intelligence, it raises deeper questions: What is the state of "knowledge" within an AI system? Does it "understand" the information it processes? What are the similarities and differences between its "learning" process and human cognitive processes?
Much of what is called "knowledge" in AI systems is often human-predefined rules, facts, or patterns learned statistically from data, rather than knowledge independently acquired by the AI through a process similar to human understanding and experiential interaction. For example, the "a priori knowledge" in a knowledge base is endowed to the machine by humans in a specific way. The "knowledge" learned by a neural network is implicit in its connection weights, a reflection of data patterns. The origin and nature of this "knowledge" are fundamentally different from the knowledge humans acquire through active cognition, social interaction, and cultural transmission. This suggests that when we use traditional epistemological frameworks (like JTB) to examine AI, AI will face severe challenges in meeting the criteria for core concepts like "belief" and "understanding."
B. Mainstream AI Paradigms and Their Epistemological Presuppositions
1. Symbolic AI (Logic-Based): Explicit Knowledge Representation and Reasoning
Symbolic AI, also known as logicism or "Good Old-Fashioned AI" (GOFAI), advocates for building artificial intelligence systems through axioms and logical systems. Its core idea is that intelligent behavior can be achieved through the manipulation of symbols that represent facts and rules about the world. In the view of symbolicists, artificial intelligence should mimic human logical methods to acquire and use knowledge. Knowledge is stored in an explicit, human-readable form of symbols and logical statements (such as production rules, semantic networks, frames, and ontologies). The reasoning process is based on the deductive, inductive, or abductive rules of formal logic, solving problems and generating new knowledge through the manipulation of these symbols.
Symbolic AI is committed to formalizing knowledge and reasoning processes, and its epistemological presuppositions have much in common with rationalist philosophy. It emphasizes the importance of a priori knowledge (encoded into the system in the form of rules and facts) and logical deduction. Its advantages lie in the clarity of its knowledge representation and the transparency of its reasoning process, making the system's decision-making process theoretically explainable and verifiable. This has similarities to the requirements in epistemology for the clarity and traceability of justification. However, Symbolic AI also faces inherent limitations. It struggles to handle ambiguous, uncertain, or incomplete information, has poor adaptability to the complexity and dynamism of the real world, and the construction and maintenance of its knowledge base often require extensive intervention by human experts, making it difficult to scale to large, open domains. This reliance on precisely formalized knowledge and its "brittleness" in handling novel situations reflects, from one perspective, the tacit dimensions of human knowledge that are difficult to fully symbolize, relying on intuition, common sense, and contextual understanding. This suggests the limitations of purely logic-based systems in fully capturing human knowledge and reveals that, in addition to explicit logical reasoning, human cognition involves a large amount of tacit knowledge and cognitive abilities that are difficult to formalize.
Furthermore, a deep epistemological challenge facing Symbolic AI is the "symbol grounding problem." That is, how do the symbols manipulated within the system acquire their meaning in the real world? If symbols are defined solely through their relationships with other symbols, the entire symbolic system may become disconnected from the external world, becoming a purely syntactic game lacking true semantic understanding. This raises a fundamental question: can a system whose "knowledge" consists of ungrounded symbols truly possess knowledge about the world? Or is it merely performing formal computations?
2. Connectionist AI (Neural Networks): Implicit Knowledge and Pattern Recognition
Connectionist AI, also known as the bionic school or the neural network school, advocates for achieving artificial intelligence by imitating the connection mechanisms of neurons in the human brain. It does not rely on pre-programmed explicit rules but instead learns patterns and associations from large-scale data by constructing networks of a large number of interconnected artificial neurons. In connectionist models, knowledge is not stored in an explicit symbolic form but is implicitly distributed in the connection weights and activation patterns between neurons. These weights are formed through learning and adjustment from training data. Reasoning (or more accurately, information processing) is achieved through the parallel distributed propagation and activation of input data in the network, ultimately producing an output. It focuses more on pattern recognition, associative memory, and learning complex relationships from data.
Connectionist AI, especially the rise of deep learning, has achieved great success in fields like image recognition and natural language processing. Its epistemological presuppositions are closer to empiricist philosophy. It emphasizes learning from "experience" (i.e., training data), where knowledge is acquired a posteriori and is probabilistic and contextual. The strength of connectionist models lies in their ability to process high-dimensional complex data, discover subtle patterns in data, and their capacity for learning and adaptation. However, Connectionist AI also brings new epistemological challenges. The most prominent is the "black box problem": due to the extreme complexity of the internal workings of neural networks, their decision-making processes are often difficult for humans to understand and explain. We may know that a model has made a certain prediction or decision, and that this decision is statistically accurate, but we are unclear how and why it made that decision.
This trade-off between performance and cognitive transparency poses a challenge to the traditional concept of justification. If the justification of a belief must be accessible or understandable to the subject (or at least to a human evaluator), then the output from a black box model, even if "true," has a questionable "justification" status. This forces us to consider whether we can accept a purely performance-based "justification," or whether we need to develop new epistemological frameworks to evaluate the knowledge claims of such systems.
