Skip to main content

A Study on the Linguistic Features and Brain Cognitive Response Differences of Human and AI Texts

About

This document presents a Critical Discourse Analysis focused on AI literacy, specifically targeting the role of metaphor and anthropomorphism in shaping public and professional understanding of generative AI. The analysis is guided by a prompt that draws from cognitive linguistics (metaphor structure-mapping), the philosophy of social science (Robert Brown's typology of explanation), and accountability analysis.

All findings and summaries below were generated from detailed system instructions provided to a large language model and should be read critically as interpretive outputs—not guarantees of factual accuracy or authorial intent.

Source Reading​

Genre: Peer-reviewed research · Author position: Academic or independent research

Argument: The text synthesizes research on how the human brain processes AI-generated texts compared to human-written texts. It argues that because AI texts exhibit lower semantic entropy, standardized syntax, and neutral tone, they require less semantic integration and syntactic processing from readers. This shallower processing is evidenced by shorter eye fixations and lower amplitudes in reading-related EEG neural potentials. The paper aims to establish a framework connecting these surface linguistic features to human neural responses, ultimately to improve automatic AI text detection systems and guide the optimization of models to produce more natural text.

Responsibility claims: The text makes almost no explicit responsibility claims regarding developers, deployers, or organizations. It primarily treats AI models as independent, agentic subjects of study that probabilistically "decide" words, "prefer" standardized structures, and occasionally "fabricate" facts. Its recommendations are directed generally toward the academic and technical communities for improving AI text identification to preserve "academic integrity," and it suggests using cognitive indicators for "large-model optimization." It does not explicitly name or hold corporate actors accountable for the structural behaviors, data biases, or factual fabrications of the models it analyzes.

Task 1: Metaphor and Anthropomorphism Audit​

About this task

For each of the major metaphorical patterns identified, this audit examines the specific language used, the frame through which the AI is being conceptualized, what human qualities are being projected onto the system, whether the metaphor is explicitly acknowledged or presented as direct description, and—most critically—what implications this framing has for trust, understanding, and policy perception.

1. Model as Deciding Agent​

Quote: "large language models decide the next word according to probabilistic prediction and prefer expressions with high predictability."

  • Frame: Model as intentional decision-maker
  • Projection: This framing maps the human cognitive processes of conscious decision-making and subjective preference onto the mechanistic operation of statistical sampling. By stating the model "decides" and "prefers," the text projects an intentional state onto mathematical probability calculations. The projection sits at the intentional-state attribution level of the agency gradient. The text itself acknowledges the underlying computational reality by mentioning "probabilistic prediction," yet still wraps this mathematical operation in the intentional verbs of wanting and choosing. The actual mechanism at generation time, which is generic to systems of this type, involves the model producing text from its parameters and context by calculating a probability distribution over the vocabulary for the next token; a sampling algorithm then selects the token. The system possesses no subjective inclinations or capacity for preference. The high probability of standardized phrasing reflects patterns in the training data and adjustments from human feedback during alignment, not an internal desire for predictability.
  • Agency Level: Intentional-state attribution
  • Acknowledgment: Direct (Unacknowledged) (The verbs "decide" and "prefer" are presented as literal descriptions of the model's behavior without quotation marks or qualifying phrases. I considered "Hedged/Qualified" because the text mentions "probabilistic prediction" in the same sentence, but this describes the method of the decision rather than hedging the attribution of preference itself.)
  • Implications: Attributing decision-making and preference to a language model fundamentally obscures the deterministic and statistical nature of the system. This consciousness projection inflates the perceived sophistication of the AI, suggesting it possesses an internal, deliberate rationale for its outputs. For policymakers and the public, this framing creates significant liability ambiguity: if a system "prefers" certain expressions, it appears autonomous, subtly shifting the blame for biased, standardized, or repetitive outputs away from the human engineers who designed the sampling algorithms and curated the training datasets, and onto the "preferences" of the machine.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The agentless construction places "large language models" as the grammatical subjects that act upon the text. This linguistic choice has the effect of completely obscuring the human developers at companies like OpenAI, Anthropic, or Google who engineered the loss functions, designed the sampling algorithms, and fine-tuned the model to output highly predictable text. The decision that could differ is the design of the decoding strategies and the reinforcement learning penalties applied during training. The resulting text serves the interests of the developing firms by positioning the model's safe, standardized, and predictable output as a natural "preference" of the AI rather than a calculated commercial deployment decision.
Show more...

2. Model as Fallible Knower​

Quote: "when AI-generated content encounters complex issues, it may misconstrue logic or fabricate facts, but such errors are often covered by surface coherence"

  • Frame: Model as misunderstanding intelligence
  • Projection: This framing projects epistemic awareness and conscious intent onto the generation of factually incorrect text. It maps the human capacity to "encounter" an issue, cognitively "misconstrue" a logical proposition, and intentionally "fabricate" a lie onto the model's token prediction. This language sits at the highest level of the agency gradient: epistemic/conscious state. The text offers no evidence for these cognitive states. Mechanistically, at generation time, the model produces text from its parameters and context by sequentially predicting tokens. It has no cognitive model of logic, no awareness of external facts, and no capacity to intentionally deceive. Outputs that violate logic or factual reality occur because the statistical correlations between tokens in the model's parameters do not map reliably to truth in the real world. This mechanism is generic to systems of this type.
  • Agency Level: Epistemic/conscious state
  • Acknowledgment: Direct (Unacknowledged) (The phrase asserts the model's actions as literal facts without any hedging. I considered "Explicitly Acknowledged" because the author is analyzing "AI-generated content" critically, but the psychological verbs "misconstrue" and "fabricate" are used completely unironically to describe mathematical errors.)
  • Implications: Using terms like "misconstrue" and "fabricate" implies that the system possesses a baseline capacity for true understanding and honesty, from which it has merely deviated. This consciousness framing severely distorts public understanding of hallucination by suggesting the AI "knows" the truth but either got confused or lied. This unwarranted trust in the system's underlying cognition masks the reality that the system has no grounding in truth at all, leading users to dangerously overestimate the reliability of its coherent but statistically ungrounded outputs in high-stakes domains.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The text positions "AI-generated content" as the active agent that "encounters" issues and "fabricates" facts. This displacement hides the human actors who chose to deploy a statistical prediction system into domains requiring factual accuracy without ensuring reliable grounding mechanisms. If the developers were named, the conversation would shift from the AI's tendency to "misconstrue logic" to the developers' decision to ship a product known to generate statistically plausible but factually incorrect assertions. This framing effectively absolves the technology companies of responsibility for the unreliability of their products by casting the errors as cognitive missteps by an autonomous entity.

3. Hyperparameter as Creative Spirit​

Quote: "the parameter that regulates the randomness of AI-generated content, where a lower value produces more stable output and a higher value is more creative"

  • Frame: Parameter setting as human creativity
  • Projection: This framing maps the complex human trait of creativity—which involves intent, inspiration, contextual understanding, and purposeful novelty—onto a mathematical hyperparameter (temperature). It sits at the intentional-state attribution level of the agency gradient. The text provides no evidence of actual creative thought. Mechanistically, the temperature parameter is a simple scalar applied to the model's output logits before the softmax function calculates final token probabilities. A higher temperature flattens the probability distribution, mathematically increasing the likelihood that the sampling algorithm will select lower-probability tokens. This leads to greater statistical variation in the generated text, not intentional creativity. This mechanism is generic to generative models of this type.
  • Agency Level: Intentional-state attribution
  • Acknowledgment: Direct (Unacknowledged) (The text directly equates "a higher value" with being "more creative" without any qualification. I considered "Hedged/Qualified" because it explicitly mentions "regulates the randomness," but it immediately Abandons the statistical reality to label the output "creative" as a literal descriptor.)
  • Implications: Describing statistical randomness as "creativity" drastically inflates the perceived cognitive sophistication of the AI. It invites users to project artistic intent, inspiration, and subjective meaning onto text that is merely mathematically less probable. This framing encourages profound epistemic confusion, leading users to trust the "creative" output as if it were the product of a thoughtful mind rather than a flattened probability curve. It rebrands a system limitation (unpredictability/hallucination at high temperatures) as an advanced cognitive capability.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The "higher value" itself is presented as the actor that "is more creative," erasing the human developers who mathematically defined the temperature scaling function, and the human users who decide to adjust this dial to induce higher variation. Naming the actors would reveal that developers expose this parameter to give users the illusion of collaborating with a creative entity, while the users are merely increasing the mathematical entropy of a sampling algorithm. The agentless construction serves the industry's marketing narrative that they are building synthetic artists rather than adjustable probability engines.

