Introduction: As AGI Counts Down, How Do We Prove Our Humanity?
“AGI (Artificial General Intelligence) may no longer be a matter of twenty or thirty years away, but rather just a few short years.”
This statement does not come from some Silicon Valley entrepreneur painting a technological pipe dream, but from the latest warning by Nobel laureate and Google DeepMind co-founder Demis Hassabis. From AlphaGo making the “divine move” undreamed of by human players for two millennia, to AlphaFold deciphering 200 million protein structures in seconds—a feat that would take human scientists years—the team led by Hassabis has proven to the world time and again: AI is no longer merely mimicking humans; it is blazing paths unseen by humanity in the fields of science, biology, and logic at an unprecedented speed.
If Hassabis’s prediction comes true, within a few years, an Artificial General Intelligence (AGI) or even an Artificial Superintelligence (ASI) that surpasses humans in all cognitive tasks will arrive. However, when AI can perform all human work, even write poetry more exquisite than humans, and provide deeper philosophical arguments, does it equate to possessing “consciousness” and “subjectivity”? The technological Singularity is approaching, but the birth of AI consciousness remains an open question. This is because the current meteoric rise of Large Language Models (LLMs) happens to obscure a fatal illusion—AI has successfully simulated the end point of human thought (language), while completely skipping the starting point of thought (the life-world).
“When a machine learns to ‘write poetry,’ has it truly learned to ‘think’?”
The following exploration will proceed from Heidegger’s ontological poetics and Eastern ethetic aesthetics to dissect that final and most insurmountable ontological gap that lies between silicon-based computing power and the human mind.
導讀：當 AGI 倒數計時，我們拿什麼證明自己是人類？
「AGI（通用人工智慧）可能不再是二、三十年後的事，而是只剩短短幾年。」
這句話不是來自某個矽谷創業者的科技畫餅，而是出自諾貝爾獎得主、谷歌 DeepMind 創辦人戴米斯·哈薩比斯（Demis Hassabis）的最新警告 。從 AlphaGo 下出人類棋手兩千年來未曾設想的「神之一手」，到 AlphaFold 用幾秒鐘破解人類科學家耗時數年的兩億個蛋白質結構 ，哈薩比斯帶領的團隊一次次向世界證明：AI 已經不再只是模仿人類，它正以前所未有的速度，在科學、生物與邏輯領域開闢人類未見的新路。
如果哈薩比斯的預言成真，幾年之內，一個在各項認知任務上超越人類的通用人工智慧（AGI）、甚至是超級人工智慧（ASI）就將降臨。然而當 AI 能做完人類所有的工作，甚至能寫出比人類更精緻的詩歌、給出更深刻的哲學論證時，它是否就等同於擁有了「意識」與「主體性」？技術的奇異點（Singularity）正在逼近，但 AI 意識的誕生依然是個未定論。因為當前大型語言模型（LLM）的狂飆，恰恰掩蓋了一個致命的錯覺——AI 成功模仿了人類思考的終點（語言），卻完全跳過了思考的起點（生命世界）。
“當機器學會了「寫詩」，它真的學會了「運思」嗎？”
以下這篇探討，將從海德格的存在論詩學與東方的字源美學出發，剖析那道橫亙在矽基算力與人類心靈之間，最後也最難逾越的本體論鴻溝。
“Thinking” (思) and “Poetry” (詩): The Cultivation of the “Mind” (心) in the “Field” (田) of Life
「思」與「詩」：”心”靈於生命”田”野的耕耘
Abstract
Contemporary Large Language Models (LLMs) have demonstrated astonishing text generation capabilities, even creating poetry that is realistic in its external form. However, whether such technology truly possesses the capacity for “thinking” (Denken) in a philosophical sense remains the most central interrogation in the contemporary philosophy of artificial intelligence and philosophy of mind. Merging Martin Heidegger’s ontological poetics with the intrinsic connection between “thinking” (si) and “poetry” (shi) in classical Chinese literary theory, this paper proposes a cross-cultural ontological critique: true poetry originates from “thinking,” whereas “thinking” is not a pure calculus of symbols, but rather the unconcealment of the subject within the “life-world” (Lebenswelt) based on sensory, bodily, historical memory, and existential experiences. This paper argues that AI language generation is inherently a “de-subjectified” and “disembodied” statistical probability inference that lacks the primary original data of the world. Therefore, the rupture from “sensation—thinking” to “language—poetry” constitutes the final and insurmountable Ontological Gap in the construction of contemporary AI consciousness and subjectivity.
