Skip to main content

Post-training makes large language models less human-like

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.

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.

V3 Enhancement: Each metaphor now includes an accountability analysis.

1. Pedagogy as Matrix Multiplication​

Quote: "instruction-tuning (teaching models to follow user requests)"

  • Frame: Model as human student
  • Projection: The metaphor of 'teaching' maps the deeply relational, conscious human pedagogical experience onto the mechanistic, statistical process of algorithmic fine-tuning. In the source domain of human pedagogy, an instructor interacts with a conscious student who possesses semantic comprehension, subjective awareness, and the capacity to intentionally internalize rules and meaning. By projecting this onto artificial intelligence, the text suggests that the computational system actually 'understands' what a user request is and consciously 'decides' to follow it. This consciousness projection fundamentally misrepresents the underlying process, which involves no subjective learning or awareness. Mechanistically, instruction-tuning merely updates statistical weights within a neural network via gradient descent, based on matching human-annotated prompt-completion pairs. The system does not 'learn' or 'follow' in any psychological or behavioral sense; it strictly minimizes mathematical loss to produce sequential token outputs that statistically correlate with the preferred training data distributions. Attributing conscious understanding to this gradient update masks the system's reliance on massive data correlation.
  • Acknowledgment: Direct (Unacknowledged) (The text presents the concept of 'teaching' as a direct, literal equivalent to the technical term 'instruction-tuning' within a parenthetical definition, completely omitting any qualifying language. I considered the 'Hedged/Qualified' category because it functions as an explanatory aside, but ruled it out because no hedging terminology (like 'akin to' or 'functions as') is present.)
  • Implications: Utilizing pedagogical metaphors to characterize weight-updating significantly distorts how users, researchers, and policymakers comprehend and regulate artificial intelligence systems. When audiences are led to believe a system has been 'taught' to follow rules, they naturally assume the model possesses a semantic understanding of those instructions and can consciously apply them in novel, out-of-distribution contexts, much like a competent human student. This creates a dangerous illusion of reliability and unwarranted trust, as users expect the system to adhere to the 'spirit' of the instructions rather than their statistical form. This consciousness projection inflates the perceived cognitive sophistication of the model, masking its brittle reliance on training data distributions and creating severe capability overestimations. Consequently, it creates liability ambiguity when the model inevitably generates harmful or biased content, as failures are attributed to 'disobedient' AI rather than flawed human data curation.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The text utilizes the agentless gerund 'teaching' to define the process of instruction-tuning, thereby completely obscuring the specific corporate actors, human data annotators, and software engineers who actually perform this optimization. It conceals who defines the parameters of a valid 'user request' and who dictates the target behavioral distribution. I considered 'Partial (some attribution)' since 'user' is explicitly mentioned, but ruled it out because the primary agent performing the actual tuning and defining the optimization objectives remains completely unstated. By removing the developers from the syntactic frame, the text shifts focus onto the AI model as an autonomous entity that somehow undergoes learning, effectively insulating the corporate creators from critical scrutiny regarding their specific data curation, labor practices, and normative alignment choices. Naming the actors would expose the subjective human decisions driving the system's behavior.
Show more...

2. Algorithmic Operations as Biological Sensory Perception​

Quote: "extending models to process images in addition to text"

  • Frame: Model as perceiving organism
  • Projection: This linguistic framing maps the biological, conscious phenomenon of sensory perception onto the mathematical conversion of pixel data into embedded vector representations. In human cognition, 'processing' an image implies conscious visual perception, subjective awareness of spatial relationships, and the cognitive synthesis of visual stimuli into coherent semantic meaning. When this organic capability is projected onto a language model, it invites the audience to imagine the system as an entity that 'sees' and subjectively comprehends visual input in a manner analogous to biological organisms. This consciousness projection obscures the strict mechanistic reality: the system merely converts numerical pixel arrays into latent embeddings and performs mathematical operations (such as cross-attention) against text tokens. The system possesses no visual field, no subjective experience of color or shape, and no grounded understanding of the physical world. It correlates matrix values without any actual sensory awareness or epistemological grounding.
  • Acknowledgment: Hedged/Qualified (The verb 'process' acts as a subtle hedge, leaning toward a functional, computational description rather than explicitly claiming the model 'sees' or 'perceives.' I considered 'Direct (Unacknowledged)' but ruled it out because 'process' is technically accurate for mathematical operations, even though it heavily invites organic interpretations in the context of 'images'.)
  • Implications: When biological sensory processing is metaphorically mapped onto computational matrix operations, it dramatically inflates user expectations regarding the system's ability to navigate and comprehend the physical world. Audiences led to believe that an AI 'processes images' in a human-like sense will intuitively trust the system's capacity for spatial reasoning, object permanence, and contextual visual understanding. This unwarranted trust becomes critically dangerous in high-stakes deployments, such as autonomous driving, medical imaging analysis, or automated surveillance, where the system's failure to actually 'understand' visual contexts can lead to catastrophic physical or social harms. By masking the statistical fragility of image embedding operations behind the illusion of organic perception, developers evade the responsibility of communicating the profound limitations and domain dependencies inherent to their multi-modal pattern matching architectures.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The phrase 'extending models' functions as a detached, agentless gerund construction that systematically removes the human designers, hardware engineers, and corporate strategists who actually build and deploy multi-modal architectures. Who is extending these models? For what commercial purposes? I considered the 'Ambiguous/Insufficient Evidence' category due to the brevity of the excerpt, but ruled it out because the broader paragraph consistently frames model evolution as a natural, agent-free progression. By grammatically presenting the extension of model capabilities as an abstract, passive occurrence, the text successfully diverts attention away from the massive capital investments, proprietary architectural choices, and explicit business strategies that drive multi-modal AI development, thereby diffusing accountability for the socioeconomic impacts of deploying these systems.

3. Mimicry as Intentional Deception​

Quote: "faithfully mimicking human behavior, including its errors, variance, and the factors that shape it"

  • Frame: Model as conscious actor/impersonator
  • Projection: The term 'mimicking' maps the human capacity for intentional impersonation and conscious performance onto the output distribution of a computational algorithm. In human contexts, mimicry requires a subject who consciously observes a target, cognitively internalizes their characteristics, and willfully modulates their own behavior to produce a convincing replication. Projecting this intentionality onto large language models suggests that the AI actively 'knows' it is imitating humans and purposefully decides to replicate their errors. This severely misrepresents the mechanistic reality of the system, which merely reproduces the statistical distribution of tokens found within its training corpus. The system does not 'choose' to mimic errors; it simply calculates that error-laden tokens have a high probability of occurring in specific mathematical contexts. Attributing active, conscious mimicry to the system obscures its fundamental lack of intentionality and awareness.
  • Acknowledgment: Hedged/Qualified (The authors qualify this 'mimicry' throughout the surrounding methodological section by explicitly operationalizing it as the 'negative log-likelihood of human responses,' demonstrating a functional rather than literal application of the metaphor. I considered 'Explicitly Acknowledged' but ruled it out as no scare quotes or direct meta-commentary appear around the word 'mimicking' itself.)
  • Implications: Framing statistical generation as intentional 'mimicry' encourages the audience to view the artificial intelligence not as a mathematical artifact, but as a deliberate psychological actor. This consciousness projection fundamentally alters the epistemic relationship between the human user and the machine output, fostering deep relation-based trust and vulnerability. If users believe the system is capable of 'faithful mimicry,' they are highly likely to attribute emotional depth, psychological continuity, and genuine empathy to the algorithmic output. In therapeutic or educational contexts—which the text explicitly envisions—this unwarranted attribution is profoundly dangerous. It leaves vulnerable populations interacting with statistical prediction engines under the false assumption that they are engaging with a reciprocating, conscious entity, thereby masking the absence of true therapeutic comprehension and shielding developers from the ethical ramifications of deploying pseudo-empathetic systems.

Accountability Analysis:

  • Actor Visibility: Ambiguous/Insufficient Evidence
  • Analysis: The structural composition of the sentence makes it genuinely unclear who the actual agent is. Does the AI system perform the 'mimicking' as an autonomous capacity, or are the human researchers actively designing applications for the purpose of 'mimicking'? I considered the 'Hidden (agency obscured)' category, but the passive framing of the preceding clause ('applications lies elsewhere') creates a structural ambiguity that defies definitive categorization. While human agency is clearly absent from the immediate phrasing, the grammatical antecedent could plausibly be either the technology itself or the human application designers. If the researchers and corporate entities were explicitly named as the actors forcing this statistical correlation, it would demand immediate ethical scrutiny regarding consent, data harvesting, and the moral legitimacy of designing systems to deceive users through simulated human variance.