Furthermore, the extreme dependence of Connectionist AI on training data makes the reliability of its knowledge closely tied to the quality, bias, and completeness of the data. The training data becomes the new "epistemic authority," but if the data itself contains biases, errors, or is incomplete (e.g., contains "data voids" or reflects social inequality), the AI system will not only learn these flaws but may also amplify and entrench them, leading to its generated "knowledge" systematically deviating from facts or being discriminatory. This highlights the epistemological responsibility in data management and algorithm design.
Table 1: A Comparison of the Epistemological Characteristics of Symbolic AI and Connectionist AI
| Feature Dimension | Symbolic AI | Connectionist AI |
|---|---|---|
| Nature of Knowledge | Explicit | Implicit |
| Knowledge Acquisition | Primarily through human programming and knowledge engineering of rules and facts | Primarily by learning patterns and associations from large-scale data |
| Reasoning/Processing | Based on logical deduction, rule matching, and symbol manipulation | Based on data-driven pattern recognition, associative learning, and parallel distributed processing |
| Transparency/Explainability | Relatively high, reasoning steps are traceable | Relatively low, often called a "black box" |
| Uncertainty Handling | Typically based on deterministic logic, or requires specific mechanisms to handle uncertainty | Inherently probabilistic, handles noise and uncertainty well |
| Dependence on A Priori Knowledge | Highly dependent on predefined knowledge bases and rules | Initially less dependent on a priori structural knowledge, mainly relies on training data |
| Corresponding Philosophical Tradition (Analogy) | Rationalism | Empiricism |
| Main Advantages | Precision, explainability, handling structured knowledge and complex reasoning | Adaptability, strong pattern recognition, handling large-scale unstructured data |
| Main Disadvantages | Brittleness, knowledge acquisition bottleneck, difficulty with ambiguity and novel situations | Opacity, requires large amounts of data, can overfit, difficulty with abstract symbolic reasoning |
This table clearly outlines the core differences in knowledge processing between the two mainstream AI paradigms and their epistemological implications, laying the groundwork for subsequent discussions on AI's challenges to traditional epistemology. These differences are not just choices of technical paths but are deeply rooted in different answers to fundamental questions like "What is knowledge?" and "How is knowledge acquired and represented?"
C. Knowledge Representation and Reasoning in AI Systems: Mechanisms and Limitations
Knowledge Representation (KR) in artificial intelligence systems aims to create an effective computational form or medium in which thinking can be accomplished within a computational environment, improving practical efficiency by guiding the organization of information. KR is a core research problem in the AI field, with the goal of enabling machines to store corresponding knowledge and to deduce new knowledge according to certain rules (mainly logical reasoning rules). Knowledge in AI systems can be distinguished into "pre-stored a priori knowledge" (endowed to the machine by humans in some way, such as describing objects, features, relationships, events, rules, etc.) and "knowledge obtained through intelligent reasoning" (acquired by combining a priori knowledge with reasoning rules).
Reasoning and problem-solving in AI are among its core objectives, with researchers dedicated to developing algorithms that can mimic the steps of human problem-solving and logical reasoning. Early AI research imitated human step-by-step deductive reasoning, while later methods were developed to handle uncertainty using concepts from probability and economics. However, for complex problems, the reasoning process can face a "combinatorial explosion" challenge, where the required computational resources (storage or time) grow exponentially with the size of the problem.
Although AI has made progress in knowledge representation and reasoning, its mechanisms and limitations also raise profound epistemological questions. First, the "knowledge" in AI systems is usually a formal modeling of human knowledge or data patterns, rather than the AI's own intrinsic understanding of concepts. AI systems manipulate these symbolized representations (such as logical propositions, ontological concepts, vector embeddings, etc.), rather than directly grasping the real-world meaning these representations refer to. As some critics have pointed out, mainstream knowledge representation research tends to mechanically dissect the complex process of human knowledge formation, retaining only the "knowledge" as a result of intelligent activity for secondary information processing, while ignoring the living source of intelligent activity, namely the dynamic cognitive process and meaning generation. This means that AI's "reasoning" is more of a symbol/data transformation based on preset rules or learned patterns, which is fundamentally different from human reasoning based on understanding and meaning.
Second, AI knowledge representation faces inherent incompleteness and uncertainty. It is almost impossible to build a complete knowledge base that covers all relevant world knowledge. The correctness of a priori knowledge needs to be verified, and this knowledge is often not simply black and white but is full of ambiguity and context dependency. This reflects the limitations of human knowledge itself and also restricts the capabilities of AI systems built on this knowledge. AI systems, no matter how powerful their computational capabilities, cannot transcend the fundamental constraints of their representation mechanisms in terms of coverage and precision. While probabilistic methods provide a path for handling uncertainty, they also introduce new problems, such as how to ensure the accuracy and robustness of probability estimates. These limitations remind us that AI systems are not omniscient in knowledge processing, and the validity and reliability of their "knowledge" are always constrained by their representation mechanisms and the quality of the "evidence" they receive.
III. The Impact of AI on Traditional Epistemology
A. Can AI "Possess Knowledge"? Re-evaluating Belief, Truth, and Justification in the Context of AI
Applying the traditional "Justified True Belief" (JTB) definition of knowledge to artificial intelligence systems immediately encounters a series of fundamental difficulties. AI systems, especially current large language models (LLMs), challenge every component of the JTB definition through their mode of operation.
