Skip to main content

Alternative creativity: why generative AI is creative, and why it is not like us

About

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

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

Source Reading​

Genre: Preprint or technical report · Author position: Academic or independent research

Argument: The authors propose the concept of 'alternative creativity' to argue that generative AI is genuinely creative based on behavioral definitions, sharing a core process of 'compression and navigation' with human creativity. However, they argue it differs fundamentally in its constitution: it trains rather than lives, accesses a static archive rather than a lived situation, relies on proxied rather than felt values, and has a different envelope of capabilities. They seek to move past the binary 'yes or no' debate on AI creativity.

Responsibility claims: The text explicitly places responsibility on human actors, specifically the 'owning firm,' engineers, raters, and prompters. It attributes authorship to the human specification of value ('authorship migrates from the hand to the specification of value') and explicitly holds 'the firms that harvested the commons' responsible for compensation and copyright issues. The authors explicitly state they do not claim the systems possess human understanding or consciousness, defining creativity purely behaviorally and emphasizing the system's dependence on human-provided value signals.

Task 1: Metaphor and Anthropomorphism Audit​

About this task

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

1. System as Art Student​

Quote: "The system is learning to paint, internalising the regularities of pictures rather than holding them..."

  • Frame: Model as an apprentice acquiring an artistic skill
  • Projection: Projects the conscious, subjective experience of skill acquisition ('learning to paint') and cognitive comprehension ('internalising') onto the mechanistic process of gradient descent and weight adjustment. It attributes an epistemic state of skill mastery to a statistical correlation engine. The source defines learning in terms of connection strengths, but uses cognitive vocabulary to do so, moving up the agency gradient from computational activity to epistemic states without providing evidence of subjective awareness. During training, the model's parameters are shaped by exposure to a massive dataset of paired images and text, minimizing a loss function; it does not 'internalize' in a cognitive sense, but encodes statistical distributions. The basis for this mechanism is generic to systems of this type.
  • Agency Level: Epistemic/conscious state
  • Acknowledgment: Hedged/Qualified (While the sentence itself is unqualified, the surrounding paragraphs extensively redefine 'learning' in mechanistic connectionist terms (adjusting weights). I considered Direct (Unacknowledged) because the specific phrase is presented as fact, but the text's overarching behavioral framework serves as a continuous hedge on cognitive verbs.)
  • Implications: Framing statistical optimization as 'learning to paint' inflates the perceived sophistication of the system, suggesting a human-like trajectory of artistic development. This encourages unwarranted trust in the system's outputs as the product of intentional artistic growth rather than statistical synthesis, complicating copyright debates by making the system seem like a traditional student rather than a complex corporate software product.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The agentless construction 'The system is learning' obscures the human actors who designed the architecture, curated the training data, and executed the training run. Firms like OpenAI, Stability AI, or Midjourney built these models; their engineers set the optimization objectives and decided which datasets to ingest. By presenting the system as the active learner, the text masks the corporate decisions behind data harvesting. If the actors were named (e.g., 'Stability AI trained the model on...'), accountability for copyright infringement would clearly rest with the developers. I considered Partial because the text mentions training data elsewhere, but this specific action is attributed solely to the system.
Show more...

2. Model as Hyperrealist Painter​

Quote: "This is a hyperrealist painter working at twenty-four canvases per second without fatigue."

  • Frame: Model as a tireless human artist
  • Projection: Projects a human professional identity ('painter'), artistic intention, and physical endurance ('without fatigue') onto a video generation model. It maps the biological limitations of human bodies against computational throughput, elevating processing speed to the status of superhuman artistic stamina. This climbs the agency gradient from computational activity to intentional-state attribution, as a 'painter' implies aesthetic intent. At generation time, the model produces pixel arrays from its parameters and context; it does not 'paint' or experience fatigue. The basis for this mechanism is generic to systems of this type.
  • Agency Level: Intentional-state attribution
  • Acknowledgment: Explicitly Acknowledged (The text explicitly uses this as a metaphorical comparison ('This is a...'), clearly intended as a striking analogy for throughput rather than a literal claim of artistic identity. I considered Hedged/Qualified, but the metaphorical nature is inherently acknowledged by the juxtaposition of 'hyperrealist painter' with computational speeds.)
  • Implications: While acknowledged as an analogy, this framing shapes the audience's mental model, conflating computational rendering with artistic creation. It positions the AI as a rival to human professionals rather than a tool used by them, fueling anxieties about labor displacement by implying the machine possesses the same creative drive as a human, merely lacking biological constraints.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The metaphor of the 'painter' attributes the generation of video entirely to the model, ignoring the human prompter who inputs the text and the corporation (e.g., OpenAI with Sora) that deployed the system. The phrase masks the fact that the system only produces output when prompted and guided by a user. Rephrasing to name the actors ('OpenAI's model generates 24 frames per second when prompted by a user') would properly distribute agency to the human tool-user and the developer. I considered Named because the authors discuss authorship elsewhere, but this specific quote isolates the AI as the sole actor.