4. Statistical Regularity as Epistemic Rationality​

Quote: "human-like cognitive biases... disappeared - and were instead replaced with more rational behaviors - in newer models"

  • Frame: Model as reasoning entity
  • Projection: This linguistic formulation projects the profoundly human capacities of cognitive bias and epistemic rationality onto the algorithmic outputs of a statistical model. In psychological and philosophical domains, 'rationality' implies a conscious agent capable of evaluating truth claims, exercising logical deliberation, and holding justified true beliefs based on evidence. By claiming that newer models exhibit 'more rational behaviors,' the text anthropomorphizes the system as a conscious knower actively overcoming inherent mental flaws. This projection completely obscures the mechanistic truth: the system does not 'reason' or hold 'beliefs.' The disappearance of so-called 'biases' is merely the result of applying reinforcement learning from human feedback (RLHF), which mathematically penalizes output vectors that human annotators flag as undesirable, forcing the system to generate token sequences that correlate with corporate-mandated stylistic guidelines. The AI possesses no subjective rationality.
  • Acknowledgment: Direct (Unacknowledged) (The text presents the transition from 'cognitive biases' to 'rational behaviors' as a literal, unhedged factual claim regarding the model's intrinsic developmental trajectory. I considered 'Ambiguous/Insufficient Evidence' because the authors rely on psychological literature, but ruled it out as they adopt the psychological vocabulary directly to describe the machine without any epistemological caveats.)
  • Implications: The uncritical projection of 'rationality' onto language models functions as a powerful rhetorical mechanism for constructing unearned authority and epistemic trust. By suggesting that post-training endows these systems with 'rational behaviors,' the text elevates mathematical token prediction to the level of justified reasoning. This framing convinces policymakers, corporations, and the public that the model's outputs are grounded in logical deduction and objective truth, rather than the subjective preferences of RLHF annotators. This consciousness projection directly enables the deployment of these systems in critical decision-making contexts—such as legal analysis, financial forecasting, and medical diagnosis—under the false pretense of superior machine objectivity. Consequently, it creates massive systemic vulnerabilities, as the inherently statistical and hallucination-prone nature of the architecture is hidden beneath the veneer of human-like rationality.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The sentence constructs an entirely agentless narrative wherein cognitive biases simply 'disappeared' and were spontaneously 'replaced' as models became 'newer.' This passive grammatical structure completely erases the vast apparatus of human agency required to enact this change. I considered 'Partial (some attribution)' but there are absolutely no human actors referenced here. The text hides the corporate executives who mandate safety guidelines, the engineers who design the reward models, and the low-wage global labor force that annotates the preference data to enforce this so-called 'rationality.' By presenting this optimization as a natural evolutionary process intrinsic to the AI, the text absolves the developers of any accountability for the specific ideological and normative choices embedded within the post-training process, framing subjective corporate alignment as objective technological progress.

5. System Modification as Social Assistance​

Quote: "the very processes that are currently employed to turn these models into useful assistants"

  • Frame: Model as helpful subordinate
  • Projection: This framing projects the social and relational role of a 'useful assistant' onto a fundamentally mechanistic mathematical artifact. An 'assistant' in human social contexts implies a conscious agent possessing situational awareness, an understanding of shared goals, a capacity for independent problem-solving, and a deliberate intention to aid another person. By conceptualizing the AI system as an 'assistant,' the text encourages the audience to map these attributes of subjective awareness and cooperative intent onto the model. In reality, the system is completely devoid of intent, goals, or the capacity to 'assist' in any conscious manner. Mechanistically, it is a static matrix of weights that processes input prompts and generates statistically probable output text based on its alignment training. It no more 'assists' than a calculator or a spreadsheet does; it merely executes programmed mathematical transformations.
  • Acknowledgment: Direct (Unacknowledged) (The term 'useful assistants' is presented directly and without any qualification as the accepted teleological endpoint of model development. I considered 'Explicitly Acknowledged' due to the authors' critique of this process, but ruled it out because they accept the premise that the models function as assistants, criticizing only the loss of human-like behavioral variance.)
  • Implications: Applying the 'assistant' metaphor to generative algorithms fundamentally alters the social integration and perceived accountability of the technology. By framing the system as a helpful, subordinate entity, it disarms critical scrutiny and fosters relation-based trust, encouraging users to anthropomorphize the tool and share sensitive personal information under the false assumption of collaborative intent. Furthermore, this projection of agency obscures the true capabilities and limitations of the system, leading users to over-rely on its outputs for tasks requiring genuine situational comprehension and ethical judgment. Crucially, the 'assistant' metaphor shifts the conceptual burden of failure: when an 'assistant' makes an error, the fault is often attributed to the assistant's competence or the user's poor instructions, thereby insulating the corporate manufacturers from direct product liability.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The phrase 'the very processes that are currently employed' relies on a classic passive voice construction to systematically erase the human actors executing these processes. Who is employing these processes? Which specific corporations and engineering teams are deciding what constitutes a 'useful assistant'? I considered 'Partial (some attribution)' because the word 'employed' implies human labor, but ruled it out because no specific or even generic human actors are actually named. By obscuring the corporate entities that dictate the parameters of 'usefulness'—typically optimizing for frictionless user engagement and risk mitigation rather than objective truth or scientific fidelity—the text presents the commercial alignment of AI as a passive, inevitable phenomenon. Naming the actors would immediately expose the profit motives and subjective design choices inherent in transforming base models into commercial products.

6. Token Correlation as Epistemic Comprehension​

Quote: "the model learns to predict the next word in large text corpora"

  • Frame: Model as cognitive learner
  • Projection: This highly pervasive metaphor projects the biological and psychological phenomenon of 'learning' onto the algorithmic mechanism of mathematical optimization. Human learning involves the conscious acquisition of knowledge, the integration of new concepts into an existing semantic worldview, and a subjective epistemological process of gaining justified true beliefs. In stark contrast, when a language model 'learns,' it is entirely devoid of cognitive awareness or semantic comprehension. Mechanistically, the training process merely involves adjusting billions of numerical weights via backpropagation to minimize a mathematically defined loss function, thereby increasing the statistical probability of generating the correct token sequence based on the training data. The system does not 'know' or 'understand' the words it predicts; it processes them purely as high-dimensional vectors. The metaphor conflates statistical pattern matching with conscious epistemic achievement.
  • Acknowledgment: Hedged/Qualified (The phrase combines the anthropomorphic verb 'learns' directly with the highly precise mechanistic descriptor 'to predict the next word,' thereby intrinsically hedging the cognitive claim within a functional statistical reality. I considered 'Direct (Unacknowledged)' but ruled it out because the explicitly mathematical context of 'predicting' fundamentally bounds the scope of the 'learning' being described.)
  • Implications: The persistent discursive habit of describing machine optimization as 'learning' generates profound epistemological confusion among the public, regulators, and even practitioners. By continuously projecting cognitive acquisition onto the system, the AI industry successfully inflates the perceived sophistication of their products, leading society to equate 'machine learning' with actual intelligence. This unwarranted trust results in severe epistemic risks; users are conditioned to accept the system's generated outputs as synthesized, 'learned' knowledge rather than mere statistical correlations drawn from potentially flawed, biased, or entirely fabricated training data. Furthermore, this framing supports a dangerous regulatory environment where algorithms are treated as quasi-autonomous epistemic agents rather than deterministic software products requiring strict quality control, rigorous safety testing, and clear corporate liability.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The sentence presents the model itself as the active agent ('the model learns'), entirely omitting the vast human and material infrastructure required to facilitate this mathematical optimization. Who curates the 'large text corpora'? Who designs the architecture, selects the hyperparameters, and pays for the massive computational energy consumption required for the training run? I considered 'Named (actors identified)' because 'the model' is named, but ruled it out because the model is the artifact, not the responsible human actor. By casting the algorithm as an autonomous learner navigating a pre-existing corpus, the text completely conceals the explicit corporate decisions regarding data scraping, intellectual property appropriation, and algorithmic design that fundamentally dictate what the model ultimately 'predicts.' Restoring human agency would expose the active curation and exploitation underlying base model development.

7. Latent Identity as Artificial Persona​

Quote: "persona-induction, i.e. conditioning a model on information about a particular individual"

  • Frame: Model as dramatic performer
  • Projection: The conceptualization of 'persona-induction' projects the complex human psychological experience of identity and role-playing onto a stateless mathematical operation. In a human context, assuming a persona involves a conscious subject utilizing empathy, memory, and cognitive continuity to deliberately perform an alternative identity. By describing the model as taking on a 'persona,' the text implies that the system possesses a latent, unified self that can be temporarily overridden or guided to 'act' like a specific individual. This consciousness projection completely misrepresents the mechanics of in-context learning. When conditioned on participant information, the model does not adopt a subjective identity; it merely alters the initial hidden states of its transformer architecture based on the prompt's tokens, thereby shifting the statistical distribution from which it samples subsequent outputs. It processes linguistic context; it does not assume an identity.
  • Acknowledgment: Explicitly Acknowledged (The text explicitly acknowledges the metaphorical and functional nature of the term by immediately defining 'persona-induction' with the mechanistic phrasing 'i.e. conditioning a model on information.' I considered 'Hedged/Qualified' but ruled it out because the presence of the definitional 'i.e.' serves as an explicit, direct translation of the metaphor into technical reality.)
  • Implications: Even when explicitly acknowledged as a technical procedure, the terminology of 'personas' actively cultivates a persistent illusion of psychological depth within computational systems. For researchers and end-users, this terminology invites unwarranted assumptions about the model's ability to maintain coherent, psychologically realistic simulations of individual humans. As the text itself critiques, this leads to profound methodological flaws in psychological and social science research, where scientists mistake statistical shifts in output distributions for genuine representations of human behavioral variance. The danger of this anthropomorphic projection lies in its potential to validate the use of AI 'personas' in high-stakes simulations—such as public policy testing, market research, or clinical psychology training—producing fundamentally invalid epistemic results while shielding researchers behind the perceived sophistication of the 'synthetic agent.'