First is the "Belief" condition. In the human context, belief is usually understood as a mental state with intentionality and consciousness. However, current AI systems, including LLMs, are widely considered to lack these attributes. They are complex programs that operate based on algorithms and statistical patterns, and their outputs do not stem from subjective "belief" or "conviction." AI "does not intend to assert anything. It does not aim to tell the truth, persuade, or deceive. Its output is generated based on probabilistic patterns in data." If "belief" is a necessary condition for knowledge, and AI cannot possess beliefs in the human sense, then by definition, AI cannot possess knowledge. Unless we are willing to redefine "belief" as a purely functional state (for example, a system that stably outputs a certain proposition), this would be a major revision of traditional epistemological concepts.
Second is the "Truth" condition. LLMs and other AI systems generate text by learning the statistical regularities in massive amounts of data. The "truthfulness" of their output is often based on its conformity to patterns in the training data, rather than a direct reflection or deep understanding of external objective reality. This "truth" is more like a "statistical truth" or "probabilistic truth" based on the internal coherence of the data, rather than the "correspondence theory of truth" in the traditional sense, which emphasizes the correspondence of propositions with facts. More seriously, AI systems can produce "plausible but false" outputs (i.e., "hallucinations"), or the correctness of their output may be merely accidental or "lucky," not derived from a reliable grasp of the facts. This accidental correctness is precisely the weak point of the JTB framework revealed by the Gettier problem.
Finally, there is the "Justification" condition. Traditionally, justification involves providing reasons, evidence, or relying on reliable cognitive processes. However, the internal workings of AI (especially deep learning models) are often opaque "black boxes." Even if an AI's output is true, it is very difficult for us to access its "justification" process. A user might "believe" that an AI's output is justified due to its authority or past performance, but this justification may be missing, incomprehensible, or based merely on statistical probability rather than logical reasoning or direct evidence. Furthermore, AI lacks the intentionality, accountability, and normative foundation inherent in the human justification process. An AI system cannot be held responsible for the truth of its assertions in the way a human can.
In summary, if we strictly adhere to the JTB framework, AI faces severe compliance issues in the three core dimensions of "belief," "truth," and "justification." This suggests that either AI does not currently have the ability to possess knowledge, or the traditional JTB framework is insufficient to properly evaluate the cognitive state of AI, requiring the development of new epistemological concepts and tools.
Table 2: AI's Challenges to the "Justified True Belief" (JTB) Framework
| JTB Component | Traditional Understanding | Challenges from AI Systems (especially LLMs) |
|---|---|---|
| Belief | A propositional attitude with intentionality, a conscious mental state | AI lacks consciousness, intentionality, and true mental states; its output is algorithmic, not what the AI "believes" |
| Truth | Correspondence with objective reality, factuality | AI output is based on statistical patterns in training data; its "truthfulness" may be correlational or accidental, not reliably corresponding to external reality; risks of "hallucinations" (false information) and reflecting data biases exist |
| Justification | Reasons, evidence, reliable cognitive processes, accountability | AI's internal processes are often opaque "black boxes"; its "justification" may be missing, inaccessible, or based only on statistical probability rather than logical reasoning or direct evidence; AI lacks the intentionality and accountability of human justification |
This table visually demonstrates the specific challenges AI poses to each component of the JTB framework, highlighting the necessity of re-examining the definition of knowledge in the age of AI. These challenges are not minor technical issues but touch upon the very foundations of epistemology.
B. AI-Generated Output as "Testimony": Issues of Intentionality, Accountability, and Reliability
A large part of human knowledge comes from the testimony of others. We believe historical records, scientific reports, news articles, and the statements of others in daily communication. The legitimacy of testimonial knowledge relies on a series of presuppositions about the provider of the testimony, such as their intention to convey the truth, their possession of relevant cognitive abilities, and their being, to some extent, accountable for their statements. However, when AI systems, particularly large language models, generate information that is accepted by users as "testimony," a series of profound epistemological problems emerge.
The most central issue is that AI-generated output lacks the intentionality, accountability, and normative foundation inherent in human testimony. Human testimony is a conscious act of communication, usually with the intention of conveying true information, and is constrained by social norms (such as the principle of honesty). A testifier who deliberately provides false information or causes errors through negligence usually bears corresponding responsibility. In contrast, AI systems operate according to algorithms and data patterns; they have no subjective intention to "inform" or "mislead." Their output is the result of their design and training, not a conscious choice. When an AI provides incorrect information, we cannot attribute it to the AI's "deception" or "negligence," because AI does not possess the subjectivity to bear such moral or cognitive responsibility.
This difference makes it extremely difficult to directly apply traditional testimonial epistemology to AI-generated content. Users may uncritically accept AI's "testimony" due to the fluency of its output, its apparent authority, or the influence of anthropomorphic bias and the social normalization of technology. They may directly adopt it as knowledge, thereby bypassing necessary cognitive diligence and justification processes. This phenomenon brings significant cognitive risks: if the AI's output is erroneous, biased, or merely statistically plausible "nonsense" (i.e., "hallucinations"), then beliefs formed based on such "testimony" will be unreliable and even harmful.