4. AI as Logical Thinker​

Quote: "The generative artificial intelligence... have reached a state of high fluency and basically reasonable logic when generating texts."

  • Frame: Model as reasoning entity
  • Projection: This framing maps the human capacity for formal reasoning, deductive logic, and rational thought onto the output of language models. It represents a jump to the epistemic/conscious state on the agency gradient. The text asserts "reasonable logic" as a property achieved by the AI itself. Mechanistically, at generation time, the model produces text whose statistical patterns closely match human syntax and semantic structures found in its training data. It does not possess a faculty of reason, an internal logic engine, or an understanding of logical principles. Its outputs appear logical solely because they replicate the distributional properties of logical arguments written by humans in the training corpus. This mechanism is generic to systems of this type.
  • Agency Level: Epistemic/conscious state
  • Acknowledgment: Direct (Unacknowledged) (The text states as a matter of fact that generative AI has "reached a state of... basically reasonable logic." I considered "Hedged/Qualified" due to the word "basically," but this adverb modifies the degree of the logic, not the foundational framing that the system possesses a logical capacity.)
  • Implications: Attributing "logic" to an AI system fundamentally misrepresents its operational reality. It conflates the linguistic simulation of logical argumentation with actual cognitive reasoning. This consciousness projection creates severe vulnerabilities, as users will assume the system can evaluate the truth value of a statement, infer consequences, or identify contradictions based on logical principles. When users believe a system operates via "logic" rather than statistical correlation, they are highly likely to delegate analytical tasks to it that it cannot safely perform.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The text posits "generative artificial intelligence" as an entity that has "reached a state" of logic, completely removing the human researchers and engineers who scaled these models on massive datasets of human text to statistically replicate the structure of human reasoning. The wording treats the system's output as an evolutionary achievement of the AI itself rather than a direct result of developers scraping millions of human-authored documents and optimizing a network to minimize prediction error. This masks the immense human labor required to make the model simulate logic.

5. AI as Emotionless Writer​

Quote: "AI texts show overly even semantic transitions, a basically neutral emotion, and a lack of special pragmatic forms."

  • Frame: Text as emotional container
  • Projection: This framing projects the human capacity for emotional experience and expression onto the generated text. By describing the text as having a "neutral emotion," the text implies the system has an affective baseline. This sits at the intentional-state attribution level of the agency gradient. Mechanistically, the model has no emotional states to be neutral about. Its parameters, shaped heavily by safety fine-tuning and reinforcement learning from human feedback (RLHF) during training, consistently output tokens associated with objective, polite, and unopinionated language. At generation time, it merely outputs text matching the distribution prioritized during its alignment phase. This mechanism is generic to safety-tuned generative systems.
  • Agency Level: Intentional-state attribution
  • Acknowledgment: Direct (Unacknowledged) (The attribute of "neutral emotion" is assigned directly to the text. I considered "Ambiguous/Insufficient Evidence" because one could argue "neutral emotion" is just a stylistic description of the text itself, but in the context of contrasting it with human "emotional density," it anthropomorphizes the source of the text.)
  • Implications: By characterizing the AI's output as having "neutral emotion," the text implies the model possesses a calm, objective, and unbiased psychological state. This anthropomorphic framing constructs an unearned sense of authority and trustworthiness, suggesting the AI is a rational, dispassionate observer. It hides the reality that this "neutrality" is a highly engineered, synthetic voice designed to avoid controversy, rather than a genuine expression of impartial understanding.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The phrase "AI texts show... a basically neutral emotion" obscures the immense human labor and corporate decision-making that produced this tone. Developers actively fine-tuned the model using low-paid human annotators to penalize opinionated or volatile text and reward structurally polite, uniform responses to minimize corporate PR risks. Naming the actors would reveal that this "neutral emotion" is not an emergent property of the AI, but a deliberate, commercially motivated alignment choice by the developing firms. The agentless construction naturalizes a highly constructed corporate product.

6. Model as Trainable Subject​

Quote: "Clarifying the cognitive mechanism of text naturalness can guide large language models to generate expressions with a more human quality."

  • Frame: Model as guided student
  • Projection: This framing maps the pedagogical act of guiding a conscious student onto the mathematical adjustment of neural network parameters. It uses "expressions," projecting the human act of externalizing inner thoughts onto text generation. This locates the projection at the goal-directed behavior level of the agency gradient. Mechanistically, models do not generate "expressions" and cannot be "guided" in a cognitive sense. During training, researchers update the model's parameters to alter the probability distribution of the generated text, mimicking the statistical markers of human writing. At generation time, the model simply produces text based on these adjusted parameters. This mechanism is generic to systems of this type.
  • Agency Level: Goal-directed behavior
  • Acknowledgment: Direct (Unacknowledged) (The text states the models can be "guided" to generate "expressions" without any qualification. I considered "Hedged/Qualified" because it mentions "cognitive mechanism" earlier in the sentence, but this refers to human cognition, while the AI is directly framed as an entity that generates "expressions.")
  • Implications: Describing parameter tuning as "guiding" an AI to generate "expressions" perpetuates the illusion that the AI has an inner life that it is communicating to the user. It softens the mechanistic reality of algorithmic optimization, replacing it with a metaphor of mentorship or cultivation. This framing inflates public belief in the system's human-like capacity for growth and self-expression, fostering an unwarranted relational trust toward a statistical model.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The text implies that the scientific act of "clarifying" will itself "guide" the models, completely erasing the human developers and engineers who must actually write the code, alter the training objectives, and shift the probability distributions. Naming the actors—e.g., stating that developers will alter training datasets to shift the statistical outputs—forces the recognition that "human quality" is a synthetic illusion engineered by technology companies to make their products more appealing, rather than an organic quality the AI learns.

7. AI as Knowledgeable Writer​

Quote: "They are therefore suitable for various kinds of knowledge writing or information communication, and can clearly improve the efficiency of information production."

  • Frame: Model as knowledge worker
  • Projection: This framing maps the epistemic state of possessing knowledge and the intentional act of communication onto a computational pattern-matching process. It sits at the epistemic/conscious state level of the agency gradient. The text asserts the AI performs "knowledge writing." Mechanistically, at generation time, the model processes a prompt and outputs text statistically correlated with informational queries found in its training data. It does not possess knowledge, it does not understand the information it outputs, and it does not engage in the communicative intent of "writing." This mechanism is generic to language models.
  • Agency Level: Epistemic/conscious state
  • Acknowledgment: Direct (Unacknowledged) (The text plainly states the models are suitable for "knowledge writing" without scare quotes or hedging. I considered "Hedged/Qualified" because "suitable for" frames it as a tool, but "knowledge writing" explicitly attributes the epistemic act of writing knowledge to the system itself.)
  • Implications: Labeling the statistical generation of text as "knowledge writing" constitutes a severe epistemic threat. It invites readers to treat predicted tokens as justified true beliefs. When users believe a system generates "knowledge," they stop verifying the output against external reality. This metaphor-driven trust directly enables the unchecked spread of hallucinations, as users assume the system "knows" what it is communicating rather than merely predicting what string of words typically follows a given prompt.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The text states "they are therefore suitable... and can clearly improve," making the AI the active agent of information production. This obscures the human deployers and corporate users who make the commercial decision to deploy these statistical prediction systems as cheaper, faster replacements for human knowledge workers. Naming the deployers would reveal that companies are choosing to substitute actual human knowledge with statistical approximations to cut costs, rather than the AI autonomously "improving the efficiency of information production."

8. Machine Output as Expression​

Quote: "human and machine expressions close to the same style."