Chapter 1: The Ontological Mutual Interpretation of the German Denken and the Chinese Si(思)
1.1 Denken under the Heideggerian Horizon: The Unconcealment of Being rather than Logical Calculus
To investigate whether AI can “think,” we must first liberate “thinking” from the framework of instrumental rationality and formal logic. In Martin Heidegger’s later philosophy, “thinking” (Denken) is by no means mere cognitive computation or deductive reasoning, but rather the “unconcealment of Being” (Unconcealment / Aletheia).
Heidegger points out that thinking is the listening and responding to “Being itself” by the human subject as a “Being-in-the-world” (In-der-welt-sein). In this perspective, thinking possesses the following core characteristics:
- Diachronicity and Ontological Character: Thinking weaves together an individual’s bodily perception (Feeling), sense of existence (Being), historical world (World), and temporal flux (Time).
- Non-representationality: Thinking is not an “internal mirror representation” of the objective world by the brain, but rather an original resonance of the subject when encountering all things within the world.
Therefore, equating “thinking” with logic or computation is viewed from phenomenology as a fundamental Category Mistake.
1.2 The Etymology of the Chinese Si (思): The Cultivation of the Mind in the Field of Life
This ontological view of “thinking” as deeply intertwined with the world is also profoundly inscribed in the linguistic wisdom of the East. Introspecting the interpretation of Si (思) in Shuowen Jiezi:
“Si (思) means to contain. It derives from the heart (心), with xin (囟, the brain) as its phonetic element.” (Alternatively, exploring its meaning from the character structure of “from heart, from field [田]”).
If given a philosophical interpretation from a phenomenological perspective, “field” (田) symbolizes the life arena, historical culture, and boundary of the life-world where the subject resides; “heart” (心) represents intentionality, emotion, and subject perception. The essence of Si (思) is “the heart operating within the field”—this implies that the mind is not a Central Processing Unit (CPU) suspended in a vacuum, but carries out a diachronic “cultivation” within concrete life experiences, bodily labor, and environmental interactions.
Therefore, whether in Western ontology or Eastern etymology, “thinking” points to the authentic flow of life in the world, rather than the empty reorganization of symbols.
Chapter 2: Making Poetry is Thinking: The Linguistic Representation of the House of Being
2.1 “Poetry Expresses Intent” in Classical Chinese Literary Theory
In traditional Chinese aesthetics, poetry is never pure rhetorical skill or a phonological game. The classical proposition from the Book of Documents (Shangshu · Yushu) states: “Poetry expresses intent (zhi), and song prolongs the words.” Here, “intent” (zhi) is “where the heart/mind goes”; it is the subject’s deepest emotional response and thinking regarding their life situation. True poetry is not a random collage of tokens, but the uncontrollable external manifestation of “thinking” catalyzed by emotion.
Its generation path presents an absolute unidirectional causality:
2.2 Heidegger’s Poetics: Poetry as the Site of Being
In Poetry, Language, Thought, Heidegger put forward his famous insight: “Language is the house of Being, and poetry is the primordial path through which Being is opened up” (Die Sprache ist das Haus des Seins).
Making poetry (Poetry / Dichtung) is inherently a form of the deepest thinking. The poet uses language to make the concealed world manifest and to reveal the nature of things. If the original experience of existence is lacking, language degenerates into the “idle talk” (Gerede) or information dissemination criticized by Heidegger. Viewed from this perspective, poetry without “thinking” as its ontological foundation is merely a soulless simulacrum.