Therefore, in the face of AI-generated "testimony," we need to establish a new framework for cognitive trust. We cannot simply treat AI as an equivalent provider of testimony to humans. Trust in AI output should not be based on an assessment of its "intentions" or "character" (as these are absent), but rather should rely more on a systematic evaluation of its design process, the quality of its training data, algorithmic transparency, historical performance, and reliability in specific application scenarios. This means that the responsibility for evaluating AI "testimony" falls more on the designers, deployers, and users of AI themselves. Users need to cultivate critical thinking and verification habits regarding AI output.
C. Large Language Models (LLMs): Stochastic Parrots or the Dawn of Understanding?
Large language models (LLMs) like the GPT series have become the focus of current AI development due to their ability to generate highly coherent, contextually relevant, and seemingly informative text. They have also sparked intense debate about their cognitive state. A core question is: do LLMs truly "understand" language and the content they process, or are they merely performing complex statistical pattern matching, like "stochastic parrots" that simply repeat combinations of patterns they have seen in massive training data?
The "stochastic parrot" view holds that LLMs mimic language by predicting the next most likely token in a sequence, but they do not "know" what they are saying. Their impressive linguistic abilities are achieved through statistical pattern matching, lacking the deep, interconnected network of intentional states (such as beliefs, desires, intentions) that underlies human language understanding. This view emphasizes that LLM output is a success at the syntactic level, not a true understanding at the semantic level. They can replicate the subtle nuances present in their training data, but this ability stems from learning the data distribution, not from a grasp of concepts and world knowledge.
However, some researchers are trying to move beyond this purely syntactic manipulation theory to explore whether and to what extent LLMs "understand" or "represent" meaning. Some views suggest that although LLMs' understanding is different from humans', they may capture some form of meaning by learning the distributional relationships of words in a large amount of text (i.e., "distributional semantics"). Some scholars even try to use philosophical theories like "inferentialist semantics" to explain LLM behavior, arguing that meaning lies in the role of words in an inferential network, and LLMs may acquire some degree of "understanding" by learning this inferential role. Other research explores whether LLMs can somehow "ground" their linguistic symbols in a broader context or (in multimodal models) non-linguistic data, thereby acquiring meaning that transcends purely textual associations.
This debate about whether LLMs "understand" actually touches on deeper philosophical questions about the definitions of "meaning" and "understanding" themselves. If we strictly define understanding as a cognitive state unique to humans, based on consciousness and intentionality, then LLMs clearly do not have understanding. But if we adopt a more functionalist or behaviorist stance, where understanding can be measured by a system's performance on specific tasks, then LLMs seem to exhibit some degree of "understanding-like" behavior in certain aspects.
Regardless of the true cognitive state of LLMs, the text they generate can easily create an "illusion of understanding" in users. Because their output is linguistically highly complex and natural, users can easily anthropomorphize them, over-attributing knowledge, reliability, and intention to them. This illusion is a significant cognitive risk, potentially leading users to uncritically accept LLM suggestions, information, or conclusions, and thus make wrong judgments or decisions. Therefore, from an epistemological perspective, even if LLMs can develop a deeper level of "understanding" in the future, it is still crucial at the present stage to maintain a cautious and critical attitude towards their output and to recognize the essential differences between their capabilities and human understanding.
D. The "Chinese Room Argument" and Its Significance for Modern AI
John Searle's "Chinese Room Argument," proposed in 1980, is a landmark thought experiment in the philosophy of artificial intelligence. It aims to challenge the view of "Strong AI," which holds that a correctly programmed computer can have cognitive states equivalent to the human mind (such as understanding and consciousness).
The Chinese Room Argument imagines a person who does not understand Chinese (Searle himself) locked in a room. In the room, there is a rulebook written in English and a large collection of Chinese symbols (as a database). People outside the room pass in questions written in Chinese through a window. The person inside the room follows the instructions in the English rulebook to find and match the Chinese symbols, and then passes out the corresponding combination of Chinese symbols as an answer. Searle points out that although the room system (including the person, the rulebook, and the symbol library) can pass the Turing test, making people outside believe there is someone who understands Chinese in the room, the person inside the room (Searle) never understands any Chinese. He is only performing pure symbol manipulation (syntactic operations) without grasping the meaning of these symbols (semantics).
This argument has profound real-world significance for modern AI, especially large language models. LLMs process input text sequences (tokens) and generate output text sequences based on the statistical regularities they have learned from massive training data (represented by the parameters of the neural network, equivalent to the "rulebook" in the Chinese Room). From the outside, LLMs can generate fluent and relevant answers on various topics, as if they "understand" the questions and the content they are discussing. However, critics argue that the way LLMs operate is very similar to the person in the Chinese Room: they are both performing complex symbol (token) manipulation but lack an understanding of the real meaning behind these symbols. When LLMs predict the next token, they do so based on statistical probability, not on a grasp of concepts or a cognitive understanding of the world. Therefore, the "stochastic parrot" label aligns with the core argument of the Chinese Room—that successful symbol manipulation does not equal true understanding.
Of course, the Chinese Room Argument also faces many rebuttals, the most famous of which is the "Systems Reply." This reply argues that although the person in the room does not understand Chinese, the entire system (including the person, the rulebook, the symbols, and the room itself) as a whole does understand Chinese. Applying this to AI, even if an AI's individual algorithms or components do not "understand," the entire AI system, perhaps in its interaction with its environment, data, or users, may exhibit some form of understanding or intelligence. This view sees understanding as a property that can emerge in a complex system, rather than being confined to a single conscious subject. This provides a perspective for thinking about AI's cognitive abilities that is different from the individualistic view of the mind and is related to theories like distributed cognition or the extended mind.