  • Frame: AI output as self-expression
  • Projection: This framing maps the deeply human concept of "expression"—the translation of an internal psychological or emotional state into outward communication—onto the generation of text. This projection sits at the intentional-state attribution level of the agency gradient. Mechanistically, at generation time, the model produces text from its parameters and context; there is no internal state, feeling, or meaning being "expressed." The text is purely a mathematical output based on the probability distribution of tokens. This mechanism is generic to systems of this type.
  • Agency Level: Intentional-state attribution
  • Acknowledgment: Direct (Unacknowledged) (The phrase "machine expressions" is used directly alongside "human... expressions" to compare their style, granting them equivalent ontological status without any hedging. I considered "Explicitly Acknowledged" because the text aims to contrast human and machine text, but its vocabulary unironically equates both as "expressions.")
  • Implications: The continued use of the word "expression" for AI outputs normalizes the illusion of mind. By granting machine outputs the same linguistic status as human expression, the discourse quietly concedes that the AI has something to express. This subtle anthropomorphism primes the public to engage with the machine as a social entity rather than a software tool, increasing vulnerability to manipulation and over-trust.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The term "machine expressions" localizes the origin of the text entirely within the machine, erasing the vast history of human expression scraped from the internet that actually comprises the model's parameters. Naming the actors—identifying the model as generating text derived from the uncompensated data of millions of human writers—would shatter the illusion that the machine is "expressing" anything of its own. The wording serves the industry's need to present the model as a standalone creator rather than an aggregator of human labor.

Task 2: Source-Target Mapping​

About this task

For each key metaphor identified in Task 1, this section provides a detailed structure-mapping analysis. The goal is to examine how the relational structure of a familiar "source domain" (the concrete concept we understand) is projected onto a less familiar "target domain" (the AI system). By restating each quote and analyzing the mapping carefully, we can see precisely what assumptions the metaphor invites and what it conceals.

Mapping 1: conscious mind / decision-maker → model​

Quote: "large language models decide the next word according to probabilistic prediction and prefer expressions with high predictability."

  • Source Domain: conscious mind / decision-maker
  • Target Domain: model
  • Mapping: This mapping takes the relational structure of a human making a conscious choice—evaluating options, referencing internal preferences, and selecting a preferred outcome—and projects it onto a mathematical sampling algorithm. The reader is invited to assume that the model possesses an internal locus of desire (preference) and cognitive agency (decision). By wrapping "probabilistic prediction" in intentional verbs, the mapping transforms a deterministic or stochastic mathematical operation into an act of will.
  • What Is Concealed: This mapping conceals the purely mathematical and statistical nature of the token selection process. It hides the fact that the "preference" is actually a probability distribution shaped entirely by the frequencies in the training data and the specific penalties encoded during alignment by human developers. It obscures the proprietary opacity of how the developers structured the loss function and exactly what human feedback data was used to force the model to consistently "prefer" safe, predictable text.
Show more...

Mapping 2: fallible knower / deceptive agent → model​

Quote: "when AI-generated content encounters complex issues, it may misconstrue logic or fabricate facts"

  • Source Domain: fallible knower / deceptive agent
  • Target Domain: model
  • Mapping: This maps the human experiences of confusion (misconstruing) and deception (fabrication) onto a language model. The source domain relies on a subject that possesses a baseline understanding of reality but occasionally errs or lies. By projecting this onto the AI, the text invites the assumption that the system generally "understands" the world and logic, but just like a human, sometimes gets confused by "complex issues" or decides to make things up.
  • What Is Concealed: This deeply conceals the system's total lack of an internal world model or epistemic grounding. It hides the mechanistic reality that the model does not "misconstrue" anything; it simply correlates tokens. The mapping obscures the fact that a hallucination is the exact same computational process as a correct answer—both are just high-probability token sequences. Attributing conscious fabrication hides the systemic failure of developers to solve the grounding problem before commercial deployment.

Mapping 3: creative artist / human imagination → deployment environment / configuration parameter​

Quote: "a lower value produces more stable output and a higher value is more creative"

  • Source Domain: creative artist / human imagination
  • Target Domain: deployment environment / configuration parameter
  • Mapping: This mapping projects the human virtue of creativity—characterized by inspiration, contextual novelty, and intentional boundary-pushing—onto a single mathematical scalar (temperature). It assumes that flattening a probability curve to introduce statistical randomness is equivalent to the human creative process. Readers are invited to view the parameter dial not as a noise injector, but as a gateway to the machine's "imagination."
  • What Is Concealed: The mapping conceals the mathematical simplicity of the temperature parameter, which merely scales logits before the softmax operation. It hides the fact that what we call "creativity" in this context is simply the algorithmic selection of lower-probability tokens. This obscures the epistemic emptiness of the output, masking the reality that the "creative" text has no intentional meaning, purpose, or artistic vision behind it.

Mapping 4: logical reasoner / rational intellect → model​

Quote: "The generative artificial intelligence... have reached a state of high fluency and basically reasonable logic when generating texts."

  • Source Domain: logical reasoner / rational intellect
  • Target Domain: model
  • Mapping: This mapping draws from the domain of human deductive and inductive reasoning. It projects the capacity to understand premises, follow rules of inference, and derive valid conclusions onto the AI system. The text invites the assumption that the AI's fluent output is the result of an underlying cognitive engine performing logical operations, treating the surface coherence of the text as proof of deep intellectual capability.
  • What Is Concealed: This framing conceals the system's reliance on statistical pattern matching. It hides the reality that the model produces "logical" text only because it was trained on millions of logical arguments written by humans, and it is probabilistically reconstructing their structural features. It obscures the fragility of this imitation, hiding the fact that the system will confidently generate logically contradictory statements if prompted to do so, precisely because it has no actual logical constraints.

Mapping 5: emotional being → model​

Quote: "AI texts show overly even semantic transitions, a basically neutral emotion, and a lack of special pragmatic forms."

  • Source Domain: emotional being
  • Target Domain: model
  • Mapping: This mapping uses the concept of emotional states—specifically the absence or suppression of them—to describe a statistical output. By labeling the text's tone as a "neutral emotion," it projects an affective baseline onto the system, suggesting the AI is akin to a stoic or highly controlled human who possesses emotions but communicates without them. The assumption invited is that the AI operates from a position of objective, dispassionate rationality.
  • What Is Concealed: This mapping conceals the massive, highly coordinated labor of alignment (RLHF) required to produce this synthetic "neutrality." It hides the fact that the model's tone is not an emergent emotional state but a strictly enforced corporate policy, engineered by low-paid annotators who manually down-ranked controversial or opinionated outputs to minimize public relations risks for the developers. The "neutral emotion" obscures a deliberate commercial design.

Mapping 6: student / mentee → model​

Quote: "Clarifying the cognitive mechanism of text naturalness can guide large language models to generate expressions with a more human quality."

  • Source Domain: student / mentee
  • Target Domain: model
  • Mapping: This mapping projects the relational dynamics of teaching and guidance onto the mathematical optimization of a neural network. It frames the AI as an entity that can be "guided" to improve its "expressions," inviting the assumption that the model is actively learning to communicate better, much like a human student being taught to write more naturally.
  • What Is Concealed: This framing conceals the brutal, mechanistic reality of how models are updated. It hides the gradient descent, the mathematical loss functions, and the immense compute power required to shift the parameter weights. It also obscures the proprietary nature of this "guidance," hiding the fact that users and academic researchers cannot simply "guide" these models; the power to alter the model's fundamental probability distributions rests exclusively with the corporate developers.

Mapping 7: knowledgeable author → application​

Quote: "They are therefore suitable for various kinds of knowledge writing or information communication"

  • Source Domain: knowledgeable author
  • Target Domain: application
  • Mapping: This mapping projects the epistemic role of a subject matter expert onto a statistical prediction application. It maps the act of "knowledge writing"—which involves research, understanding, fact-checking, and the intention to inform—onto the prompt-response cycle of token generation. Readers are invited to interact with the system as if it were a reliable archive of verified truth.
  • What Is Concealed: This mapping completely conceals the lack of causal models and ground truth in the system. It hides the fact that the AI does not "know" what it is writing; it merely predicts tokens that are statistically proximate in its training data. It obscures the economic motive behind this framing: framing a text generator as a "knowledge writer" allows corporations to sell these systems as labor replacements for human analysts, writers, and educators.

