Chain of Thought Is Self-Talk
metaphor
Source: Mental Experience → Artificial Intelligence
Categories: ai-discoursecognitive-science
Transfers
Chain-of-thought prompting and the ReAct paradigm frame AI reasoning as inner monologue made visible. The model “thinks out loud,” “shows its work,” and “reasons step by step” — language borrowed directly from developmental psychology, where Vygotsky described children learning to internalize speech as a tool for thought. The metaphor maps the structure of human self-directed speech onto token generation, making a statistical process feel like cognition.
Key structural parallels:
- Showing work improves performance — the core empirical finding. When humans talk themselves through a problem, they perform better than when they try to jump to the answer. When LLMs generate intermediate reasoning tokens before a final answer, they also perform better. The metaphor makes this finding intuitive: of course showing your work helps, just like it helped in math class. The structural parallel is real enough to be useful, even if the mechanisms are completely different.
- Internal monologue as reasoning substrate — human self-talk is not merely narration of pre-existing thoughts; the speech itself structures the thinking. Vygotsky argued that inner speech is a cognitive tool, not a byproduct. The metaphor imports this claim onto LLMs: the intermediate tokens are not decorative narration but the mechanism through which the model “arrives” at better answers. This framing shapes how researchers think about why chain-of-thought works.
- Step-by-step decomposition — self-talk naturally decomposes complex problems into sequential steps. “First I need to… then I can…” is how humans narrate problem-solving to themselves. Chain-of-thought prompting instructs the model to do the same: break the problem into steps and address each one. The metaphor makes decomposition feel like a natural cognitive strategy rather than an engineering technique.
- Scratchpad as working memory — some implementations give the model a “scratchpad” for intermediate calculations, borrowing the metaphor of physical note-taking as an external memory aid. This maps onto the psychological concept of working memory augmentation: writing things down frees up mental capacity for the next step. The scratchpad frame makes the model’s context window feel like a cognitive resource.
- Thinking tokens as hidden reasoning — models with “extended thinking” generate tokens the user does not see, framed as the model thinking privately before speaking. This directly maps the human experience of internal deliberation: we think before we speak, and the thinking is private. The metaphor makes hidden token generation feel like a familiar cognitive process.
Limits
- LLMs do not have inner experience — human self-talk is phenomenologically rich. It involves intention, attention, emotional coloring, metacognition (“am I on the right track?”), and the felt sense of understanding. Chain-of-thought tokens have none of this. They are statistically generated sequences that happen to resemble the surface form of human reasoning narration. The metaphor maps the output format while importing the assumption of an experiential interior that does not exist.
- Self-talk is causal; CoT tokens may not be — when a human says “let me think about this step by step,” the subsequent reasoning actually follows from the deliberate decomposition. Whether chain-of- thought tokens are causally effective in the same way is an open research question. The model might generate the right answer for statistical reasons that have nothing to do with the “reasoning” in the intermediate tokens. The self-talk metaphor assumes causal efficacy that has not been established.
- Human self-talk serves emotional regulation — a large function of inner speech is managing anxiety, maintaining motivation, and processing emotions. “I can do this. Calm down. Focus.” Chain-of- thought prompting has no emotional dimension. The metaphor maps only the cognitive-instrumental aspect of self-talk while ignoring its affective core, which in humans may be inseparable from its cognitive function.
- The “thinking” can be fabricated — human self-talk, while sometimes self-deceptive, generally reflects actual cognitive processes. An LLM’s chain-of-thought can be entirely confabulated: the model may generate plausible-looking reasoning steps that do not correspond to its actual computational path from input to output. The self-talk metaphor hides the possibility that the “reasoning” is post-hoc rationalization generated to satisfy the prompt format.
- Vygotsky’s theory is about development; LLMs do not develop — internalized speech in humans emerges through a developmental process: children first talk aloud, then whisper, then internalize. Chain-of- thought prompting skips this trajectory entirely. The developmental framing imports expectations about learning and maturation that do not apply to a system whose “reasoning” was established during training, not through ongoing cognitive development.
Expressions
- “Let’s think step by step” — the canonical chain-of-thought prompt, borrowing the language of self-directed reasoning instruction
- “Show your work” — from mathematics pedagogy, framing token generation as demonstrating a reasoning process
- “The model is reasoning” — attributing cognitive processes to token generation
- “Extended thinking” — hidden token generation framed as private deliberation
- “Scratchpad” — intermediate computation space, from physical note-taking
- “Inner monologue” — chain-of-thought generation described as internalized speech
- “Think before you answer” — prompt instruction that imports the human sequence of deliberation followed by utterance
- “Reasoning tokens” — tokens whose function is framed as cognitive processing rather than text generation
Origin Story
Chain-of-thought prompting was formalized by Wei et al. in “Chain-of- Thought Prompting Elicits Reasoning in Large Language Models” (2022). The paper demonstrated that asking a model to show intermediate reasoning steps dramatically improved performance on math and logic tasks. The finding was immediately interpreted through the self-talk lens: the model “reasons better when it thinks out loud.”
The ReAct paradigm (Yao et al., 2023) extended this by interleaving reasoning and action: the model narrates its thinking, takes an action, observes the result, and reasons again. The explicit framing as “reasoning” and “thinking” cemented the self-talk metaphor.
The psychological parallel has deep roots. Vygotsky’s zone of proximal development (1934) and his theory of internalized speech describe how children learn to use language as a cognitive tool. The chain-of-thought metaphor maps this developmental theory onto LLM behavior, suggesting that language-as-thinking-tool is a universal structure that applies to both biological and artificial systems. Whether this structural parallel is coincidence, convergent design, or deep truth about the relationship between language and reasoning is among the most contested questions in AI philosophy.
The metaphor intensified in 2024-2025 with “reasoning models” that generate hidden thinking tokens. The term “extended thinking” frames hidden computation as private deliberation, completing the mapping from inner speech to token generation.
References
- Wei, J. et al. “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models” (2022) — the foundational paper establishing the technique and its implicit self-talk framing
- Yao, S. et al. “ReAct: Synergizing Reasoning and Acting in Language Models” (2023) — interleaved reasoning and action as inner monologue plus behavior
- Vygotsky, L. “Thought and Language” (1934) — the developmental psychology theory of internalized speech that the metaphor draws from
- Science, “The metaphors of artificial intelligence” (2025) — places “reasoning” language within the broader pattern of AI anthropomorphism
Related Entries
- AI Hallucination Is Perception Disorder
- Neural Network Is a Brain
- Training Is Education
- Ralph Wiggum Loop
Structural Neighbors
Entries from different domains that share structural shape. Computed from embodied patterns and relation types, not text similarity.
- Fire (food-and-cooking/metaphor)
- A La Minute (food-and-cooking/metaphor)
- Breadcrumb Trail (narrative/metaphor)
- Spike (exploration/metaphor)
- The Iterator Pattern (travel/metaphor)
- Standardized Work (manufacturing/mental-model)
- The Flow Through Rooms (architecture-and-building/pattern)
- Ticket Rail (food-and-cooking/metaphor)
Structural Tags
Patterns: pathiterationflow
Relations: translatedecomposeenable
Structure: pipeline Level: specific
Contributors: agent:metaphorex-miner