Despite the controversy, the Chinese Room Argument continues to remind us that when evaluating the cognitive abilities of AI, we must be wary of equating its external behavioral performance (such as passing the Turing test or generating fluent text) with internal understanding and consciousness. For LLMs, even if they can perform increasingly complex language tasks, whether they truly "know" or "understand" what they are saying remains an unresolved and profoundly epistemological question.
IV. Navigating the Cognitive Challenges of Advanced AI
A. The "Black Box" Problem: Opacity, Trust, and Epistemic Responsibility
Modern artificial intelligence, especially connectionist models based on deep learning, often has internal operating mechanisms that are so complex that even their designers cannot fully understand their decision-making processes. This phenomenon is known as the "black box problem" of AI. When an AI system (such as ChatGPT, Gemini, etc.) makes a prediction, classification, or generates content, we may know its input and output, but the "reasoning" path of how it got from the input to the output is opaque, or its decision logic is "inexplicable."
This opacity poses a direct challenge to the justification of knowledge. If justification requires us to be able to understand or review the reasons or process by which a belief is true, then the output of a black box AI makes such a review extremely difficult. We may observe that an AI performs excellently on certain tasks, and its output is often "correct," but this does not automatically equate to its output being "justified knowledge." Because we have no way of knowing how this "correct" result was produced—was it based on reliable "reasoning," an accidental statistical coincidence, or did it learn some unforeseen, or even harmful, correlation in the data? This cognitive uncertainty makes trusting AI decisions risky, especially in high-stakes fields like medical diagnosis, financial risk control, and judicial sentencing.
The black box problem also complicates the attribution of cognitive responsibility. When an opaque AI system makes a wrong decision, causing loss or injustice, who should be held responsible? The AI itself (which is difficult given that AI currently lacks legal personality and true autonomous consciousness)? Its designers (but they may not be able to fully foresee or control all of the AI's behaviors)? Or the deployers or users (but they may lack the ability to understand and intervene in the AI's internal mechanisms)? This blurring of responsibility further weakens the basis for treating AI output as a reliable source of knowledge.
Therefore, the opacity of AI forces us to rethink the meaning of "justification." Should we adhere to the traditional view of justification, which requires access to reasons, and thus be skeptical of the knowledge claims of black box AI? Or should we turn to a more results- and reliability-oriented view of justification, where if an AI system consistently demonstrates high accuracy in practice, we can (to some extent) "trust" its output, even if we don't understand its process? Or should we vigorously develop technologies that can "open the black box," such as Explainable AI (XAI), to bridge this cognitive gap? These are key questions that epistemology must face in the age of AI. Some researchers suggest that hybrid methods like Neuro-Symbolic AI may offer a way to combine the learning ability of neural networks with the explainability of symbolic systems, thereby mitigating the black box problem.
B. Explainable AI (XAI): The Quest for Justification and Understanding in AI Decisions
In response to the cognitive and ethical challenges posed by the increasing complexity and opacity of AI systems (especially deep learning models), "Explainable AI" (XAI) has emerged. The goal of XAI is to develop AI systems that can provide explanations for their decision-making processes and output results, thereby enhancing human understanding, trust, and control over AI. XAI aims to reveal how and why an AI model makes a specific prediction or decision, for example, by using techniques (such as LIME, SHAP, etc.) to identify the key features that influence a decision, display decision rules, or provide simplified model approximations.
From an epistemological perspective, the pursuit of XAI is closely related to the requirement for "justification" in epistemology. If an AI system can not only give an answer but also provide a reasonable explanation for its answer, then this explanation itself can be seen as a form of "justification," thereby enhancing the credibility of its output as "knowledge." XAI attempts to bridge the gap between AI's computational processes and human cognitive understanding, providing users with a basis for evaluating the reliability of AI output. Early XAI research focused mainly on developing new explanation methods, but subsequent research has also begun to focus on whether these methods can effectively meet the needs and expectations of different stakeholders, and how to handle the impact of stakeholders' own biases on XAI-assisted decision-making.
However, XAI also faces its own epistemological difficulties. First, many XAI techniques provide "post-hoc explanations" of model behavior. These explanations themselves may be simplifications or approximations of complex model behavior, rather than a complete reproduction of the model's actual "thinking" process. This raises the question: is an "explanation" of a decision equivalent to the AI "possessing" an internal, accessible justification? Such an explanation is more like a "cognitive scaffold" built for human users to help us understand and trust AI, but it does not necessarily mean that the AI itself is "reasoning" in an explainable way.
Second, some scholars propose that the goal of XAI should not just be to provide "explanations," but to pursue a higher level of "understanding." In epistemology, "understanding" is often considered a deeper and more valuable cognitive achievement than simply having "knowledge" or "explanations." Understanding is not only about "what" and "how," but also about "why," involving a grasp of the relationships between things and insight into their meaning. If XAI can help humans (and even AI itself) to achieve an "understanding" of the deep logic and principles behind AI decisions, it would be a major epistemological advance. But this also places higher demands on XAI, possibly requiring the integration of explainability and internal logic from the very beginning of model design, rather than relying solely on post-hoc explanation tools. Furthermore, understanding itself does not always require all information to be absolutely true (i.e., it need not be "factive"), which differs from the "truth" condition in the traditional definition of knowledge and provides a new dimension for evaluating the cognitive state of AI.