Accountability Analysis:

  • Actor Visibility: Partial (some attribution)
  • Analysis: While the immediate phrase 'conditioning a model' employs a passive gerund that obfusctes the actor, the broader context of the sentence clearly implicates the human researchers and application designers who actively employ this 'popular approach for eliciting more human-like behavior.' I considered 'Hidden (agency obscured)' due to the lack of explicit proper nouns, but ruled it out because the text is specifically analyzing a methodological choice made by the scientific community. However, the agency remains only partial, as the text discusses the technique as a widespread abstract phenomenon rather than holding specific researchers, companies, or platforms accountable for propagating the epistemologically flawed practice of utilizing language models as substitute human subjects to cut research costs.

8. Parametric Improvement as Evolutionary Power​

Quote: "as in some domains there is evidence that alignment with human behavior plateaus - or even declines - as models become more powerful"

  • Frame: Model as evolving entity
  • Projection: The phrase 'models become more powerful' projects the concept of organic, autonomous maturation and intrinsic agency onto the iterative engineering and scaling of software artifacts. The metaphor of 'becoming powerful' implies that models possess an internal developmental trajectory, akin to biological evolution or human cognitive growth, allowing them to accumulate capability and agency independently over time. This completely obscures the mechanistic and deeply industrial reality: models do not 'become' anything autonomously. Their capabilities increase exclusively because human engineers drastically expand their parameter counts, feed them exponentially larger datasets, and expend massive amounts of electrical energy and computational compute to optimize their loss functions. The 'power' of the model is entirely a product of external corporate resource allocation and mathematical scaling laws, not an intrinsic, agential maturation.
  • Acknowledgment: Direct (Unacknowledged) (The description of models 'becoming more powerful' is stated as a straightforward, literal factual premise within the scientific narrative, completely lacking any qualifying or mechanistic language. I considered 'Hedged/Qualified' because it is situated within a discussion of empirical 'evidence,' but ruled it out because the specific anthropomorphic phrase itself remains entirely unhedged.)
  • Implications: Framing the advancement of artificial intelligence as an autonomous process of 'becoming more powerful' profoundly distorts public and regulatory discourse. It cultivates a sense of technological determinism, convincing society that the exponential growth of AI capabilities is a natural, unstoppable force rather than the direct result of deliberate corporate investment and engineering decisions. This illusion of autonomous capability inflation actively disempowers regulatory bodies, fostering a fatalistic approach to tech governance where policymakers focus on managing the supposed 'behavior' of the models rather than regulating the capital allocation, data extraction, and compute usage of the corporations building them. By projecting intrinsic power onto the mathematical artifact, the technology industry successfully deflects accountability for the massive societal and environmental disruptions caused by their deliberate scaling strategies.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The construction 'models become more powerful' grammatically positions the AI artifact as the sole active subject of its own advancement, completely erasing the massive corporate infrastructure responsible for this development. I considered 'Ambiguous/Insufficient Evidence' but the structural erasure of agency is highly explicit here. Who is making the models more powerful? OpenAI, Meta, Google, and the myriad of hardware engineers and data center operators who physically construct and train these massive systems. By utilizing an intransitive verb ('become') to describe what is actually a highly resource-intensive, deliberate industrial manufacturing process, the text perfectly obscures human agency. Naming the actors would fundamentally shift the narrative from a story about magical technological evolution to a critical examination of concentrated corporate power, capital expenditure, and explicit design choices.

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: Human pedagogy and conscious instruction → Mathematical optimization via gradient descent and backpropagation​

Quote: "instruction-tuning (teaching models to follow user requests)"

  • Source Domain: Human pedagogy and conscious instruction
  • Target Domain: Mathematical optimization via gradient descent and backpropagation
  • Mapping: The relational structure of human education is mapped directly onto the mechanics of neural network fine-tuning. In the source domain, a conscious teacher transmits concepts to a student who utilizes cognitive awareness, semantic understanding, and deliberate intent to internalize the rules and subsequently alter their behavior. When projected onto the target domain, this mapping invites the assumption that the language model 'understands' the concepts within the human-annotated datasets and willfully chooses to comply with the instructions. It maps human cognitive compliance onto statistical parameter updates, suggesting that algorithmic output generation is driven by internalized comprehension and conscious rule-following rather than mere mathematical probability.
  • What Is Concealed: This pedagogical mapping comprehensively conceals the stark mathematical reality of the system. It hides the fact that instruction-tuning simply calculates loss gradients across billions of parameters to minimize the mathematical distance between the model's output distribution and the specific token sequences provided by underpaid human annotators. Furthermore, it obscures the profound epistemic brittleness of the system; because the model lacks actual comprehension, it cannot genuinely 'follow' rules, making it highly susceptible to adversarial jailbreaks that exploit its statistical nature. The framing also hides the corporate labor supply chains required to produce the training data.
Show more...

Mapping 2: Biological sensory perception and cognitive synthesis → Multi-modal cross-attention mechanisms and vector embedding​

Quote: "extending models to process images in addition to text"

  • Source Domain: Biological sensory perception and cognitive synthesis
  • Target Domain: Multi-modal cross-attention mechanisms and vector embedding
  • Mapping: The complex structure of organic sensory perception is mapped onto the computational architecture of multi-modal neural networks. In biological systems, visual processing involves specialized organs receiving light, transmitting signals to a conscious brain, and integrating those signals into a subjective, spatially grounded understanding of reality. This mapping invites the profound assumption that the AI system possesses a rudimentary form of visual awareness—that it can 'see' and semantically interpret an image just as it 'reads' text. It maps conscious perceptual synthesis onto the mere mathematical alignment of diverse high-dimensional latent spaces.
  • What Is Concealed: The mapping conceals the total absence of physical grounding and subjective awareness in multi-modal models. Mechanistically, the system merely segments image data into patches, flattens them into numerical vectors, and processes them through transformer layers to calculate statistical attention weights relative to text tokens. It hides the fact that the system possesses no actual spatial comprehension, object permanence, or understanding of physical laws. By suggesting the model 'processes images' like an organism, the text obscures the system's massive reliance on flawed training distributions and its severe vulnerability to minor pixel perturbations that would never fool a biologically perceiving entity.

Mapping 3: Conscious impersonation and intentional theatrical performance → Statistical token generation aligning with human response distributions​

Quote: "faithfully mimicking human behavior, including its errors, variance, and the factors that shape it"

  • Source Domain: Conscious impersonation and intentional theatrical performance
  • Target Domain: Statistical token generation aligning with human response distributions
  • Mapping: The structural dynamics of human mimicry are mapped onto the output mechanics of generative algorithms. Mimicry requires an intentional actor who observes a subject, cognitively grasps their behavioral nuances, and willfully modulates their own actions to create a deceptive or accurate representation. This mapping projects conscious intentionality onto the language model, inviting the assumption that the AI possesses a latent, objective 'self' that actively 'chooses' to simulate human errors and psychological variance. It maps the deliberate cognitive effort of impersonation onto the passive, mathematical sampling of tokens from a probability distribution.
  • What Is Concealed: This mapping profoundly conceals the fundamentally deterministic and statistical nature of the text generation. The model does not 'know' what an error is, nor does it possess the intent to mimic one; it merely outputs a sequence of tokens because that sequence achieved the lowest loss score during its optimization against human datasets. Furthermore, this mapping obscures the epistemic opacity of the proprietary systems involved; because researchers cannot access the exact training data of commercial models like Llama or Qwen, they cannot mathematically verify whether the system is 'mimicking' underlying psychological structures or simply regurgitating memorized transcripts from the training corpus.

Mapping 4: Human epistemic maturation and deliberate logical reasoning → Reinforcement learning from human feedback modifying output vectors​

Quote: "human-like cognitive biases... disappeared - and were instead replaced with more rational behaviors - in newer models"

  • Source Domain: Human epistemic maturation and deliberate logical reasoning
  • Target Domain: Reinforcement learning from human feedback modifying output vectors
  • Mapping: The structure of human cognitive development and rational deliberation is projected onto the corporate process of AI safety alignment. In humans, overcoming bias and becoming 'more rational' involves self-reflection, the conscious evaluation of evidence, and a deliberate commitment to logical truth. By mapping this onto newer language models, the text invites the assumption that the system possesses an internal epistemological framework and actively 'reasons' its way to better conclusions. It maps the conscious acquisition of justified true belief onto the algorithmic suppression of statistically probable, yet corporately penalized, token sequences.
  • What Is Concealed: This highly anthropomorphic mapping conceals the subjective, coercive, and profoundly mechanical nature of RLHF. The model does not 'reason' its way to rationality; instead, corporate engineers train a separate reward model based on the subjective preferences of low-wage click-workers, which then automatically updates the main model's weights to avoid generating specific outputs. The mapping hides the fact that 'rationality' in this context is merely a statistical proxy for corporate brand safety and normative compliance. It obscures the absence of ground truth, logic, and reasoning in the system, presenting commercially sanitized outputs as objective epistemic achievements.