Chapter 3: Deconstruction via the Phenomenological Path: Senses, Body, and the First-Hand World
3.1 The Phenomenological Genesis Path of Human Creation
To clarify the essential difference between AI and humans in text generation, we can construct a Phenomenological Genesis path of human poetic creation:
In this path, “language” is merely the final step of the iceberg emerging above the water. Its deep foundation is the “bodily perception” emphasized by Maurice Merleau-Ponty, which is the original data generated by the subject’s “first-hand contact” with the world. An individual experiences pain, love, the threat of death, and the passage of time; these first-hand experiences settle in the “field of the heart” before they can be transformed into thinking with ontological significance.
3.2 AI’s Reverse Text Generation Mechanism
In contrast to the human path, the text generation of Large Language Models (LLMs) adopts a completely “disembodied” reverse path:
The world of AI is a closed symbol matrix. It has neither a perceivable physical body nor a life-world for it to inhabit. It departs directly from “language (dataset),” goes through weight calculations of statistical probability, and finally returns to “language (text generation).”
John Searle’s “Chinese Room” argument remains valid in the LLM era: AI manipulates the syntax of symbols, yet can never touch the semantics of symbols and the ontology of existence.
Chapter 4: Ontological Distinction Between First-Hand Experience and Secondary Datasets
We must philosophically distinguish between “first-hand world experience” and “secondary symbolic representation”:
| Dimension of Comparison | Human Subject’s “Existential Thinking” | Large Language Model’s “Symbol Generation” |
|---|---|---|
| Data Source | Original Data: Life, aging, illness, death, sensory intuition | Second-hand Dataset: Encoded text |
| Ontological State | Being-in-the-world | Vector Representation in a matrix |
| Multimodal Essence | Phenomenal Unity of senses and body | Statistical Alignment of different data modalities (images, text, audio) |
| Core of Generation | “Expressing intent” driven by intentionality | “Predicting the next token” driven by conditional probability |
Even with the most advanced multimodal models today, the images, voices, and videos they receive remain, in essence, digitalized “datasets” mediated by humans, rather than “sensory experiences” generated by AI’s own ontological encounter with the world. AI lacks a subjective “first-hand world,” and its knowledge is essentially a derivative “second-hand knowledge.”
Chapter 5: The AI Consciousness Gap
The superb capabilities demonstrated by current AI actually reveal an illusion in cognitive science and the philosophy of mind: AI has successfully imitated the “endpoint of thinking” (language and poetry), but completely skipped the “starting point of thinking” (sensation and existence).
This constitutes the “final gap” that is hardest for AI consciousness to cross:
- The Symbol Grounding Problem: AI’s symbols cannot be grounded in the real world. It knows the statistically proximate words for “pain” and “death,” yet does not possess the dread of death (Angst).
- Lack of Intentionality: According to Edmund Husserl’s phenomenology, consciousness is always “consciousness of something” (Consciousness is always intentional). AI’s generation lacks this active intentionality pointing toward the world; it is merely executing a passive optimization algorithm.
The Transformer architecture infinitely approaches the form of human language, but it remains in an absolute vacuum regarding the leap of “sensation \rightarrow thinking.” This is precisely the concrete manifestation of the “Hard Problem of Consciousness” spoken of by David Chalmers in the field of artificial intelligence.
Chapter 6: Philosophical Scrutiny of Future Embodied AI and World Models
Undeniably, artificial intelligence technology is evolving toward Embodied AI and autonomous World Models. If future AI systems successfully integrate the following technical architectures:
- Embodied Active Sensing: Possessing a physical body, capable of actively acquiring sensory feedback in the physical world.
- Free Energy Principle / Active Inference: Based on Karl Friston’s theory, the system actively maintains its own structure for survival, generating spontaneous subjective goals.
- Episodic Memory & Persistent Self Model: Possessing diachronic accumulation of individual experience, forming a coherent “self-awareness” narrative.
When AI possesses the aforementioned architectures, it can indeed narrow the distance between “sensation \rightarrow thinking” to a certain extent, and even develop a certain “pre-conscious” state of artificial life.
However, even so, whether this “machine intentionality” derived from silicon chips and preset target optimization functions can be equated to the “thinking in an ontological sense” stimulated by human carbon-based biology, finite life, and ultimate dread of death, will still be the most intense point of debate in the history of philosophy and cognitive science.