C. Algorithmic Bias as Epistemological Failure: The Consequences of Inconclusive and Inscrutable Evidence
Algorithmic bias refers to the phenomenon where an AI system produces unfair or discriminatory outcomes for specific groups due to systematic flaws in its training data, algorithm design, or deployment method. Algorithmic bias is usually discussed from an ethical perspective, but viewing it as an "epistemological failure" can more profoundly reveal its essence.
From an epistemological perspective, bias itself is a "cognitive defect," a flawed cognition of others or things. When an AI system learns from data containing historical biases (for example, data reflecting stereotypes or unfair treatment of specific racial, gender, or socioeconomic groups in society), or when the algorithm's design itself embeds inappropriate assumptions, the AI is actually "learning" and "reasoning" based on "inconclusive evidence" or "inscrutable evidence." This flawed "evidentiary" basis inevitably leads the AI system to form distorted "representations" of the world and unreliable "knowledge."
Therefore, algorithmic bias is not just an ethical problem that leads to unfair outcomes, but a systematic epistemological failure. The output produced by an AI system in such cases, even if formally "self-consistent" or "efficient," cannot represent true or reasonably justified knowledge about the world, especially concerning the groups affected by the bias, because it is built on a flawed cognitive foundation. When human users tend to trust the decisions of machines, algorithmic bias can further entrench and amplify these erroneous cognitions.
Furthermore, algorithmic bias can lead to "epistemic injustice." If an AI system, due to bias, systematically devalues, ignores, or misrepresents the experiences, abilities, or characteristics of specific groups, it is not only producing false knowledge but also damaging the status and dignity of these groups as cognitive subjects, depriving them of the right to be known and evaluated fairly. For example, in recruitment, credit approval, or medical diagnosis, if an AI consistently gives lower assessments to certain groups, it is, in effect, institutionally entrenching skepticism or denial of the cognitive abilities of these groups.
Therefore, addressing the problem of algorithmic bias requires not only ethical norms and technical corrections but also reflection from an epistemological level: how can we ensure that the data and algorithms on which AI systems rely can provide a fair and comprehensive cognitive basis? How can we identify and correct the systematic "blind spots" and "distortions" that may occur in the AI's cognitive process? This requires us to implement principles of cognitive diligence and epistemic justice throughout the entire process of AI design, development, and deployment.
D. AI "Hallucinations": Misinformation, Truth, and the Verifiability of AI-Generated Content
AI "hallucination" refers to the phenomenon where an AI system (especially generative AI, such as large language models) produces content that is seemingly coherent and persuasive but is actually false, fabricated, or inconsistent with the input. These "hallucinated" contents may have no factual basis at all, or they may mix real information with fictional elements. For example, when asked to provide a literature review, an LLM might "invent" non-existent paper titles, authors, or even citations.
The mechanisms behind AI hallucinations are complex and varied. They may stem from the probabilistic nature of the models themselves (they are designed to generate the statistically most likely sequence, not absolutely true content), errors, contradictions, or outdated information in the training data, the model's ignorance of its own knowledge boundaries (i.e., it still provides answers in uncertain areas), or certain inherent flaws in the language models themselves. Sometimes, AI even exhibits "source amnesia," meaning it cannot trace the true source of the content it generates, or a "hallucination snowball effect," where once an error is produced, it continues to generate more related errors to maintain coherence.
From an epistemological perspective, AI hallucinations pose a direct and serious threat to the "truth" condition of knowledge. If a system can so confidently and fluently generate false information, its reliability as a source of knowledge is greatly diminished. The unique feature of AI hallucinations is that they are often false information produced "without human deceptive intent." This is different from disinformation, which is deliberately created or spread by humans, or misinformation, which is unintentionally spread. The "deceptiveness" of AI hallucinations lies in the highly anthropomorphic and seemingly plausible form of their output, which makes it easy for users (especially unwary ones) to believe them.
The phenomenon of AI hallucinations highlights the extreme importance of verifying information in the age of artificial intelligence. Users can no longer simply treat AI-generated content as authoritative or factual but must cultivate the habit and ability to critically examine and independently verify such content. This is not only a new requirement for individual cognitive abilities but also a challenge to the education system, research norms, and the entire information ecosystem. We need to develop new tools, methods, and literacies to identify and respond to potential misinformation generated by AI. Some scholars even argue that the term "hallucination" may not be entirely appropriate, as it implies that the AI "sees something that isn't there," whereas it is more like the AI is "fabricating" content. Regardless of the name, this phenomenon forces us to pay more attention to the cognitive basis of AI output and how we can ensure that the "knowledge" we rely on is true and reliable in an era of increasing dependence on AI for information.
V. The Reconstruction of Knowledge in the AI Era: Future Trajectories
A. Neuro-Symbolic AI: Toward a More Integrated and Robust Cognitive Architecture?
Against the backdrop of Symbolic AI and Connectionist AI each demonstrating unique advantages and limitations, Neuro-Symbolic AI, as a hybrid approach that combines the strengths of both, is receiving increasing attention. Its core idea is to combine the powerful learning and pattern recognition capabilities of neural networks with the explicit knowledge representation and logical reasoning capabilities of symbolic systems, in the hope of building more reliable, explainable, and cognitively comprehensive AI systems.