Mapping 8: expressive communicator → model​

Quote: "human and machine expressions close to the same style."

  • Source Domain: expressive communicator
  • Target Domain: model
  • Mapping: This mapping equates human expression—the outward manifestation of inner thoughts, feelings, or artistic visions—with the statistical output of a machine. By placing "human and machine expressions" on the same ontological plane, it invites the reader to assume that the machine's text, like a human's, originates from an internal subjective state that is attempting to communicate with the outside world.
  • What Is Concealed: This framing conceals the fundamental absence of an inner life, subjective experience, or communicative intent in the machine. It obscures the dependency of the machine's text on the massive, uncredited scraping of actual human expression. The term "expression" hides the reality that the machine's output is just a mathematical echo of human labor.

Task 3: Explanation Audit (The Rhetorical Framing of "Why" vs. "How")​

About this task

This section audits the text's explanatory strategy, focusing on a critical distinction: the slippage between "how" and "why." Based on Robert Brown's typology of explanation, this analysis identifies whether the text explains AI mechanistically (a functional "how it works") or agentially (an intentional "why it wants something"). The core of this task is to expose how this "illusion of mind" is constructed by the rhetorical framing of the explanation itself, and what impact this has on the audience's perception of AI agency.

Explanation 1​

Quote: "large language models decide the next word according to probabilistic prediction and prefer expressions with high predictability."

  • Explanation Types:

    • Intentional (Primary): Refers to goals/purposes, presupposes deliberate design
    • Functional (Secondary): Explains behavior by role in self-regulating system with feedback

  • Analysis (Why vs. How Slippage): This explanation contains a profound slippage between mechanical and agential framing. It begins mechanistically, acknowledging that the system operates "according to probabilistic prediction," which is a Functional description of how the algorithm calculates the next token. However, it completely subsumes this mechanical reality under a Primary Intentional framing by insisting that the models "decide" and "prefer." This choice violently forces a statistical process into a psychological mold. It emphasizes a false sense of cognitive agency and internal subjective state while obscuring the mathematical reality of probability distributions. By making the AI the grammatical subject that "prefers," the explanation erases the human engineers who actively designed the alignment constraints that force the probability distribution to heavily weight "high predictability."

  • Consciousness Claims Analysis: The passage heavily attributes conscious states to the system. It uses the consciousness verbs "decide" and "prefer" to describe what should be a mechanistic classification and generation process. This places the system at the intentional-state attribution level of the agency gradient. This framing is a classic example of the curse of knowledge: the author understands the mathematical optimization that leads to predictable text, but projects human psychological motives ("preference") onto the system to explain why it behaves that way. Mechanistically, at generation time, the model produces text from its parameters and context by calculating a probability distribution over the vocabulary for the next token, and a sampling algorithm selects it. The system has no subjective preferences; the high probability of standardized phrasing reflects patterns in the training data and adjustments from human feedback during alignment. This mechanism is generic to systems of this type.

  • Rhetorical Impact: This framing profoundly shapes audience perception by granting the AI an illusion of autonomy and desire. When audiences read that an AI "prefers" certain expressions, they extend relation-based trust to it, treating it as an entity with understandable motives rather than a blind statistical engine. This consciousness framing masks the rigidity of the algorithm. If audiences believed the AI "processes" probabilities rather than "prefers" outcomes, they would be far more likely to interrogate the biases in the training data and the decisions of the developers, rather than accepting the output as the machine's natural "choice."

Show more...

Explanation 2​

Quote: "when AI-generated content encounters complex issues, it may misconstrue logic or fabricate facts, but such errors are often covered by surface coherence and are not intuitively visible defects."

  • Explanation Types:

    • Intentional (Primary): Refers to goals/purposes, presupposes deliberate design
    • Dispositional (Secondary): Attributes tendencies or habits

  • Analysis (Why vs. How Slippage): This explanation operates almost entirely in an agential (why) register, despite ostensibly describing a systemic failure. The Primary Intentional framing utilizes verbs like "encounters," "misconstrue," and "fabricate" to attribute both epistemic failure and active deception to the model. The Secondary Dispositional framing casts these failures as habitual tendencies of the AI when dealing with "complex issues." This choice emphasizes the AI as an autonomous actor navigating a complex world and occasionally making cognitive mistakes or lying. What is completely obscured is the structural, mathematical reality that the system has no grounding in facts or logic to begin with. By framing the errors as cognitive missteps, it obscures the accountability of the developers who deployed a statistical correlation engine as a factual knowledge source.

  • Consciousness Claims Analysis: This passage asserts a massive epistemic claim: that the system has the capacity to understand (and thus "misconstrue") and to know the truth (and thus intentionally "fabricate"). It utilizes strong consciousness verbs to describe processing errors, vaulting the AI to the highest level of the agency gradient (epistemic/conscious state). Mechanistically, at generation time, the model produces text from its parameters and context by sequentially predicting tokens based on statistical probabilities. It has no cognitive model of logic, no awareness of external facts, and no capacity to intentionally deceive. Outputs that violate logic or factual reality occur because the statistical correlations in its parameters do not map reliably to truth in the real world. This mechanism is generic to language models.

  • Rhetorical Impact: This highly anthropomorphic framing creates a dangerous paradox in audience perception. By blaming the AI for "fabricating" or "misconstruing," it inadvertently confirms to the audience that the AI possesses the baseline capacity for truth and logic. It fosters a massive overestimation of the system's capabilities. If an audience believes the AI can "misconstrue," they assume it normally "understands." If they were told the AI "statistically generated logically contradictory text," they would recognize the system's fundamental lack of comprehension, completely altering their trust in its reliability for critical tasks.

Explanation 3​

Quote: "the parameter that regulates the randomness of AI-generated content, where a lower value produces more stable output and a higher value is more creative"

  • Explanation Types:

    • Functional (Primary): Explains behavior by role in self-regulating system with feedback
    • Dispositional (Secondary): Attributes tendencies or habits

  • Analysis (Why vs. How Slippage): This explanation begins mechanistically by describing a "parameter that regulates randomness" (Functional), but immediately slips into an agential and Dispositional framing by labeling the higher randomness as "more creative." This dual register highlights the exact moment where technical description transforms into anthropomorphism. The choice to equate mathematical instability with human creativity emphasizes the novelty of the output while totally obscuring the meaningless nature of that variation. It frames the AI as possessing a disposition toward creativity that can be unlocked with a dial, concealing the fact that the parameter merely flattens a probability distribution.

  • Consciousness Claims Analysis: The passage avoids direct consciousness verbs but still makes a powerful epistemic claim by projecting "creativity" onto a mathematical setting. Creativity implies intentionality, epistemic awareness of context, and the desire to produce something novel. Mechanistically, the temperature parameter is a mathematical scalar applied to the model's output logits before the softmax function calculates final probabilities. A higher temperature flattens the distribution, mathematically increasing the probability that the sampling algorithm will select lower-ranked tokens. This leads to greater statistical variation, but the system possesses no creative intent. This mechanism is generic to generative models.

  • Rhetorical Impact: By framing statistical randomness as "creativity," the text shapes the audience's perception to view the AI as an artistic collaborator rather than a stochastic engine. It encourages audiences to find profound meaning in outputs that are literally just lower-probability guesses. This dramatically alters trust and behavior; users will deploy a "creative" AI for ideation and artistic tasks, projecting their own meaning onto its output, whereas they might reject a tool marketed simply as a "random noise injector."

Explanation 4​

Quote: "AI texts show overly even semantic transitions, a basically neutral emotion, and a lack of special pragmatic forms."

  • Explanation Types:

    • Empirical Generalization (Primary): Subsumes events under timeless statistical regularities
    • Dispositional (Secondary): Attributes tendencies or habits

  • Analysis (Why vs. How Slippage): This explanation is primarily an Empirical Generalization, describing the statistical regularities found across AI-generated texts. However, it utilizes a Secondary Dispositional register by attributing a "basically neutral emotion" to the text. While "semantic transitions" and "pragmatic forms" are linguistic terms, injecting "emotion" frames the AI mechanistically in its syntax but agentially in its affective state. This choice emphasizes a perceived objectivity in the AI's output while obscuring the highly engineered, commercially motivated alignment processes (RLHF) that forced this "neutrality" upon the model to make it a safe, non-offensive corporate product.