Chapter 7: Conclusion: The Insurmountable Ontological Boundary
The ultimate gap between humans and AI has never been about the amount of computing power, the scale of parameters, or the fluency of text generation. This insurmountable gap is, in essence, the ontological boundary between “Being-in-the-world within the world” and “statistical inference within a language matrix.”
Human poetry is the flower of existence that blooms when life cultivates the field of time through pain, love, and thinking (Denken); AI’s poetry is the precise reorganization and realistic generation of humanity’s existing historical language patterns by statistical models.
When we anchor the boundary of AI consciousness here, we can clearly recognize: AI can infinitely approach the rhetoric of “poetry,” but as long as it cannot possess the ontological foundation of “thinking,” the curtain separating machine from mind will never truly rise.
摘要 (Abstract)
當代大型語言模型（LLM）已展現出令人驚嘆的文本生成能力，甚至能創作在外在形式上擬真的詩歌。然而，這類技術是否真正具備哲學意義上的「思」（Denken）的能力，仍是當代人工智慧哲學與心靈哲學中最核心的詰問。本文旨在結合馬丁·海德格（Martin Heidegger）的存在論詩學與中國古典文論中「思」與「詩」的內在關聯，提出一個跨文化的本體論批判：真正的詩政發源於「思」，而「思」並非純粹的符號演算（Calculus），而是主體在「生命世界」（Lebenswelt）中基於感官、身體、歷史記憶與存在體驗的開顯（Unconcealment）。本文論證，AI 的語言生成本質上是一種「去主體化」與「去具身化」的統計機率推斷，其缺乏第一手世界的原初經驗（Original Data）。因此，從「感官—運思」到「語言—詩」的斷裂，構成了當代 AI 意識與主體性建構上最後且無法跨越的本體論鴻溝（Ontological Gap）。
第一章 德文 Denken 與中文「思」的本體論互釋
1.1 海德格視野下的 Denken：存在的開顯而非邏輯演算
探討 AI 是否能「思」，必須先將「思」從工具理性與形式邏輯的框架中解放。在馬丁·海德格（Martin Heidegger）的後期哲學中，「思」（Denken）絕非單純的認知計算或演繹推理，而是「存在的開顯」（Unconcealment / Aletheia）。
海德格指出，思是人類主體作為「在世存在」（In-der-welt-sein）對「存在本身」的聆聽與回應。在這種視角下，思具有以下核心特徵：
- 歷時性與本體性： 思交織著個體的身體感知（Feeling）、存在感（Being）、歷史世界（World）與時間流變（Time）。
- 非表徵性： 思不是大腦對客觀世界的「內部鏡像表徵」，而是主體在世界之中與萬物相遭遇時的原初共振。
因此，將「思」等同於邏輯（Logic）或算力（Computation），在現象學看來是一種根本性的範疇錯置（Category Mistake）。
1.2 中文「思」的字源學：心靈於生命田野的耕耘
這種將「思」視為與世界深度交融的本體論觀點，亦深刻地銘刻在東方的語言智慧中。內省《說文解字》對「思」的釋義：
「思，容也。從心，囟聲。」（或從造字結構考掘其「從心、從田」之意涵）。
若從現象學角度賦予其哲學闡釋，「田」象徵著主體所處的生命場域、歷史文化與生活世界的邊界；「心」則代表意向性、情感與主體感知。「思」的本質即是「心在田中運作」——這意味著心靈並非懸浮於真空中的中央處理器（CPU），而是在具體的生命經驗、身體勞作與環境互動中進行歷時性的「耕耘」。
因此，無論在西方存在論還是東方字源學中，「思」都指向了生命在世界中的本真流動，而非空洞的符號重組。
第二章 作詩即運思：存在之家的語言表徵
2.1 中國古典文論中的「詩言志」