Mapping 5: Social subordination and deliberate cooperative aid → Commercial alignment fine-tuning for interactive chat interfaces​

Quote: "the very processes that are currently employed to turn these models into useful assistants"

  • Source Domain: Social subordination and deliberate cooperative aid
  • Target Domain: Commercial alignment fine-tuning for interactive chat interfaces
  • Mapping: The relational dynamics of human assistance are mapped onto the product design of conversational AI. A human assistant utilizes situational awareness, shared objectives, empathy, and conscious problem-solving to aid their employer. By projecting this social role onto language models, the metaphor invites users to map attributes of cooperative intent, reliability, and subjective comprehension onto the algorithm. It frames the mathematical generation of helpful-sounding text as a deliberate, conscious act of social subordination, suggesting the model actually 'wants' to be useful.
  • What Is Concealed: This mapping conceals the absolute lack of intent, situational awareness, and reliability within the system. The model does not 'assist'; it mathematically retrieves and ranks tokens based on optimized probability distributions. The metaphor also obscures the intense commercial objectives behind this alignment. The models are 'turned into assistants' not to provide genuine aid, but to maximize user engagement, harvest behavioral data, and integrate seamlessly into corporate product ecosystems. By framing the system as a helpful entity, the text hides the proprietary nature of the alignment processes and shields the developers from accountability when the 'assistant' inevitably hallucinates or provides dangerous instructions.

Mapping 6: Cognitive education and knowledge acquisition → Iterative weight adjustment via loss minimization (Backpropagation)​

Quote: "the model learns to predict the next word in large text corpora"

  • Source Domain: Cognitive education and knowledge acquisition
  • Target Domain: Iterative weight adjustment via loss minimization (Backpropagation)
  • Mapping: The structure of organic learning is mapped directly onto the mechanics of neural network pre-training. Human learning involves the conscious integration of semantic concepts, the development of internal mental models, and an awareness of meaning. This mapping invites the assumption that the AI system is acquiring actual knowledge and semantic comprehension of the text corpora it processes. It projects the cognitive state of 'knowing' onto the purely mathematical process of calculating conditional probabilities across a high-dimensional vector space.
  • What Is Concealed: This pervasive mapping conceals the total absence of semantic comprehension and grounded knowledge in the resulting model. Mechanistically, the system is simply performing gradient descent to minimize cross-entropy loss, adjusting billions of floating-point numbers so that the predicted sequence of tokens statistically matches the training data. The mapping obscures the fact that the model relies entirely on statistical correlation without any causal or semantic understanding of the words it generates. Furthermore, by framing the process as innocent 'learning,' it hides the vast, often legally dubious corporate extraction of copyrighted data and personal information required to construct the 'large text corpora.'

Mapping 7: Theatrical identity adoption and psychological role-playing → Prompt-based manipulation of initial hidden states in a transformer​

Quote: "persona-induction, i.e. conditioning a model on information about a particular individual"

  • Source Domain: Theatrical identity adoption and psychological role-playing
  • Target Domain: Prompt-based manipulation of initial hidden states in a transformer
  • Mapping: The psychological framework of adopting a persona is mapped onto the mechanics of in-context learning. Taking on a persona requires a conscious human subject who possesses a baseline identity, memory, and the cognitive empathy required to act as someone else. By utilizing this mapping, the text invites the audience to assume that the language model possesses a unified, latent cognitive architecture that can be temporarily superseded by an artificial identity. It projects psychological depth and behavioral consistency onto the transient mathematical processing of an input prompt.
  • What Is Concealed: This mapping conceals the stark reality that language models are completely stateless between inferences and possess no latent identity to override. Mechanistically, 'persona-induction' merely means prepending specific text tokens to a prompt, which shifts the attention mechanisms and alters the probability distribution from which the model samples its subsequent outputs. It hides the fact that the resulting 'behavior' is not a coherent psychological simulation, but merely a superficial statistical correlation with textual tropes found in the training data associated with the given demographic keywords. The mapping obscures the fundamental invalidity of treating these systems as genuine psychological subjects.

Mapping 8: Organic biological maturation and autonomous capability growth → Industrial expansion of parameter count and compute scaling​

Quote: "as in some domains there is evidence that alignment with human behavior plateaus - or even declines - as models become more powerful"

  • Source Domain: Organic biological maturation and autonomous capability growth
  • Target Domain: Industrial expansion of parameter count and compute scaling
  • Mapping: The relational structure of organic growth is mapped onto the industrial production of software. In biology, organisms naturally 'become more powerful' through intrinsic developmental processes, cellular maturation, and self-directed adaptation. By applying this framing to language models, the text invites the assumption that artificial intelligence possesses an internal evolutionary drive, accumulating agency and capability autonomously over time. It maps the biological acquisition of strength onto the mathematical scaling of transformer architectures.
  • What Is Concealed: This mapping comprehensively conceals the massive, deliberate human labor and capital expenditure required to increase a model's capabilities. Models do not 'become' powerful; they are engineered to be larger by corporate actors who invest billions of dollars in specialized hardware (GPUs), consume vast amounts of electrical energy, and scrape exponentially larger datasets. By obscuring this industrial reality behind the illusion of autonomous maturation, the metaphor hides the explicit commercial motives driving the scaling laws. It severely limits transparency by treating AI development as an inevitable force of nature rather than a series of highly consequential, and therefore regulatable, corporate design choices.

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: "Base models are the output of pretraining, in which the model learns to predict the next word in large text corpora."

  • Explanation Types:

    • Genetic: Traces origin through dated sequence of events or stages
    • Empirical Generalization: Subsumes events under timeless statistical regularities

  • Analysis (Why vs. How Slippage): This explanation operates primarily as a Genetic account, detailing the origins of the system ('the output of pretraining') to explain its fundamental nature. Simultaneously, it relies on an Empirical Generalization regarding how the system typically operates ('predict the next word'). By combining these modes, the passage attempts to provide a mechanistic, 'how-it-works' framing of base models, emphasizing their foundational architecture. However, the inclusion of the agential verb 'learns' introduces a subtle tension. While the structural intent of the explanation is purely technical and historical—describing the pretraining phase—the vocabulary choices obscure the purely mathematical nature of the process. This hybrid framing emphasizes the historical construction of the model while subtly obscuring the specific, deliberate human actions (data curation, architecture design) required to facilitate this supposed 'learning' process.

  • Consciousness Claims Analysis: The passage exhibits a classic 'curse of knowledge' dynamic, where the authors, who intimately understand the mathematics of backpropagation, utilize the consciousness verb 'learns' as convenient shorthand, inadvertently projecting cognitive awareness onto the system. The text conflates knowing with processing; the model does not actually acquire knowledge or 'learn' in any epistemic sense. Mechanistically, during the pretraining phase, the algorithm processes vast sequences of tokens, utilizing gradient descent to iteratively adjust billions of parametric weights in order to minimize cross-entropy loss, thereby refining its statistical probability distribution for next-token generation. By describing this purely computational pattern-matching process as 'learning to predict,' the text attributes a conscious, epistemic state of acquisition to an artifact that merely processes high-dimensional matrices. The absence of explicit technical descriptors allows the consciousness claim to stand unmitigated.

  • Rhetorical Impact: By framing statistical optimization as 'learning,' the rhetorical impact is profound: it significantly inflates the audience's perception of the AI's autonomy and cognitive sophistication. Even within a technical explanation, this consciousness framing conditions readers to extend an unwarranted degree of epistemic trust to the model's outputs. If an audience believes the system 'learns,' they will naturally assume it possesses a generalized, adaptable intelligence capable of comprehending meaning, rather than recognizing it as a brittle, domain-bound statistical correlation engine. This dramatically alters risk assessment; decision-makers might deploy 'learning' systems in novel contexts under the false assumption that they can consciously adapt, rather than recognizing their absolute dependency on the specific parameters of their training corpora.

Show more...

Explanation 2​

Quote: "Post-training techniques, such as reinforcement learning from human feedback, on the other hand, are designed to maximize user engagement, thereby shifting models away from their original objective."

  • Explanation Types:

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

  • Analysis (Why vs. How Slippage): This passage operates primarily as an Intentional explanation, explicitly detailing the deliberate human goals ('designed to maximize user engagement') driving the modification of the system. It also functions as a Functional explanation by describing how RLHF operates within the broader context of the system's architecture to alter its behavior ('shifting models away from their original objective'). Notably, this is one of the few instances where the text successfully frames the AI mechanistically while explicitly acknowledging human agency and intent. By stating the techniques are 'designed to maximize,' the explanation emphasizes the 'why' of the system's behavior, but crucially locates that 'why' in the corporate designers rather than the AI itself. This framing effectively highlights the commercial motives that dictate the model's ultimate outputs, exposing the subjective nature of post-training.