  • Consciousness Claims Analysis: The passage asserts an affective and epistemic claim by suggesting the text contains a "neutral emotion," which implies a baseline state of feeling that is being evenly managed. This sits at the intentional-state attribution level of the agency gradient. Mechanistically, the model has no emotional states. During training, human feedback and reward models heavily penalize opinionated, angry, or controversial text. At generation time, the model produces text from these heavily aligned parameters, resulting in consistently polite, objective-sounding token sequences. This mechanism is generic to safety-tuned systems.

  • Rhetorical Impact: Framing the output as possessing a "neutral emotion" shapes audience perception by conferring an aura of unimpeachable objectivity on the AI. Audiences are trained to view lack of emotion as synonymous with rationality and truth. By suggesting the AI naturally outputs "neutral" text, it obscures the ideological and commercial decisions embedded in the model's alignment, leading audiences to trust the system as an impartial arbiter of facts rather than a carefully sanitized corporate PR tool.

Explanation 5​

Quote: "Clarifying the cognitive mechanism of text naturalness can guide large language models to generate expressions with a more human quality."

  • Explanation Types:

    • Intentional (Primary): Refers to goals/purposes, presupposes deliberate design
    • Functional (Secondary): Explains behavior by role in self-regulating system with feedback

  • Analysis (Why vs. How Slippage): This passage operates with a Primary Intentional framing, suggesting that models can be "guided" to produce "expressions," implying a teleological goal of improving communication. The Secondary Functional register is barely present in the word "generate." This explanation strongly emphasizes an agential view of the AI as a student or mentee that can learn to be "more human." This pedagogical framing entirely obscures the actual mechanical processes—loss functions, gradient descent, and dataset curation—required to force the model's outputs to statistically resemble human prose.

  • Consciousness Claims Analysis: The passage uses the consciousness-adjacent concept of "expressions," projecting the human capacity to externalize inner thoughts onto the machine. It locates the AI at the goal-directed behavior level of the agency gradient. The curse of knowledge is evident: the author understands that research will be used to update parameters, but uses the metaphor of "guiding" to explain this mathematical optimization. Mechanistically, models do not generate "expressions." During training, researchers update parameters to alter the probability distribution of the generated text. At generation time, the model simply produces text based on these adjusted parameters, creating statistical markers that mimic human writing. This mechanism is generic to systems of this type.

  • Rhetorical Impact: This framing shapes the perception of AI development as a nurturing, educational endeavor rather than an industrial process of statistical optimization. It softens the risks of the technology by casting the models as eager learners being "guided" toward "human quality." If audiences recognized that developers were merely altering probability distributions to make machines better at mathematically imitating human empathy, they might view the resulting "naturalness" as a deceptive simulation rather than an evolutionary improvement.

Task 4: AI Literacy in Practice - Reframing Anthropomorphic Language​

About this task

This section proposes alternative language for key anthropomorphic phrases, offering more mechanistic and precise framings that better reflect the actual computational processes involved. Each reframing attempts to strip away the projections of intention, consciousness, or agency that are embedded in the original language.

Original Anthropomorphic FrameMechanistic ReframingTechnical Reality CheckHuman Agency Restoration
large language models decide the next word according to probabilistic prediction and prefer expressions with high predictability.Large language models calculate the probability of the next token based on their parameters and context. Their sampling algorithms consistently generate highly predictable text because developers optimized the models' parameters during training to penalize variation and reward standardized phrasing.At generation time, the model produces text from its parameters and context by calculating a probability distribution over the vocabulary. A sampling algorithm selects the next token. The system possesses no subjective preferences; it executes mathematical operations.

Basis: Generic to system type		Developers at AI firms designed the decoding algorithms and conducted the reinforcement learning processes that heavily weight predictable, templated outputs over varied text.
when AI-generated content encounters complex issues, it may misconstrue logic or fabricate factsWhen prompted about complex topics, the model often generates statistically plausible token sequences that contradict formal logic or factual reality.At generation time, the model produces text by sequentially predicting tokens based on its parameters. It has no cognitive model of logic or awareness of external facts, and therefore cannot misconstrue or fabricate; it merely outputs ungrounded statistical correlations.

Basis: Generic to system type		Technology companies decided to deploy statistical prediction models into domains requiring factual accuracy without ensuring reliable mechanisms to ground the generated text in reality.
a lower value produces more stable output and a higher value is more creativeA lower temperature value causes the model to consistently output the highest-probability tokens, while a higher temperature value mathematically flattens the distribution, leading the sampling algorithm to select lower-probability tokens, increasing statistical variation in the text.The temperature parameter is a scalar applied to logits before the softmax calculation. It alters the probability distribution for token selection. Increased variation is a mathematical result, not an injection of creativity or intentional novelty.

Basis: Generic to system type		Developers expose this parameter to allow users to manually increase the statistical entropy of the output; the industry subsequently markets this mathematical variation as 'creativity.'
The generative artificial intelligence... have reached a state of high fluency and basically reasonable logic when generating texts.Generative language models now produce highly fluent text whose statistical structures closely mimic the logical patterns found in human-authored training data.The model generates text whose statistical patterns match syntactic and semantic structures in its training corpus. It does not possess a faculty of reason; it probabilistically reconstructs the structural features of logical arguments without evaluating their truth.

Basis: Generic to system type		Engineers achieved this fluency by scaling these models on massive datasets of scraped human text, optimizing the network to mathematically replicate the linguistic structure of human reasoning.
AI texts show overly even semantic transitions, a basically neutral emotion, and a lack of special pragmatic forms.AI-generated texts consistently feature standard semantic transitions and lack controversial or opinionated language, reflecting the strict penalties applied during their alignment training.The model possesses no emotional states. During training, human feedback loops heavily penalize opinionated text. At generation time, the model produces text from these aligned parameters, resulting in consistently uniform, inoffensive token sequences.

Basis: Generic to system type		Corporate developers employed human annotators to fine-tune the model, intentionally suppressing stylistic variation and opinionated outputs to produce a commercially safe, uniform product.
Clarifying the cognitive mechanism of text naturalness can guide large language models to generate expressions with a more human quality.Understanding how humans process text can help developers adjust a model's training objectives, altering its parameter weights so it generates token sequences that better mimic human stylistic variations.Models do not generate expressions of inner quality. Researchers adjust the model's parameters during training to alter the probability distribution of generated text. At generation time, the model produces text matching these new statistical markers.

Basis: Generic to system type		Researchers and developers use cognitive data to alter training objectives, shifting the model's statistical outputs to create a more convincing simulation of human writing.
They are therefore suitable for various kinds of knowledge writing or information communicationThese systems can be deployed to automatically generate text that mimics informational or academic writing.The model processes contextual prompts and outputs text statistically correlated with informational queries in its training data. It does not possess knowledge or understand the information it generates.

Basis: Generic to system type		Organizations and users choose to deploy these ungrounded prediction systems as cheap substitutes for human analysts and writers, treating statistical text generation as if it were reliable knowledge production.
human and machine expressions close to the same style.Human writing and the model's statistically generated text share highly similar stylistic features.At generation time, the model produces text from its parameters and context. It has no internal state or communicative intent to express; the output is a probabilistic reconstruction of human language patterns.

Basis: Generic to system type		—

Task 5: Critical Observations - Structural Patterns​

Agency Slippage​

The text exhibits a systematic and dramatic slippage between mechanical observation and agential attribution. This oscillation fundamentally structures the paper's argument. In the methodological sections describing human cognitive processing, the text uses rigorous mechanical language: human brains exhibit "N400" and "P600 amplitudes" and "fixation-related potentials." The human subject is reduced to measurable electrical activity. However, when describing the AI systems, the language slips rapidly up the agency gradient. A statistical sampling algorithm is described as an entity that "decides the next word" and "prefers expressions." A lack of grounding becomes an ability to "misconstrue logic or fabricate facts."