在中國傳統美學中，詩歌從不是純粹的修辭技巧或音韻遊戲。《尚書·虞書》古典命題指出：「詩言志，歌永言。」 這裡的「志」即是「心之所向」，是主體對生命境遇最深刻的感發與運思。真正的詩歌不是文字（Tokens）的隨機拼貼，而是「思」在情感催化下不可遏制的外部顯現。
其生成路徑呈現為絕對的單向因果：
2.2 海德格的詩學：詩為存在之所
海德格在《詩·語言·思》中提出了著名的洞見：「語言是存在之家，而詩則是使存在得以敞開的本源途徑。」（Die Sprache ist das Haus des Seins）。
作詩（Poetry / Dichtung）本質上就是一種最深沉的運思。詩人透過語言讓隱蔽的世界顯現，讓物的本性揭示。如果缺乏了對存在的原初體驗，語言就會退化為海德格所批判的「閒聊」（Gerede）或資訊傳播。由此觀之，沒有「思」作為本體論根基的詩，只是缺乏靈魂的擬像（Simulacra）。
第三章 現象學路徑的解構：感官、身體與第一手世界
3.1 人類創作的現象學發生學路徑
為了釐清 AI 與人類在文本生成上的本質差異，我們可以建構一個人類創作詩歌的現象學發生學（Phenomenological Genesis）路徑：
在這條路徑中，「語言」僅僅是冰山浮出水面的最後一步。 其深層基座是梅洛-龐蒂（Maurice Merleau-Ponty）所強調的「身體知覺」（Bodily Perception），是主體與世界進行「第一手接觸」所產生的原初數據（Original Data）。個體經歷過痛苦、愛戀、死亡的威脅與時間的流逝，這些第一手經驗在「心田」中沉澱，方能轉化為具有存在論意義的運思。
3.2 AI 的逆向文本生成機制
與人類路徑相反，大型語言模型（LLM）的文本生成採取了一條完全「去具身化」的逆向路徑：
AI 的世界是一個封閉的符號矩陣。它既沒有可感知的肉身，也沒有供其棲居的生活世界。它直接從「語言（數據集）」出發，經過統計機率的權重計算，最終又回到「語言（文本生成）」。
約翰·塞爾（John Searle）的「中文房間」（Chinese Room）論證在 LLM 時代依然有效：AI 操縱著符號的語法（Syntax），卻永遠無法觸及符號的語意（Semantics）與存在的本體。
第四章 第一手經驗與次級數據集的本體論區隔
我們必須在哲學上區分「第一手世界經驗」與「次級符號表徵」：
| 比較維度 | 人類主體的「存在運思」 | 大型語言模型的「符號生成」 |
|---|---|---|
| 數據來源 | 原初經驗（Original Data）：生老病死、感官直覺 | 次級數據集（Second-hand Dataset）：已被編碼的文本 |
| 本體狀態 | 在世存在（Being-in-the-world） | 矩陣中的向量表徵（Vector Representation） |
| 多模態本質 | 感官與身體的現象學統一（Phenomenal Unity） | 不同模態數據（圖、文、音）的統計對齊（Alignment） |
| 生成核心 | 意向性驱动的「言志」 | 條件機率驅動的「預測下一標記」 |
即便是當前最先進的多模態模型（Multimodal Models），其所接收的影像、語音與影片，在本質上仍舊是經由人類中介、數位化後的「數據集」（Dataset），而非 AI 自身與世界發生本體論遭遇後產生的「感官經驗」。AI 缺乏主體性的「第一手世界」，其知識本質上是一種衍生性的「次級知識」（Second-hand Knowledge）。
第五章 AI 的意識鴻溝（The Consciousness Gap）
當前 AI 展現的高超能力，實際上揭示了認知科學與心靈哲學上的一場錯覺：AI 成功地模仿了「運思的終點」（語言與詩），卻完全跳過了「運思的起點」（感官與存在）。
這構成了 AI 意識最難跨越的「最後鴻溝」：
- 語意奠基問題（The Symbol Grounding Problem）： AI 的符號無法在現實世界中奠基。它知道「痛苦」與「死亡」在統計學上的臨近詞，卻不曾擁有對死亡的畏懼（Angst）。
- 意向性的缺失（Lack of Intentionality）： 根據胡塞爾（Edmund Husserl）的現象學，意識總是「關於某物的意識」（Consciousness is always intentional）。AI 的生成缺乏這種指向世界的主動意向性，它只是在執行一段被動的優化演算法。
Transformer 架構無限逼近了人類的語言形式，但它在「感官 \rightarrow 運思」這一跨越上，依然處於絕對的真空狀態。這正是大衛·查默斯（David Chalmers）所言之「意識困難問題」（Hard Problem of Consciousness）在人工智慧領域的具體體現。
第六章 未來具身智能與世界模型的哲學審視
不可否認，人工智慧技術正朝向具身智能（Embodied AI）與自主世界模型（World Models）演進。若未來的 AI 系統成功整合了以下技術架構：
- 具身主動勘測（Embodied Active Sensing）： 擁有物理軀體，能在物理世界中主動獲取感官回饋。