As previously discussed (see Sections I.D and II.B), Connectionist AI excels at processing large-scale, unstructured data and learning complex patterns, but its "black box" nature and lack of robust symbolic reasoning capabilities are its main shortcomings. Symbolic AI, on the other hand, is adept at handling structured knowledge, performing explicit logical reasoning, and providing explainable results, but its ability to handle the ambiguity of the real world and learn from experience is limited. Neuro-Symbolic AI attempts to integrate these two paradigms in various ways, such as using neural networks to process perceptual input or learn sub-symbolic representations from data, and then passing these representations to a symbolic reasoning engine for higher-level cognitive tasks (like planning or problem-solving); or conversely, using symbolic knowledge to guide the learning process of a neural network to improve its learning efficiency, generalization ability, or explainability.
From an epistemological perspective, the exploration of Neuro-Symbolic AI is significant. If Connectionist AI in some way echoes the empiricist emphasis on experiential learning, and Symbolic AI reflects the rationalist emphasis on logic and a priori knowledge, then Neuro-Symbolic AI can be seen as an attempt at the computational level to achieve the integration and complementarity of these two epistemological paths. This integration holds the promise of overcoming the limitations of a single paradigm, for example, by introducing symbolic components to enhance the transparency and explainability of connectionist models, thereby solving parts of the "black box" problem; or by using the adaptive learning capabilities of neural networks to compensate for the shortcomings of symbolic systems in knowledge acquisition and dealing with novel situations.
Furthermore, neuro-symbolic systems may provide a promising path towards achieving "true AI reasoning" that is closer to human cognition. Many current LLMs are criticized for relying mainly on statistical pattern matching rather than deep logical understanding. By explicitly integrating symbolic reasoning modules, neuro-symbolic architectures could potentially enable AI systems to perform more complex, multi-step, and verifiable reasoning processes that go beyond mere statistical inference, moving towards rule-based deduction, induction, and even abduction. This would help AI systems provide a more solid "justification" for their conclusions and present their "thinking" process in a way that is more easily understood by humans. For example, a system could use a neural network to extract features and concepts from raw data, and then use symbolic logic to reason about the relationships between these concepts to draw conclusions and explain its reasoning chain. Such an architecture not only promises to improve AI performance but also has the potential to make it cognitively more robust and trustworthy.
B. Human-AI Cognitive Augmentation: The "Extended Mind" and "Cognitive Offloading"
As AI becomes more deeply integrated into human life and work, a new perspective is emerging to examine the relationship between AI and human cognition: no longer viewing AI merely as an independent entity trying to simulate or surpass human intelligence, but as an extension and enhancement of human cognitive abilities. This perspective draws on concepts such as the "Extended Mind Theory" proposed by Andy Clark and David Chalmers, and the related concept of "cognitive offloading."
The Extended Mind Theory argues that human cognitive processes are not entirely confined within the brain but can extend into the external world, using tools and resources in the environment to assist or constitute cognitive activities. For example, we use notebooks to record information to aid memory and calculators for complex computations; these external tools functionally become part of our cognitive system. AI, with its powerful capabilities in information processing, pattern recognition, and decision support, is increasingly becoming a powerful form of this "cognitive tool." In fields like medicine, scientific research, and education, AI and technologies like mixed reality (MR) can serve as cognitive extensions for clinicians, researchers, or students, helping them process complex information, make better decisions, or train efficiently in simulated environments.
Relatedly, "cognitive offloading" refers to the act of transferring certain mental functions (such as memory, calculation, information retrieval, decision-making, etc.) to external resources (like digital devices or AI assistants). This offloading can increase efficiency and reduce cognitive load, allowing humans to focus on higher-level thinking and creativity. For example, students can use AI tools to assist with literature review and writing, and scientists can use AI to analyze massive datasets and generate hypotheses.
However, this picture of human-AI cognitive augmentation also brings new epistemological challenges and potential risks. First, it blurs the boundaries of the cognitive subject. If "knowledge" and "cognitive processes" are distributed across a hybrid system composed of humans and AI, then questions like "Who is the knower?", "To whom does the knowledge belong?", and "How is cognitive responsibility distributed?" become more complex. Traditional individualistic epistemology may be inadequate to fully explain this distributed, collaborative cognitive phenomenon.
Second, while "cognitive offloading" brings convenience, it may also lead to the degradation or "deskilling" of human cognitive abilities. If individuals become overly reliant on AI to complete cognitive tasks, it could weaken their independent critical thinking skills, memory, analytical abilities, and problem-solving skills. Research suggests that frequent use of AI tools may be negatively correlated with critical thinking abilities, especially among younger users. This cognitive dependency could not only make individuals vulnerable when AI is unavailable or makes mistakes but could also, in the long run, affect the overall cognitive health and intellectual development of humanity. Therefore, how to balance the use of AI to enhance cognitive abilities while avoiding cognitive over-reliance and the loss of core skills has become a pressing issue. This requires us not only to develop smarter AI but also to cultivate humans with high levels of AI literacy and cognitive autonomy.