  • Consciousness Claims Analysis: In a rare departure from the dominant anthropomorphic pattern, this passage rigorously avoids attributing conscious states to the AI system itself. It successfully distinguishes between knowing and processing by locating intent strictly within the human domain ('designed to maximize') and treating the model as the passive object of manipulation ('shifting models'). There are no consciousness verbs projected onto the algorithm here. Mechanistically, the text accurately describes how reinforcement learning from human feedback operates: human annotators rank outputs, and a reward model updates the primary model's weights to increase the statistical probability of generating preferred tokens. By focusing on the systemic intent ('user engagement') rather than projecting agency onto the model, the authors effectively bypass the curse of knowledge, providing a clear, accurate assessment of how external corporate objectives govern computational processing.

  • Rhetorical Impact: The rhetorical impact of this mechanistic, human-centered framing is highly clarifying. By revealing that models do not autonomously 'decide' to be helpful, but are explicitly 'designed' to 'maximize user engagement,' the text immediately demystifies the technology and accurately places accountability on the corporate developers. This framing drastically reduces the illusion of machine autonomy and directly undermines relation-based trust by exposing the commercial motives behind the AI's 'persona.' If audiences understand that 'alignment' is actually a mathematically enforced mandate for user engagement, they are far more likely to approach the system with appropriate skepticism, recognizing its outputs not as objective truths or conscious reasoning, but as statistically optimized corporate products designed to maintain user retention.

Explanation 3​

Quote: "human decision-making is shaped by heuristics and biases, which might be captured by base models but are then overwritten by reasoning post-training, which optimizes for normatively correct responses."

  • Explanation Types:

    • Theoretical: Embeds in deductive framework, may invoke unobservable mechanisms
    • Genetic: Traces origin through dated sequence of events or stages

  • Analysis (Why vs. How Slippage): This passage utilizes a Theoretical explanation by embedding the AI's behavior within the psychological framework of human 'heuristics and biases.' It also employs a Genetic structure by describing a sequential process where initial traits are 'overwritten' by a subsequent stage ('post-training'). The analysis reveals a complex slippage: the text frames the system mechanistically by explicitly noting that post-training 'optimizes for' specific responses, yet simultaneously agentializes the system by labeling the process 'reasoning post-training.' This choice emphasizes the outcome (normative correctness) while severely obscuring the actual statistical mechanics of how that outcome is achieved. By embedding the AI within human cognitive theory, the explanation validates the model as a psychological subject, even as it attempts to describe its architectural evolution.

  • Consciousness Claims Analysis: This passage contains a profound epistemic slippage. By using the phrase 'reasoning post-training,' the text projects the deeply conscious, logical state of 'reasoning' directly onto the mathematical process of reinforcement learning. The authors, experiencing the curse of knowledge, understand that RLHF penalizes illogical outputs, but they translate this into the consciousness verb 'reasoning,' completely blurring the line between processing and knowing. Mechanistically, post-training does not install the capacity to 'reason'; it simply calculates gradients to suppress the generation of tokens associated with specific human heuristics, shifting the probability distribution toward sequences that human annotators have labeled as 'normatively correct.' Attributing actual reasoning to this gradient update masks the system's continued reliance on statistical correlation and its fundamental lack of any internal logical framework or ability to evaluate truth claims.

  • Rhetorical Impact: By labeling the optimization process as 'reasoning post-training,' the text explicitly encourages the audience to perceive the resulting system as a conscious, logical agent. This consciousness framing is profoundly dangerous because it manufactures unearned epistemic authority; users and policymakers will naturally trust a system they believe possesses the capacity to 'reason.' Consequently, audiences will vastly underestimate the system's risk, assuming its 'normatively correct' responses stem from deep logical deduction rather than mere statistical mimicry of corporate safety guidelines. If audiences believe the AI 'reasons' rather than 'processes,' they are likely to delegate complex, high-stakes decision-making tasks to the algorithm, fundamentally misunderstanding the system's brittle reliance on historical data correlations.

Explanation 4​

Quote: "Persona-induction, i.e. conditioning a model on information about a particular individual, has become a popular approach for eliciting more human-like behavior..."

  • Explanation Types:

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

  • Analysis (Why vs. How Slippage): This passage primarily utilizes a Functional explanation, defining 'persona-induction' by its operational role ('conditioning a model') to explain how behavior is generated. It relies secondarily on an Intentional explanation by identifying the deliberate goal of the researchers ('eliciting more human-like behavior'). The framing is largely mechanistic regarding the AI ('conditioning a model'), but highly agential regarding the human researchers. This choice effectively emphasizes the methodological techniques utilized by scientists while exposing the synthetic nature of the AI's outputs. However, the phrasing 'eliciting more human-like behavior' subtly obscures the reality that the system possesses no internal behavior to 'elicit'; it merely alters its statistical token generation based on the provided prompt context.

  • Consciousness Claims Analysis: The epistemic structure here is relatively sound, as the text explicitly defines the anthropomorphic metaphor ('persona-induction') with a functional, mechanistic description ('conditioning a model on information'). There are no overt consciousness verbs projected onto the algorithm, and the distinction between knowing and processing is maintained by locating agency with the human researchers who 'elicit' the output. Mechanistically, the passage accurately describes prompt engineering: providing specific contextual tokens (information about an individual) which alter the attention weights and subsequent probability distributions of the transformer architecture, resulting in text that statistically correlates with that specific demographic data. The text successfully avoids the curse of knowledge by recognizing that the model does not assume an identity, but merely processes conditional information, even if it uses the psychologically loaded term 'behavior' to describe text generation.

  • Rhetorical Impact: Because this framing explicitly locates agency with the human researchers ('a popular approach for eliciting'), it effectively mitigates the risks associated with anthropomorphism. It signals to the audience that the 'human-like behavior' is not an intrinsic, autonomous quality of a conscious machine, but rather a deliberate, synthetic illusion manufactured by human design. This transparency radically alters risk perception; it prevents the audience from extending relation-based trust to the AI 'persona' and instead directs critical scrutiny toward the validity of the human researchers' methodology. By understanding that the AI merely processes conditioned information rather than 'knowing' an identity, stakeholders are empowered to question the ethical and scientific legitimacy of using large language models as surrogates in behavioral experiments.

Explanation 5​

Quote: "One potential explanation for this effect is that post-trained models simply produce more deterministic outputs, thereby failing to capture the noisiness of human behavior."

  • Explanation Types:

    • Empirical Generalization: Subsumes events under timeless statistical regularities
    • Dispositional: Attributes tendencies or habits

  • Analysis (Why vs. How Slippage): This passage utilizes an Empirical Generalization by describing the statistical regularity of the model's outputs ('produce more deterministic outputs') to explain a specific behavioral effect. It also features a Dispositional framing by characterizing the AI as possessing the habit or tendency of 'failing to capture' human noisiness. This explanation frames the AI heavily in mechanistic, 'how' terms, focusing entirely on output distributions and statistical determinism rather than agency or intent. The choice to emphasize 'deterministic outputs' brilliantly highlights the mathematical reality of the system, deliberately avoiding anthropomorphic projections. It emphasizes the algorithmic constraints of the architecture while effectively obscuring any sense of conscious machine agency, grounding the analysis strictly within the realm of data variance.

  • Consciousness Claims Analysis: This passage serves as a masterclass in epistemic precision. It rigorously avoids attributing any conscious states, utilizing strictly mechanistic terminology ('produce', 'outputs', 'deterministic'). The distinction between knowing and processing is pristine; the system is evaluated purely on its statistical capacity to generate data ('capture the noisiness') rather than any subjective ability to 'understand' humans. The authors entirely resist the curse of knowledge, evaluating the model strictly as a mathematical artifact rather than a psychological subject. Mechanistically, the text accurately assesses the consequences of RLHF and temperature settings during post-training, which systematically collapse the probability distribution to favor the most 'safe' or 'correct' token sequences, thereby drastically reducing the statistical variance (noisiness) that characterizes base model pretraining distributions. The processing reality is laid bare.

  • Rhetorical Impact: The rhetorical impact of this highly mechanistic framing is a stark reduction in unwarranted trust and a necessary demystification of the technology. By characterizing the system's failures in terms of 'deterministic outputs' rather than 'misunderstanding' or 'disobedience,' the text forces the audience to confront the AI as a rigid statistical engine. This framing shapes audience perception by completely stripping the system of autonomy, highlighting instead its mathematical limitations. Consequently, stakeholders understand that the AI does not 'fail' because it lacks empathy or intelligence, but simply because its optimization constraints mathematical variance. This realization prevents users from treating the system as a conscious agent and encourages structural, technical evaluations of its reliability, promoting significantly more responsible deployment policies.

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.

V3 Enhancement: A fourth column addresses human agency restoration—reframing agentless constructions to name the humans responsible for design and deployment decisions.