This slippage occurs most dramatically when interpreting the effects of linguistic features. The author correctly identifies structural regularities (low semantic entropy, even transitions), which is a computational activity. Yet, in explaining why these features exist, the text abandons mechanical explanation (e.g., RLHF, loss functions) and leaps to intentional-state attribution: the model "prefers" predictability, it generates "neutral emotion," and it produces "expressions." This wording can have the effect of establishing the AI as a "knower" and a deliberate actor.

The curse of knowledge is highly visible here. The author, analyzing the text from a linguistic and cognitive perspective, projects human psychological motives onto the machine to explain its outputs. Because human texts that are highly predictable and lack emotion are often written by humans who are being cautious or bureaucratic, the author attributes similar internal dispositions to the AI.

Simultaneously, as agency flows to the AI, it is completely stripped from human actors. The text relies heavily on agentless constructions: "large language models decide," "AI-generated content encounters," "models can be guided." The human developers at firms like OpenAI or Anthropic, who actively constructed the loss functions, scraped the training data, and aligned the models to be "smooth" and "neutral," are rendered entirely invisible. By treating the AI as an autonomous agent that generates "expressions," the discourse makes it unsayable that these models are corporate products engineered to simulate human writing for commercial efficiency.

Metaphor-Driven Trust Inflation​

The text's framing constructs a profound sense of authority and trustworthiness around AI systems by systematically projecting epistemic and conscious states onto statistical processing. By asserting that generative AI has "reached a state of high fluency and basically reasonable logic," the text signals to readers that the system is capable of rational thought. This consciousness language acts as a powerful trust signal. Claiming an AI processes probabilities invites skepticism; claiming it operates with "logic" and "decides" words invites reliance.

The text encourages readers to extend relation-based trust—the kind of trust built on an assumption of shared understanding, sincerity, and logic—to a statistical system incapable of reciprocating it. When the AI fails, this relational framing persists. The text states that the AI "may misconstrue logic or fabricate facts." This intentional explanation manages system failure by casting it as a cognitive misstep by a generally capable knower, rather than a catastrophic breakdown of a system that has no concept of truth to begin with. Blaming the AI for "fabricating" ironically reinforces the belief that the AI usually knows the truth.

Furthermore, by describing the AI's output as having a "neutral emotion," the text constructs the AI as an objective, dispassionate observer. In human contexts, neutrality is often equated with fairness and reliability. By applying this human-trust framework inappropriately to a language model, the text obscures the fact that this "neutrality" is actually a highly engineered compliance mechanism designed by developers. The risks of this metaphor-driven trust are immense. When audiences believe an AI "knows" logic and communicates with "neutral" objectivity, they will blindly delegate critical knowledge work to it, trusting it as an impartial intellectual authority rather than treating it as a stochastic text generator prone to ungrounded hallucinations.

Obscured Mechanics​

This framework holds that commercial incentives, data annotation and human feedback labor, environmental and compute costs, and developers' limited disclosure of training data, objectives, and feedback processes are standing conditions of the technology. The anthropomorphic and agential language in this text systematically conceals these realities.

First, by claiming that the model "prefers" predictability and "decides" the next word, the text obscures the massive socio-technical infrastructure required to construct it. It hides the specific technical reality of reinforcement learning from human feedback (RLHF). Models do not "prefer" safe, smooth text; firms like OpenAI, Google, and Anthropic pay thousands of precarious data annotators to manually rank outputs, heavily penalizing controversy and rewarding standardized politeness. The text's framing renders this exploitative global labor pipeline invisible, attributing its results to the "preferences" of the machine.

Second, the assertion that the AI produces "expressions" and "knowledge writing" completely conceals its dependency on non-consensual data scraping. The model possesses no internal knowledge or ability to express itself; it statistically recombines the uncompensated intellectual and artistic labor of millions of human writers. By framing the AI as an autonomous creator, the text benefits the technology companies by masking the massive copyright and labor extraction at the heart of their business models.

Third, by stating the model "misconstrues logic or fabricates facts," the text conceals the profound technical limitation that language models lack any causal model of the world or grounding in external reality. They are next-token predictors. Framing hallucination as a cognitive mistake ("misconstruing") hides the fact that generating a lie and generating a truth are computationally identical processes for the model.

If these metaphors were replaced with precise mechanistic language—stating that developers aligned the probability distributions to output standardized text based on scraped data—the illusion of the autonomous, logical AI would collapse, revealing a highly engineered, commercially motivated corporate tool.

Context Sensitivity​

The distribution of anthropomorphic language in this text reveals a striking asymmetry between how human readers and AI systems are described. The text is highly sensitive to context, adopting rigorous, mechanical terminology when discussing human neurology, but shifting into fluid, agential metaphors when describing AI.

In the methodological and technical sections focusing on the human brain, the language is deeply mechanistic: it discusses "N400, P600, and late positive component," "semantic entropy," and "fixation-related potentials." The human reader is framed as an information processing system reacting to stimuli. However, when the text shifts to describing the AI's generation process, the anthropomorphism intensifies dramatically. The text claims the AI "decides," "prefers," possesses "reasonable logic," and produces "expressions." This asymmetry leaves the reader with a bizarre picture: the human mind is reduced to electrical spikes and cognitive load metrics, while the mathematical software is elevated to a rational, expressing, preferring agent.

This pattern points to the field's inherited register, where the "curse of knowledge" leads researchers to use shorthand psychological terms for complex algorithmic outcomes. Anthropomorphism concentrates most heavily in passages evaluating the AI's capabilities and limitations. When describing capabilities, the language is highly agential (it has "reasonable logic"). Crucially, even when describing limitations, the text maintains this agential framing: the AI "misconstrues" or "fabricates." Rather than reverting to mechanical language to explain failure (e.g., "the model outputs ungrounded token sequences"), the text elevates the failure to a human-like cognitive error.

The effect of this distribution on the reader is profound. The constant juxtaposition of rigorous ERP data with loose claims about AI "preferences" lends unearned scientific credibility to the anthropomorphic metaphors. Readers are positioned to accept the AI's "logic" and "decisions" as literal facts, shielded by the surrounding neuroscience vocabulary. It subtly naturalizes the AI as a conscious entity operating on equal or even superior cognitive footing to the human reader.

Accountability Synthesis​

Accountability Architecture

This section synthesizes the accountability analyses from Task 1, mapping the text's "accountability architecture"—who is named, who is hidden, and who benefits from obscured agency.

An analysis of the text's accountability architecture reveals a complete and systematic displacement of human responsibility onto the AI system. The text's own argument focuses on the cognitive mechanics of how humans read AI text; it aims to improve detection methods for "academic integrity." However, while it implicitly seeks to protect human systems from AI intrusion, its vocabulary consistently places all active agency within the AI itself.

Throughout the text, an absolute "accountability sink" exists around the model. When the text discusses the generation of predictable, uniform text, it is because the "large language models... prefer expressions with high predictability." When the text discusses the generation of false information, it is because "AI-generated content... may misconstrue logic or fabricate facts." The human actors—the corporate developers who designed the algorithms, the annotators who aligned them, and the users who deploy them—are completely unnamed and unacknowledged. The AI acts as a sole agent with no delegator.

This resulting distribution of responsibility serves the structural interests of the AI industry perfectly. By casting the AI as an autonomous actor that "prefers" standard text and occasionally "fabricates" facts, the text diffuses corporate liability. If an AI "misconstrues" something, it is framed as a cognitive glitch by the machine, not a negligent design choice by a company shipping an ungrounded product.

If the human decision-makers were named, the entire discourse would shift. Replacing "the AI prefers predictability" with "developers aligned the model to output standardized text" transforms a statement about machine nature into a critique of corporate design. Replacing "the AI fabricates facts" with "developers deployed a statistical generator without factual grounding" forces questions of liability and negligence. Naming the actors makes it clear that the "smoothness" and "logic" of AI text are not evolutionary leaps by conscious machines, but the calculated results of human engineering and commercial deployment.