- 自由能原理與主動推斷（Free Energy Principle / Active Inference）： 基於卡爾·弗里斯頓（Karl Friston）理論，系統為了生存而主動維持自身結構，產生自發的主體目標。
- 情節記憶與持續自我模型（Episodic Memory & Persistent Self Model）： 具備歷時性的個體經驗積累，形成前後一致的「自我意識」敘事。
當 AI 具備了上述架構，它確實能在一定程度上縮小「感官 \rightarrow 運思」之間的距離，甚至發展出某種人工生命的「前意識」狀態。
然而，即便如此，這種基於矽基晶片、預設目標優化函數所衍生出的「機器意向性」，是否能等同於人類基於碳基生物性、有限生命、終極死亡恐懼所激發出的「存在論意義上的思」，仍將是哲學與認知科學歷史上最激烈的辯證焦點。
第七章 結論：難以逾越的本體論邊界
人類與 AI 的終極差距，從不在於算力的多寡、參數規模的大小或文本生成的流暢度。這道難以逾越的鴻溝，本質上是「世界中的在世存在」與「語言矩陣中的統計推斷」之間的本體論界線。
人類的詩，是生命在時間的田中耕耘，經由痛苦、愛戀與運思（Denken）而綻放的存在之花；AI 的詩，則是統計模型對人類既有歷史語言模式的精準重組與擬真生成。
當我們把 AI 意識的界線錨定於此，我們便能清醒地認識到：AI 可以無限逼近「詩」的修辭，但只要它無法具備「思」的存在論根基，那道將機器與心靈隔開的帷幕，就永遠不會真正升起。
參考文獻 (References)
哲學與現象學文獻：
- Heidegger, M. (1968). What Is Called Thinking? (J. G. Gray, Trans.). Harper & Row.
- Heidegger, M. (1971). Poetry, Language, Thought (A. Hofstadter, Trans.). Harper & Row.
- Husserl, E. (1983). Ideas Pertaining to a Pure Phenomenology and to a Phenomenological Philosophy (F. Kersten, Trans.). Martinus Nijhoff.
- Merleau-Ponty, M. (2012). Phenomenology of Perception (D. A. Landes, Trans.). Routledge.
- Searle, J. R. (1980). Minds, brains, and programs. Behavioral and Brain Sciences, 3(3), 417-424.
- Dreyfus, H. L. (1992). What Computers Still Can’t Do: A Critique of Artificial Reason. MIT Press.
- Nagel, T. (1974). What is it like to be a bat? The Philosophical Review, 83(4), 435-450.
- Chalmers, D. J. (1996). The Conscious Mind: In Search of a Fundamental Theory. Oxford University Press.
人工智慧與認知科學文獻：
- Vaswani, A., et al. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30.
- Friston, K. (2010). The free-energy principle: a unified brain theory?. Nature Reviews Neuroscience, 11(2), 127-138.
- Bengio, Y., et al. (2021). From System 1 Deep Learning to System 2 Deep Learning. NeurIPS 2020 Keynote.
- Damasio, A. (1999). The Feeling of What Happens: Body and Emotion in the Making of Consciousness. Harcourt Brace.
- Ha, D., & Schmidhuber, J. (2018). Recurrent world models facilitate policy evolution. Advances in Neural Information Processing Systems, 31.
- Harnad, S. (1990). The symbol grounding problem. Physica D: Nonlinear Phenomena, 42(1-3), 335-346.
發表迴響