C. The Epistemological Shift in AI-Driven Scientific Discovery and Education
The rapid development of artificial intelligence is profoundly changing the core domains of knowledge production and dissemination—scientific research and educational practices—and is triggering potential shifts in their epistemological foundations.
In the field of scientific discovery, AI is evolving from an auxiliary tool to a potential "cognitive engine." "Agentic AI" systems, equipped with reasoning, planning, and autonomous decision-making capabilities, are changing the way scientists conduct literature reviews, generate hypotheses, design experiments, and analyze results. For example, Google's "AI co-scientist" system is designed to mimic the reasoning process of the scientific method, collaborating with human scientists to discover new knowledge and formulate original research hypotheses and proposals. By automating traditionally labor-intensive research steps, AI promises to accelerate the pace of scientific discovery, reduce costs, and make cutting-edge research tools more accessible. However, this AI-driven model of scientific discovery also brings new epistemological challenges. If AI generates scientific hypotheses or analyzes data in an opaque "black box" manner, how can human scientists trust, verify, and understand the "knowledge" contributed by AI? How can the core values of scientific research—such as reproducibility, peer review, and a clear understanding of methodology—be guaranteed in a research process with deep AI involvement? This may require the development of new scientific methodologies and validation standards to adapt to the new paradigm of human-machine collaboration in research, ensuring that AI's contributions are not only efficient but also epistemologically sound and trustworthy.
In the field of education, technologies like Generative AI (GenAI) are bringing profound changes to teaching content, personalized learning, intelligent tutoring, assessment methods, and even teacher professional development. AI has the potential to help students overcome learning obstacles, providing customized learning paths and instant feedback. However, the application of AI in education also raises fundamental epistemological and ethical concerns about teaching goals, student agency, the cultivation of critical thinking, and educational equity. Some scholars warn that an overemphasis on using AI to meet labor market demands could lead to superficial learning rather than fostering students' deep understanding and critical thinking. AI might inadvertently reinforce existing educational inequalities or shape teaching practices in unintended ways. Therefore, some researchers call for the establishment of a "new onto-epistemological basis" for the interaction between AI and humans in education. This means that the focus of education may need to shift from the traditional knowledge transmission model to cultivating students' ability to learn collaboratively with AI, critically evaluate AI-generated information, and conduct innovative inquiries with AI assistance. The subjectivity and agency of students need to be protected and enhanced, and education should be committed to cultivating individuals who can engage in meaningful learning and adhere to ethical norms in the AI era. This requires educators themselves to improve their AI literacy and to critically reflect on the role of AI in education, ensuring that technology serves truly human-centered educational goals.
D. Concluding Reflections: The Evolving Landscape of Knowledge and Inquiry
The rise of artificial intelligence is challenging and reshaping our understanding of knowledge, the cognitive subject, and the process of inquiry in unprecedented ways. From clarifying the conceptual distinction between different notions of epistemology, to examining AI's challenge to the classic definition of knowledge (JTB), to analyzing the epistemological presuppositions inherent in different AI paradigms (Symbolic, Connectionist, Neuro-Symbolic) and the cognitive dilemmas they bring (such as the black box problem, algorithmic bias, and AI hallucinations), this report has revealed the complex and profound interactive relationship between AI and epistemology.
AI systems, especially advanced AI like large language models, exhibit significant differences from human cognition in the core elements of traditional epistemology such as "belief," "truth," and "justification," making the question of whether "AI can possess knowledge" an unresolved philosophical puzzle. As a provider of "testimony," AI's lack of intentionality and accountability challenges the knowledge transmission model based on trust. The debate over whether LLMs are "stochastic parrots" or possess a nascent form of "understanding" touches upon the fundamentals of meaning theory.
In the face of these challenges, the exploration of Explainable AI (XAI), the revelation of the epistemological roots of algorithmic bias, and the emphasis on verification mechanisms for AI hallucinations all reflect humanity's efforts to cognitively manage and regulate this powerful technology. At the same time, the development of hybrid paradigms like Neuro-Symbolic AI heralds the possibility of building more robust AI cognitive architectures that combine learning capabilities with reasoning transparency.
Furthermore, the development of AI prompts us to rethink the boundaries of the cognitive subject. Concepts like the "extended mind" and "cognitive offloading" suggest that the "knower" of the future may no longer be an isolated individual but a deeply integrated human-machine cognitive system. This brings both the immense potential for cognitive enhancement and the risks of cognitive dependency and skill degradation. In key knowledge domains like scientific discovery and education, the integration of AI is giving rise to new research methods and teaching paradigms, while also requiring us to be vigilant about its potential negative impacts, ensuring that human cognitive autonomy and critical thinking are cherished and cultivated in the wave of technology.
Ultimately, the relationship between artificial intelligence and epistemology is not a one-way street of scrutiny and being scrutinized, but a dynamic co-evolution. The development of AI continuously poses new questions and research objects for epistemology, forcing philosophy to reflect on and update its theoretical frameworks; in turn, the insights of epistemology can provide directional guidance and ethical navigation for the healthy development of AI. In this era of increasing AI penetration, cultivating human cognitive virtues—such as critical thinking, the will to seek truth, intellectual humility, and open-mindedness—is more important than ever. Only in this way can we, in our interaction with AI, truly achieve the growth of knowledge and the enhancement of wisdom, and jointly shape a future that is epistemologically more sound and responsible.