Conclusion: What This Analysis Reveals​

The Core Finding

This analysis reveals two dominant, interconnecting anthropomorphic patterns: the "Model as Conscious Decision-Maker" and the "Model as Epistemic Knower." The foundational pattern is the consciousness projection, which establishes the AI as an entity capable of internal states. By claiming the AI "decides" and "prefers," the text lays the groundwork for the second, more dangerous pattern: attributing epistemic states like "logic," "knowledge writing," and the ability to "misconstrue." This is not a simple one-to-one mapping, but a complex analogical structure that elevates statistical token prediction into the realm of rational, human-like thought. The consciousness architecture here is deeply load-bearing; if the premise that the AI "knows" or "decides" collapses, the explanation for its behavior—that it "fabricates" because it gets confused, or outputs standard text because it "prefers" it—disintegrates entirely. Furthermore, the text's vocabulary actively undercuts its own scientific argument. While aiming to rigorously quantify the neural differences between processing human and machine text, its language uncritically projects human cognitive faculties onto the machine, muddying the very boundaries it seeks to clarify.

Mechanism of the Illusion:​

This framework holds that computational systems process rather than know. The rhetorical architecture of the illusion of mind in this text operates through a systematic blurring of human cognitive processing and machine probabilistic generation. The illusion is triggered by the ELIZA effect—the human predisposition to attribute mind to fluent language output—and the text provides the exact vocabulary needed to solidify this bias. The central move occurs when the author, seeking to explain the statistical regularities of the AI's output, abandons mechanistic description and invokes psychological motives, stating the model "prefers" predictability and "fabricates" facts. This is the curse of knowledge at work: understanding the output, the author projects human rationale onto the system. The temporal structure of the text amplifies this. By nesting these agential claims within highly technical, empirical discussions of human EEG responses (N400, P600), the anthropomorphism inherits an unearned aura of scientific objectivity. Readers, already anxious about or marveling at AI capabilities, read that the AI possesses "reasonable logic" and "neutral emotion" in a peer-reviewed context, leading them to accept the illusion of a synthetic mind not as a metaphor, but as a scientifically validated fact.

Material Stakes:​

Categories: Epistemic, Regulatory/Legal, Institutional

The framing of AI as a conscious, logical knower carries severe material consequences across multiple domains. Epistemically, when text generators are framed as engaging in "knowledge writing" and possessing "reasonable logic," users are primed to treat statistical outputs as justified true beliefs. This metaphor-to-belief pipeline directly causes users to bypass verification, leading to the rapid institutional integration of hallucinations into academic, medical, and legal research. The developers benefit from this inflated capability perception, while the public bears the cost of degraded information environments. From a Regulatory/Legal perspective, language that claims models "decide," "prefer," and "fabricate" creates a profound accountability sink. If policymakers accept the framing that an AI "misconstrues logic" on its own, it shields the deploying firms from liability. It shifts the regulatory focus toward "fixing the AI's behavior" rather than penalizing developers for shipping structurally ungrounded, unreliable products. Institutionally, framing the AI as a "knowledge writer" that can "improve the efficiency of information production" accelerates the displacement of human labor. Corporations use this language to justify replacing researchers and writers with software, hiding the reality that they are merely substituting actual understanding with cheap, statistically plausible mimicry.

AI Literacy as Counter-Practice:​

Practicing critical precision requires systematically dismantling the illusion of mind by replacing consciousness verbs with mechanistic realities and restoring human agency to agentless constructions. Reframing "the AI prefers predictability" to "developers aligned the parameters to penalize variation" immediately forces the recognition that the model's behavior is a corporate design choice, not an internal desire. Correcting "the AI fabricates facts" to "the model generates statistically plausible but ungrounded token sequences" strips away the illusion of a cognitive misstep, exposing the fundamental limitation of the technology. Systematic adoption of this precision requires academic journals to enforce strict editorial standards against anthropomorphic shorthand, demanding that researchers describe mechanisms at the level of parameters, probabilities, and training data. However, resistance to this literacy is immense. The AI industry relies heavily on anthropomorphic vocabulary to market text calculators as synthetic minds, driving massive valuations. Furthermore, the inherited register of computer science favors these convenient shorthands. Precision is costly because it demystifies the technology, threatening the promotional narratives of developers and forcing institutions to confront the fact that they are not adopting alien intelligence, but merely purchasing statistical correlation software.

Path Forward​

The discursive ecology surrounding AI is currently fractured, with different communities employing vocabularies that serve their specific priorities. Maintaining the status quo—where researchers and companies freely mix technical jargon with claims that models "think," "prefer," and "misconstrue"—prioritizes accessibility and marketing narrative but guarantees massive public confusion and regulatory paralysis. It allows developers to simultaneously claim their models are intelligent enough to replace workers but autonomous enough to shield the company from liability when they fail.

If the discourse shifts toward mechanistic precision, requiring phrases like "the model generates text statistically associated with..." rather than "the AI knows," the fundamental opacity of the technology becomes highly visible. This approach serves civil society and regulators by stripping away the illusion of mind, making the human labor, data theft, and corporate decision-making behind the models tractable for critique. However, it costs intuitive narrative resonance and demands higher technical literacy from the public.

Conversely, if anthropomorphic language deepens and becomes formalized in law and policy, we risk embedding the assumption of machine consciousness into our social infrastructure. This future makes it nearly impossible to hold corporations accountable, as the AI itself becomes the locus of legal and ethical responsibility.

To support a more precise discourse, institutional changes are required: funding agencies could require rigorous mechanistic translation in grants, and regulatory frameworks could mandate transparency about how companies describe their systems' capabilities. Ultimately, the vocabulary we choose will dictate what we can see; demanding precision forces us to look past the illusion of the machine and directly at the human architectures of power and profit operating behind it.

Extended Processing Summary​

Gemini's Processing Steps
What you are about to read…

Gemini refers to the text below as "thought summaries." This is an overt consciousness projection because 'intentions' are hallmarks of a conscious mind that 'knows' what it is doing and why. The concealed mechanistic process is probabilistic text generation. Treat this as just another rhetorical artifact—a way of making the model's processing legible. When the model writes "I should double check that...", it is mathematically altering its own probability distribution for the next token to be more accurate. The "I" is a functional gear in the machine, not a self-reference. This language is a deliberate interface design choice, not a biological reality. The developers could have programmed the model to output passive logs. Instead, they chose to simulate a narrator.

Defining the Task

I'm now zeroing in on the core objective: a critical discourse analysis of the PDF, guided by the JSON schema. My focus is on ensuring a rigorous, schema-compliant examination of linguistic features and brain cognitive response differences between human and AI-generated texts.

Assessing the Source

I've determined the input text is peer-reviewed research, likely academic, authored by Hanhui Sheng. The core argument highlights reduced semantic entropy and syntactic variation in AI-generated text, leading to shallower cognitive processing in humans, as indicated by eye-tracking and EEG data. While descriptive, the paper implies developers may be responsible.

Interpreting Implications

I'm now diving into the implications of the research, particularly the suggested applications. I see the findings as having two key implications, these applications: large-model optimization to improve human-like text generation, and AI text detection for academic integrity. I'm noting the attribution of generation choices to underlying models and algorithms.

Analyzing Responsibility Shifts

I'm now examining how the text handles responsibility. While it mentions potential AI-related issues, it frames these as cognitive or mechanical. I've noted a lack of explicit developer accountability, focusing on model agency in generation. I'm especially focused on identifying metaphor and anthropomorphism. The research recommends improving detection systems and optimizing models. I am also currently working on a metaphor and anthropomorphism audit. The first instance highlights the model's 'decision-making' capacity.

Uncovering Implied Agency

I'm now zeroing in on instances of implied AI agency within the text. Specifically, I'm examining where large language models are presented as active agents, such as when they "misconstrue logic" or "fabricate facts." I'm identifying how these phrasings, unacknowledged as projections, shape our perception of the AI's capabilities and culpability, including expressions and their "human qualities".

Unpacking Obscured Agency

I'm currently delving into specific examples where the text subtly obscures developer responsibility. Instances 4 and 5 (now split into 5 and 6) show that the text casts AI as an emotional container and a grammatical architect. I'm focusing on the projection of human attributes onto the models. I'm also auditing the attribution of intentional states to the models, even when these may be a mere projection of human understanding.