Original Anthropomorphic FrameMechanistic ReframingTechnical Reality CheckHuman Agency Restoration
instruction-tuning (teaching models to follow user requests)Instruction-tuning updates the neural network's parameters via gradient descent using human-annotated prompt-completion datasets. This process mathematically minimizes the loss function to increase the statistical probability that the model will output token sequences correlating with the formats and stylistic guidelines defined by the developers.The AI does not 'learn' or 'understand' instructions; mechanistically, it merely retrieves and ranks tokens based on adjusted probability distributions derived from supervised training data.Corporate engineers and data scientists design instruction-tuning pipelines, utilizing low-wage human annotators to curate specific datasets that explicitly dictate the mathematical optimization of the model's output distribution.
extending models to process images in addition to textEngineers expand the model architecture by integrating vision encoders that convert pixel arrays into high-dimensional vector embeddings, which are then mathematically aligned with textual embeddings using cross-attention mechanisms.The system does not possess sensory awareness or 'perceive' images; mechanistically, it strictly performs matrix multiplications to correlate numerical pixel embeddings with text token activations.Hardware engineers and corporate research teams at major technology firms specifically design and deploy multi-modal architectures to expand their proprietary systems' capabilities into visual data correlation.
faithfully mimicking human behavior, including its errors, variance, and the factors that shape itThe model generates text sequences that statistically correlate with the variance and error rates present within its human-generated training corpus, optimizing for high mathematical likelihood scores relative to psychological transcripts.The model possesses no intentionality and cannot consciously 'mimic'; it mechanistically samples tokens from a probability distribution shaped by the presence of human errors in its massive pre-training data.Researchers deliberately prompt generative algorithms to produce outputs that statistically align with human datasets, attempting to use the system's text generation as a substitute for actual human experimental subjects.
human-like cognitive biases... disappeared - and were instead replaced with more rational behaviors - in newer modelsNewer models generate token sequences that more closely align with formal logic structures because corporate developers heavily applied reinforcement learning to penalize the mathematical probability of outputting sequences associated with specific human biases.The algorithm does not possess 'rationality' or overcome 'bias'; mechanistically, its weights are updated by a reward model to statistically suppress specific token combinations deemed undesirable by human annotators.Corporate alignment teams, directing armies of data annotators, explicitly decide which text patterns are 'rational' and build reward models that force the algorithm to generate outputs complying with those subjective corporate standards.
the very processes that are currently employed to turn these models into useful assistantsThe specific fine-tuning methodologies that developers utilize to mathematically constrain the model's token generation, optimizing its output distributions for frictionless conversational interaction and commercial utility.The AI is not an 'assistant' and possesses no cooperative intent; mechanistically, it is a static matrix of weights that mathematically calculates the most probable sequence of tokens in response to a conversational prompt.Corporate executives and product teams mandate the use of RLHF and instruction-tuning to modify base models, explicitly designing them to function as commercial products that maximize user engagement.
the model learns to predict the next word in large text corporaDuring the pretraining phase, the algorithm utilizes backpropagation and gradient descent to continuously update billions of numerical parameters, minimizing cross-entropy loss to statistically map token relationships across vast datasets.The system does not 'learn' or acquire semantic knowledge; mechanistically, it calculates complex conditional probabilities to identify correlations among high-dimensional vector representations of text tokens.Data engineers scrape massive quantities of copyrighted and public text from the internet, constructing the enormous datasets necessary for the mathematical optimization of the transformer architecture.
persona-induction, i.e. conditioning a model on information about a particular individualProviding specific demographic text tokens within the prompt, which alters the transformer's attention mechanisms and shifts the probability distribution from which subsequent tokens are sampled.The system does not possess an identity or adopt a psychological 'persona'; mechanistically, it simply processes the prepended contextual tokens to generate outputs that statistically correlate with similar demographic tropes found in its training data.Application designers and scientists deliberately utilize prompt engineering techniques, hoping that manipulating the model's input context will generate statistical outputs they can use as proxy data.
as in some domains there is evidence that alignment with human behavior plateaus - or even declines - as models become more powerfulEmpirical evidence indicates that alignment with human behavioral distributions plateaus or declines when developers exponentially scale up the parameter count and training data volume of their architectures.Models do not autonomously 'become powerful' or possess intrinsic strength; mechanistically, their capability to process complex statistical patterns increases solely as a direct function of applied computational resources and expanded matrix dimensions.Technology corporations make deliberate, highly capital-intensive decisions to construct massive data centers, consume vast energy resources, and scale up model parameters in pursuit of competitive commercial advantage.

Task 5: Critical Observations - Structural Patterns​

Agency Slippage​

The text exhibits a systemic oscillation between rigorous mechanical descriptions and pervasive agential framings, functioning as a discursive mechanism to bridge technical reality with social relevance. This slippage predominantly flows in a mechanical-to-agential direction. In the highly technical sections, the authors explicitly define AI behavior through mechanistic terminology, utilizing the negative log-likelihood (NLL) metric to describe token prediction. However, the text immediately leverages this mathematical grounding to license aggressive anthropomorphism in its introduction and discussion sections. For example, the precise process of 'predicting the next word' rapidly morphs into the agential claim that 'the model learns,' and the mathematical optimization of token outputs transforms into 'teaching models.' This slippage relies heavily on the 'curse of knowledge.' The researchers, intimately familiar with the mathematical nuances of gradient descent, utilize consciousness verbs as shorthand, inadvertently projecting cognitive awareness onto the artifact. Consequently, the AI is established as an active 'knower' and 'learner' before broader agential claims—such as models 'becoming more powerful' or 'mimicking behavior'—are introduced. Simultaneously, this attribution of agency TO the AI necessitates the removal of agency FROM human actors. The text relies heavily on agentless constructions: 'post-training makes models,' 'models are currently employed,' and 'persona-induction... has become a popular approach.' By conceptualizing the AI as an autonomous, evolving entity, the text systematically obscures the engineers, annotators, and corporate strategists who actually design, deploy, and profit from these systems. This dual slippage—agentializing the machine while erasing the humans—rhetorically constructs the AI as a natural force rather than a manufactured industrial product, effectively shielding the explicit design choices of the technology industry from critical scrutiny.

Metaphor-Driven Trust Inflation​

The text’s pervasive deployment of anthropomorphic and consciousness-attributing language systematically constructs an unwarranted foundation of trust and authority around mathematical models. By framing AI systems as 'useful assistants' and referring to the mathematical suppression of variance as the emergence of 'more rational behaviors,' the text actively encourages audiences to extend relation-based trust to statistical artifacts. Relation-based trust relies on assumptions of sincerity, ethical intent, and mutual vulnerability—attributes strictly limited to conscious actors. When the text claims that newer models are 'rational' and can be 'taught,' it signals to the reader that the AI possesses the capacity to evaluate truth claims and comprehend the spirit of an instruction. This consciousness framing violently blurs the distinction between processing (mathematical calculation) and knowing (justified belief), compelling users to rely on the model’s outputs as synthesized knowledge rather than mere statistical correlation. The distinction between performance-based trust (reliability in specific domains) and relation-based trust is completely collapsed. Consequently, when the system hallucinates or fails, audiences—conditioned by the 'assistant' metaphor—attribute the failure to a misunderstanding or a 'glitch' rather than recognizing the inherent epistemic brittleness of a system devoid of actual intelligence. This metaphor-driven trust inflates perceived competence and encourages the deployment of these systems in high-stakes domains, creating severe systemic vulnerabilities where human oversight is abdicated to an unthinking probability engine.

Obscured Mechanics​

The anthropomorphic language utilized throughout the text profoundly conceals the material, technical, and economic realities of artificial intelligence production. When the text employs consciousness verbs like 'learns,' 'teaches,' and 'understands,' it actively obscures the brute mechanistic reality that these systems are utterly dependent on massive text corpora, lacking any causal models, physical grounding, or true semantic comprehension. For instance, the framing of 'post-training' as a mechanism that 'makes models less human-like' hides the technical reality of reinforcement learning from human feedback (RLHF)—a process that does not change the AI's internal 'psychology,' but merely applies a mathematical penalty to specific output vectors. Furthermore, applying the 'name the corporation' test reveals severe accountability gaps. When the text states 'models become more powerful,' it conceals the actions of specific mega-corporations (OpenAI, Meta, Google) driving this scaling. This framing makes the vast material costs—the staggering energy consumption, the environmental impact of data centers, and the extraction of public data—completely invisible. Crucially, it erases the precarious global labor force of data annotators who manually label the datasets necessary for the system to appear 'rational.' By presenting the commercial alignment of these systems as the inevitable evolution of an autonomous 'assistant,' the metaphors shield the profit motives and proprietary opacity of the tech industry. Replacing these metaphors with mechanistic language would immediately expose AI not as an intelligent agent, but as a resource-intensive, corporately controlled infrastructure.