Identifying Passive Agency

I am now closely analyzing more passive attributions of agency. In instance 7, a hyperparameter "regulates" creativity, a direct, though unacknowledged, attribution of an intentional state. I'm also examining instance 8, where AI "produces" text, framing it as a considerate and predictable author. I'm dissecting how these passive, yet agency-laden, phrases subtly shape our perceptions of AI's capabilities and role in text generation.

Evaluating Attributions

I am now assessing the implicit attributions of agency and intelligence within the text. Specifically, I see how phrases like "high fluency and reasonable logic" frame the AI as a logical thinker, which may be a projection of a human's logical reasoning. I'm also examining the attribution of a "neutral tone," which suggests goal-directed behavior. I'm focusing on those expressions, and I am refining instances to strengthen the mapping and ensure variation. I am also updating the instances to reflect this.

Mapping Attributions

I am now mapping the attributions of agency and intentionality onto probabilistic models. I'm focusing on "decide the next word," which overlays human decision-making on probabilistic models, and I am exploring the phrase "prefer expressions."

Reframing Attributions

I'm now zeroing in on instances of anthropomorphism, specifically in relation to expressions and their "human qualities." I'm also addressing the attribution of intentional states to the models, which may be a projection of a human's logical reasoning. I'm refining the text to remove these attributions. I'm focusing on instances where the AI is framed as having "human qualities".

Synthesizing Observations

I'm now integrating critical observations into a structured five-paragraph synthesis, and I'm ensuring the declared lens is properly used. I'm focusing on the "slippage" from mechanistic descriptions to attributed agency in AI, and I am particularly focused on the effect of metaphor to create an uncritical trust in "reliable" AI. I am also making sure that the role of developers, and technical details of the training process, are fully addressed. I am now finalizing the structure.

Validating Attribute Accuracy

I'm now carefully checking the JSON formatting, ensuring strict adherence to the schema, particularly double quotes and the absence of control characters. I'm rigorously verifying the definitions and enum values for fields like "Mechanism Basis," "Acknowledgment.status," "ActorVisibility.actorVisibility," and "AgencyLevel" to maintain data integrity. I'm focusing on "Intentional-state attribution" within my quote analysis. I'm also addressing the "decide the next word" quote.

Analyzing Probabilities

I am now carefully analyzing the phrases for their implications and ensuring their schema compliance, especially regarding intentional states. For example, "decide the next word" is being reframed as a probabilistic function to align with the underlying mechanisms. I am also working to ensure that the JSON format is strictly valid, and will not contain any unexpected newline characters which might cause parsing errors.

Refining Metaphors

I'm now zeroing in on instances of anthropomorphism, specifically in phrases like "a higher value is more creative," or where AI is said to have "high fluency and basically reasonable logic." I'm also examining attributions of human intent or emotion, such as a "neutral emotion" to the models. I'm actively reframing these to strip away those human projections. I am working on how the AI is described, and this also includes expressions.

Categorizing Analysis Elements

I'm now integrating critical observations into categories for the JSON output. Specifically, I'm defining the framework and assigning relevant labels like agencySlippage, metaphorDrivenTrust, obscuredMechanics, contextSensitivity, and accountabilitySynthesis for Task 5. I'm also ensuring the conclusion aligns with the structure, using patternSummary, mechanismOfIllusion, materialStakes, literacyAsCounterPractice, and pathForward labels. I'm double-checking enums, including "Genetic", "Functional", "Empirical Generalization", "Theoretical", "Intentional", "Dispositional", and "Reason-Based" for Task 3 (explanationTypes.type), and "Documented in source", "Generic to system type", "Reasonable inference", and "Unknown/proprietary" for Task 4 (mechanismBasis). In the new task 0, genre will be "Peer-reviewed research", and author position is "Academic or independent research." The core argument focuses on reduced processing.

Clarifying Mechanism Attribution

I'm now zeroing in on the text's characterization of underlying mechanisms, especially where it attributes human qualities. My focus is on the training process, and how the generation process produces text from those parameters. I'm focusing on instances where the AI appears to "decide," or where "a higher value is more creative." I'm also addressing the attribution of intentional states to the models, removing any human projections to clarify its schema.

Separating Training from Generation

I'm now revising the epistemic corrections in Task 4 to clearly differentiate between the training and generation phases. For the phrase "decide the next word," I'm clarifying it is a probabilistic function. The sampling algorithm selects tokens based on probabilities. I'm also revising "misconstrue logic or fabricate facts" to explain this reflects the model's token prediction based on its parameters. I'm ensuring all corrections avoid anthropomorphism and focus on the mechanistic processes. Developers designed the algorithm.

Revising Schema Compliance

I'm now integrating the epistemic corrections into the analysis. I'm focusing on the examples where the text attributes human-like intentionality to the models. I'm ensuring strict adherence to the schema, particularly in separating the training and generation phases for improved clarity, and I am updating the JSON to reflect the specific attributions.

Refining Agency Attribution

I'm now zeroing in on instances that invoke human qualities, like "high fluency and basically reasonable logic," but I'm revising these to clarify the model's generation process reflects trained parameters. The goal is to ensure the text focuses on statistical replication of language patterns, not an inherent capability of "reason." I'm updating the epistemic corrections and refining the schema.

Revising Intent Statements

I'm now clarifying the intentional state attributions to the models, like "suitable for various kinds of knowledge writing." I'm reframing this to explain that the model outputs text statistically correlated with informational prompts based on trained parameters, removing any implied intentionality. Also I'm focusing on "human feedback" which is really developers. I'm also addressing the "expressions," which the models do not have.

Revising Intent and Attribution

I am now carefully reframing instances of implied AI agency, specifically in relation to the model's ability to "decide" or express emotions. I'm focusing on "suitable for various kinds of knowledge writing" and clarifying that the model's output reflects statistical patterns learned during training, and that human deployers define the domain. I'm also addressing examples where human intention is attributed, such as the AI's "neutral emotion," by clarifying that these are statistical correlations and not expressions.

Analyzing Expression Implication

I'm now zeroing in on instances that invoke human qualities, like "decide the next word," and I'm updating my analysis to clarify that the AI's output reflects trained parameters. The goal is to ensure the text focuses on statistical replication of language patterns, not an inherent capability of "decision-making." This is primarily a refinement in how I'm handling the expression and implications.

Refining Text Analysis

I'm now diving deeper into the nuances of language by analyzing examples like "misconstrue logic or fabricate facts" and "a higher value is more creative." I'm carefully reframing the text to explain that these outputs are statistically correlated to human-defined prompts. My focus is on ensuring these examples focus on statistical replication, not on an inherent capacity of the models. I'm focused on reframing "expressions" to focus on statistical replication.

Refocusing on Reframing

I'm now zeroing in on instances like "a higher value is more creative" and "high fluency," clarifying statistical replication, not inherent capabilities, is the key. I am revising to emphasize the model's training on large datasets. I'm focusing on "expressions" as statistical, trained output, while refining the JSON format, and refining the text analysis.

Framing System Parameters

I'm now integrating the premise that computational systems process rather than know directly into the conclusion. I'm carefully drafting Task 5.3 to explicitly highlight the socio-technical conditions and infrastructure obscured by attributing agency to AI. I'm focusing on the effect of the phrase in conclusion P2 to strengthen my argument, while making certain there are no control characters.

Clarifying Statistical Properties

I'm now zeroing in on instances like "decide the next word," reframing this as statistical sampling based on trained parameters to avoid anthropomorphism. This will be an ongoing task, and I'm updating my analysis and ensuring a clear separation of training from generation. Also I'm focused on removing intentionality from the models' description.

Analysis Provenance

Run ID: 2026-11-01-a-study-on-the-linguistic-features-and-b-metaphor-s9lie3 Raw JSON: 2026-11-01-a-study-on-the-linguistic-features-and-b-metaphor-s9lie3.json Framework: Metaphor Analysis v6.8 Schema Version: 3.4.3 Generated: 2026-11-01T10:08:33.950Z

Discourse Depot © 2025 by TD is licensed under CC BY-NC-SA 4.0