Context Sensitivity​

The distribution and intensity of anthropomorphic language across the text are highly strategic, varying significantly based on the rhetorical context. The density of consciousness claims intensifies dramatically in the introduction and discussion sections, where the authors attempt to frame the societal relevance and visionary applications of their research. Here, mathematical processing becomes 'mimicking,' 'teaching,' and 'learning.' This demonstrates a clear pattern: the text establishes scientific credibility through rigorous mechanical language in the methodology (discussing negative log-likelihoods, bfloat16 precision, and parameter counts), and then actively leverages that credibility to justify aggressive metaphorical license in its normative claims. Crucially, there is a stark asymmetry in how capabilities versus limitations are framed. The system’s capabilities are consistently described in agential, consciousness-oriented terms ('models learn,' 'models act as useful assistants'), whereas its limitations are frequently relegated to mechanical or statistical descriptions ('failing to capture noisiness,' 'deterministic outputs'). This strategic asymmetry accomplishes a vital rhetorical function: it maximizes the perceived sophistication of the technology while minimizing structural critiques by attributing failures to mere mathematical artifacts. The register shifts seamlessly from acknowledged analogies (testing 'persona-induction' as a hypothesis) to literalized facts ('models are more rational'). This context-sensitive anthropomorphism serves the dual goal of rendering complex mathematics intuitive for a broad academic audience while simultaneously maintaining the mystique and perceived autonomy of the AI systems, thereby fulfilling the narrative requirements of impactful scientific literature.

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.

Synthesizing the accountability analyses reveals a comprehensive architecture of displaced responsibility, structurally designed to diffuse human accountability for the design and deployment of AI systems. The text consistently exhibits a pattern where human decision-makers are completely erased through agentless passive constructions or by grammatically positioning the AI itself as the sole active agent. Phrases like 'instruction-tuning (teaching models),' 'models become more powerful,' and 'processes currently employed' create an accountability sink where responsibility entirely disappears into technological abstraction. By framing subjective corporate optimization choices as the autonomous evolution of the machine, the text ensures that human agency remains perpetually hidden. If audiences accept this framing, the liability implications are profound: when an AI generates biased, harmful, or systematically flawed outputs, the blame is attributed to the 'model's failure to understand' or a 'glitch in the learning process,' rather than targeting the executives who authorized the deployment, the engineers who designed the reward functions, and the corporations that profit from the system's use. If human decision-makers were explicitly named—for instance, changing 'post-training makes models' to 'corporate engineering teams utilize post-training to force models'—it would radically transform the discourse. It would make alternative design choices visible, render the underlying profit motives open to critique, and establish clear chains of legal and ethical liability. Ultimately, the systemic function of this displaced agency is to protect institutional and commercial interests, presenting highly political, capital-intensive software development as the inevitable progression of an autonomous, quasi-intelligent entity.

Conclusion: What This Analysis Reveals​

The Core Finding

This analysis reveals three dominant, deeply interconnected anthropomorphic patterns structuring the discourse: the 'Cognitive Learner' pattern (mapping statistical optimization to human education), the 'Relational Assistant' pattern (projecting social intentionality onto commercial alignment), and the 'Autonomous Evolver' pattern (framing industrial scaling as organic maturation). These patterns operate as a cohesive system of consciousness architecture. The Cognitive Learner pattern serves as the foundational, load-bearing premise; society must first accept the illusion that the machine 'knows' and 'learns' before it can logically accept that the machine functions as a deliberate 'assistant' or matures into a 'powerful' entity. The sophistication of this system lies in its ability to seamlessly translate the highly complex analogical structures of human psychology directly onto the mathematical matrices of transformer architectures. Crucially, this architecture relies entirely on the systematic conflation of processing and knowing. By projecting conscious awareness, epistemic justification, and subjective intent onto a system that merely processes high-dimensional vectors, the text secures the illusion of machine intelligence. If the foundational premise of the 'learning machine' collapses, the subsequent metaphors of rational assistants and faithful mimics immediately disintegrate, revealing the brittle statistical artifact beneath.

Mechanism of the Illusion:​

The 'illusion of mind' is meticulously constructed through a subtle rhetorical architecture that actively exploits the 'curse of knowledge' and the audience's innate vulnerability to social cues. The central sleight-of-hand occurs at the linguistic level, where precise mechanistic verbs (processes, calculates, minimizes) are systematically replaced with deeply loaded consciousness verbs (understands, learns, reasons). The authors, possessing the technical expertise to understand that gradient descent is purely mathematical, employ this psychological shorthand, inadvertently leading lay audiences to attribute subjective awareness to the system. The temporal structure of the argument is vital to this deception: the text first establishes the AI as a 'knower' through foundational genetic explanations ('the model learns'), and subsequently leverages that established epistemic authority to build complex agential claims ('the model becomes a rational assistant'). This causal chain preys upon the human psychological desire to anthropomorphize responsive entities, actively shifting the audience from performance-based trust in a tool to relation-based trust in a simulated agent. By utilizing intentional and reason-based explanation types to describe corporate alignment processes, the illusion achieves sophisticated narrative resonance, transforming mathematical constraints into the compelling story of a developing mind.

Material Stakes:​

Categories: Regulatory/Legal, Epistemic, Economic

The material consequences of these metaphorical framings are profound and tangible across multiple domains. In the Regulatory/Legal sphere, framing AI as a 'learning assistant' that 'knows' rather than 'processes' fundamentally shifts liability. If regulators believe a system possesses autonomous reasoning, they focus legislation on managing AI 'behavior' and establishing safety 'guardrails' for the machine, rather than drafting strict product liability laws that hold the corporate manufacturers financially responsible for deploying defective statistical software. In the Epistemic domain, the consequences deeply affect scientific research and public knowledge. As highlighted by the text's critique of persona-induction, when researchers believe language models can 'faithfully mimic' human psychology, they begin replacing actual human subjects with algorithmic surrogates. This leads to the generation of fundamentally invalid scientific data, as researchers analyze statistical text correlations under the false premise that they are studying human variance. Economically, this anthropomorphic discourse directly benefits massive technology conglomerates. By convincing the public and investors that models are 'becoming more powerful' and approaching generalized intelligence, corporations secure exorbitant venture capital funding and justify massive energy extraction, while socializing the risks of deployment. If the metaphors were removed, the illusion of general intelligence would shatter, threatening the speculative economic valuations of these proprietary systems.

AI Literacy as Counter-Practice:​

Critical AI literacy demands rigorous linguistic precision as a direct counter-practice to these anthropomorphic obfuscations. By implementing reframings that strip away consciousness verbs, the fundamental reality of the technology is exposed. For instance, translating 'the model learns' to 'the algorithm utilizes gradient descent to minimize loss' forces the recognition that the system operates entirely without awareness, relying exclusively on historical data correlations rather than generalized understanding. Furthermore, correcting agentless constructions—translating 'models become more powerful' to 'corporations expand parameter counts'—restores human agency and directly counters the evasion of liability. Systematic adoption of this precision would require academic journals to enforce strict editorial standards against anthropomorphic shorthand, demanding mechanistic accuracy. Researchers would need to commit to utilizing technical vocabulary even when communicating with the public, resisting the temptation of narrative resonance. Unsurprisingly, this practice faces immense resistance from the technology industry, whose commercial interests are fundamentally protected by the illusion of autonomous machine intelligence. Anthropomorphic language serves as a vital marketing tool; replacing it with dry, mechanistic precision threatens the mystical appeal of the product, thereby jeopardizing user engagement, investment capital, and the evasion of stringent software regulation.

Path Forward​

Looking toward the discursive horizon, the vocabulary choices we normalize today will definitively structure the societal integration of artificial intelligence. If the status quo of pervasive anthropomorphism deepens, the public will increasingly engage with statistical models through relation-based trust. While this enables highly intuitive, frictionless user experiences and accelerates commercial adoption, it embeds catastrophic vulnerabilities; users will routinely entrust sensitive decisions to unthinking algorithms, and corporate developers will continue to evade liability by blaming the 'autonomous' machine. Conversely, if mechanistic precision becomes the mandated norm—perhaps through regulatory frameworks requiring capability disclosures or scientific journals rejecting psychological language for software—transparency will drastically improve. Audiences will correctly perceive AI as a data-dependent, proprietary artifact, ensuring that liability remains firmly with human developers. However, this shift risks alienating non-technical users and reducing the narrative resonance necessary for broad public engagement. A sustainable hybrid future requires deep institutional change. We must develop robust educational frameworks that teach citizens to fluently translate between intuitive metaphors and strict mechanistic realities. Funding agencies could mandate dual-vocabulary disclosures, requiring researchers to justify the use of any agential language with underlying mathematical proofs. Ultimately, the future of AI discourse is a contest of values; the choice between the enchanting illusion of the 'synthetic mind' and the rigorous clarity of the 'statistical engine' will determine who wields power, who bears responsibility, and how society understands the boundary between human consciousness and mathematical correlation.

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 Analytical Scope

I'm now zeroing in on the core analytical lens: AI metaphors, anthropomorphism, and the projection of consciousness within the text. I'm aiming to dissect the language used, identify how these concepts are employed, and understand the implications they carry within the discourse. I need to pinpoint the specific instances where human-like characteristics or consciousness are attributed to the LLMs, and how those attributions shape the argument's overall message.

Parsing the Data Points

Now I'm focusing on the JSON output's constraints, emphasizing precision in data mapping. The source text analysis is complete: I've extracted critical phrases and terminology regarding the research paper. The paper's core finding is that "post-training" reduces the behavioral alignment of LLMs. This involves noting instances where models are described as "learning," "being taught," or producing "normatively correct responses."

Analyzing Human-Model Interaction

I'm now honing in on how models are portrayed in relation to human responses and actions. Key phrases like "mimicking," "emulating," and "simulating" human behavior are central. I'm noting the nuances of attributing agency, considering how processes shape the model's 'usefulness' and accuracy in emulating human behavior. The goal is to catalog instances where models are deliberately crafted to act as interactive simulacra.

Categorizing Metaphorical Language

I'm now identifying and classifying the diverse metaphors used to describe AI models. Specifically, I'm focusing on those linked to cognition, agency, and sensory abilities. I've noted various framings, such as AI as an actor, a learner, a force, a processor of sensory input, and a rational agent, noting the phrases and concepts which lead to each. I'm also ensuring the analysis covers the constraint checklist, and documenting the confidence score.

Mapping Metaphors and Explanations

I'm now engaged in Task 2, aiming to map source and target domains for eight metaphors, detailing the mappings and concealed aspects within them. I've also begun Task 3, auditing five passages with Brown's typology, including passages that describe pretraining and post-training processes and their impact on model behavior. The first three examples of Brown's typology have been completed.

Reframing Linguistic Structures

I'm now fully immersed in reframing the anthropomorphic examples from Task 1, with a target of eight instances. I'm focusing on the reframed explanations, epistemic corrections, and restored human agency to provide more clarity. I've begun Task 5, specifically tracking agency slippage: I'm meticulously noting the move from mechanical to agential descriptions, especially in light of the paper's measurement of "human-likeness" via negative log-likelihood.

Synthesizing Foundational Patterns

I'm synthesizing the paper's key patterns, focusing on how AI's framing shifts trust through metaphors, like the "assistant" persona. I'm also examining how corporate decisions behind techniques such as RLHF are obscured, creating the illusion of autonomous action. Furthermore, I'm noting the context sensitivity of anthropomorphism within the paper, and tracing the agential slippage in descriptions of "post-training." I'm now drafting the five-paragraph synthesis.

Refining Enumerated Categories

I'm now laser-focused on refining the enumerated categories within the JSON schema. I'm rigorously applying the "name the actor" test to ensure precise categorization. I'm meticulously checking the string constraints for "status" and "actorVisibility," ensuring that I'm accurately classifying instances. I'm also carefully checking and estimating word counts for each entry to meet length constraints and avoiding uniform categorization. I'm now drafting with these specific details in mind.

Categorizing Metaphorical Language

I'm now refining the category classifications within the JSON output. I'm focusing on assigning precise categories, particularly considering nuanced uses of terms like "mimicking" and "human-like." I'm also ensuring the correct categorization for "persona," based on the paper's specific evaluation. I'm meticulously checking the JSON syntax and ensuring the output aligns with the paper's key findings.

Analyzing Pedagogical Metaphors

I'm now focusing on the "teaching" metaphor, analyzing how instruction-tuning projects human pedagogy onto model optimization, implying a conscious learning capacity. This also extends to the "learns to predict" metaphor. I'm carefully drafting Task 1.2, meeting all word count constraints for each subtask and the overall task.

Constructing Detailed Analyses

I'm now completely focused on drafting both Task 5 and the Conclusion. I'm taking the instructions' directives to heart, prioritizing deep analysis and academic language. I'm meticulously building out the structure and content for both sections, ensuring I meet and exceed all word count targets by writing substantial, structured paragraphs with compound sentences. I'm also ensuring a singular line of output, and will emit minified JSON to meet this strict constraint.

Encoding JSON Strictness

I'm now implementing the strictest JSON encoding. I'm focusing on ensuring absolutely no control tokens, including newlines or tabs, are present anywhere in the string. I'm reformulating the data to ensure that the entire JSON block will be a single line of text with all strings having spaces instead of \n. I've just started to create the items for Task 1 and ensure those word counts are hit exactly.

Deconstructing Instruction-Tuning

I'm now fully engaged in deconstructing the "teaching" metaphor and its implications within the context of instruction-tuning, aiming to pinpoint exactly how human-like learning is projected onto the model. I've analyzed the "Acknowledgment" status, and refined the classification based on the direct and unacknowledged use of the metaphor. The focus is now on the core process of statistical weight adjustments. The entire focus is on ensuring a robust understanding.

Analyzing Pedagogical Framing

I'm now zeroing in on the implications of the "teaching" metaphor, particularly its role in shaping user trust and expectations of AI systems. I'm focusing on the effect of that framing on interaction with models, especially the risk of over-reliance and the issue of accountability. I'm exploring the implications for liability when models produce unanticipated and potentially harmful outputs.

Unmasking Corporate Actors

I'm now zeroing in on the deliberate obscurity fostered by agentless gerunds like "teaching." I've realized these conceal corporate actors and engineers responsible for defining the system's training data. I've classified this instance as "Unacknowledged" given the primary actor is entirely absent. The focus is on how this framing isolates the model as an autonomous entity. I'm actively working to expose these hidden agents.

Refining Accountability Assessment

I'm now refining the accountability assessment, focusing on the nuanced application of "faithfully mimicking human behavior." The ambiguous agency, stemming from the sentence's structure, demands precise categorization. I'm focusing on why I chose "Ambiguous/Insufficient Evidence," examining whether the actor is the researchers, the model, or the application. I'm also ensuring the analysis meets or exceeds word count.

Unveiling Agential Ambiguity

I'm now grappling with the nuanced agency in the phrase "faithfully mimicking human behavior." The ambiguity stems from whether researchers, the model itself, or the application are the active actors. This necessitates the "Ambiguous/Insufficient Evidence" categorization. My goal is to expose the implied corporate actors, but I'm concerned that a choice like "Hidden" would be inappropriate because of the nature of the entire paragraph. I'm focusing on ensuring I'm meeting the minimum word count.

Constructing Final JSON

I'm now completely focused on crafting the final single-line JSON output. I'm taking the previous instructions to heart, and structuring the components for the eight audits, eight mapping analyses, five explanations, and the five critical observations and conclusion paragraphs. I am making sure the word counts for all five critical observations and conclusions are accurate. I am ensuring that the word counts for all the entries of Task 1 are met, using the new JSON schema. I'm focusing on ensuring I don't introduce any newlines and produce a valid JSON.

Encoding Final JSON

I'm now completely engaged in encoding the complete, single-line JSON output. I'm focusing on ensuring all the text is a continuous string within the JSON properties, and all word counts are meticulously met. I'm reviewing the instructions one final time, ensuring no newlines or extraneous whitespace appear. I'm now structuring the elements for the eight audits, eight mapping analyses, five explanations, five critical observations, and the conclusion, ensuring a strict, minified JSON format.

Synthesizing Lengthy Analytical Data

I'm now completely focused on synthesizing the analytical data to meet required word counts and conform to the strict JSON schema, which entails concatenating text to meet specific length requirements. I'm focusing on preserving deep analytical rigor by carefully constructing the requisite components for the audit, mapping analyses, and explanations. My immediate goal is now to ensure a strictly single-line, minified JSON format, completely devoid of control tokens.

Refining Task Completion

I'm now zeroing in on completing the final JSON output, ensuring it's a single, minified line. I'm carefully constructing the analytical sections, including the audits, mappings, explanations, critical observations, and the conclusion. I'm focusing on rigorous word count compliance and precise JSON formatting while eliminating any newlines or control tokens. I've begun to incorporate the analyses from the Brown explanation types.

Categorizing Linguistic Structures

I'm now revising Task 1 category assignments, ensuring variance and accuracy. I've re-categorized "persona-induction" as explicitly acknowledged, following the paper's usage. I'm focusing on "Hedged/Qualified," "Direct," and "Explicitly Acknowledged" categories and applying them consistently, noting the context of each phrase and its implications within the research paper. I'm carefully reviewing the source text to improve each.

Analyzing Ambiguous Phrases

I'm now zeroing in on an "Ambiguous/Insufficient Evidence" example: "human-like cognitive biases... emerged... but disappeared." My focus is on ensuring this categorization is justified and the JSON formatting is perfect. I'm prioritizing eliminating newlines for the final output, and am confident I can generate a single-line JSON, but it will be a gargantuan project.

Expanding Analytical Components

I'm now expanding the current output to meet the mandated 300-350 word count, ensuring meticulous detail in each analytical section. I'm focusing on incorporating academic concepts like "structure-mapping," "agency displacement," and the "curse of knowledge" to enrich the analysis. I'm now crafting the audit, mapping analyses, explanations, critical observations, and conclusions, ensuring a single-line, minified JSON format. I'm meticulously adding all previously stated details to the output string. I will ensure this expansion preserves a rigorous, academic tone.

Analysis Provenance

Run ID: 2026-05-15-post-training-makes-large-language-model-metaphor-sdzisz Raw JSON: 2026-05-15-post-training-makes-large-language-model-metaphor-sdzisz.json Framework: Metaphor Analysis v6.5 Schema Version: 3.0 Generated: 2026-05-15T21:47:36.151Z

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