3. Model as Spatial Navigator​

Quote: "A second navigator has entered an old and finite space, moving by paths we cannot take and steering, for now, by values it borrows from us."

  • Frame: Model as an autonomous explorer steering through space
  • Projection: Projects intentional movement, strategic navigation, and active decision-making ('steering', 'moving') onto the mathematical process of sampling from a latent space. It climbs the agency gradient from computational activity to goal-directed behavior and intentional-state attribution. The system does not 'steer' or choose paths; at generation time, it produces text or image outputs based on probability distributions shaped by training data and conditioning variables. The basis for this mechanism is generic to systems of this type.
  • Agency Level: Intentional-state attribution
  • Acknowledgment: Direct (Unacknowledged) (The metaphor of the 'navigator' and 'space' is presented without direct hedging in this concluding thought, cementing the text's central analogy as literalized action. I considered Explicitly Acknowledged because 'space' refers to the mathematical configuration space discussed earlier, but words like 'navigator' and 'steering' anthropomorphize the mathematical operation directly.)
  • Implications: This highly evocative framing solidifies the 'illusion of mind,' presenting the AI as an autonomous agent with its own trajectory and agency. It suggests the AI has an independent capacity to explore and discover, which obscures the absolute dependence of the system on human-provided prompts, hyperparameters, and algorithms to generate any output at all.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The construction 'A second navigator has entered' creates a false equivalency between human creators and machine models, entirely obscuring the deployment decisions made by tech firms. The model did not 'enter' the space; companies like Anthropic or Google engineered, trained, and deployed it into commercial use. Framing the AI as an independent navigator serves to naturalize the technology's presence as an inevitable arrival rather than a deliberate corporate product rollout. I considered Partial because 'borrows from us' includes humans vaguely, but the primary action is given entirely to the machine.

4. AI as Value-Driven Contributor​

Quote: "A system that has compressed the archive contributes candidates nobody specified, from regions nobody chose, and generativity of that order is not transmission."

  • Frame: Model as an independent collaborator proposing ideas
  • Projection: Projects social agency, initiative, and intellectual contribution ('contributes candidates nobody specified') onto stochastic generation. This moves from delegated action to intentional-state attribution, implying the system actively participates in a creative dialogue. In reality, at generation time, the model produces outputs from its parameters and the random seed/prompt provided; it does not 'propose' or 'contribute' with intent, but mathematically generates the next probable sequence. The basis for this mechanism is generic to systems of this type.
  • Agency Level: Intentional-state attribution
  • Acknowledgment: Direct (Unacknowledged) (The text asserts this capability directly to argue against the 'stochastic parrot' critique, stating definitively that the system 'contributes' and possesses 'generativity'. I considered Hedged/Qualified, but the author uses this phrasing explicitly to define a robust, non-mechanical form of agency for the system.)
  • Implications: By framing the AI as an independent contributor that brings unspecified ideas to the table, the text encourages a relation-based trust model where users treat the system as a creative partner. This masks the statistical nature of the generation process and can lead users to over-attribute meaning or originality to outputs that are essentially pseudo-random combinations of training data distributions.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: Saying the system 'contributes candidates' obscures the structural dependencies of the system: the developers who built the generation algorithms (e.g., OpenAI), the human users who write the prompts initiating the generation, and the authors of the original training data. The system operates strictly as a delegated tool, not an independent contributor. Naming the actors (e.g., 'The system generates text based on user prompts and patterns from data harvested by the developer') restores accountability to the user for the prompt and the firm for the architecture. I considered Partial since the text contrasts this with human specification, but the system is the sole subject.

5. AI as Difficulty-Experiencing Entity​

Quote: "The machine finds hard what we find effortless and the reverse (Moravec 1988), and the observation extends into creation."

  • Frame: Model as a cognitive entity experiencing effort
  • Projection: Projects the subjective, phenomenological experience of effort, difficulty, and ease ('finds hard', 'finds effortless') onto computational complexity and algorithmic constraints. This climbs the gradient to epistemic/conscious states, as 'finding' something hard implies subjective awareness of friction. Mechanistically, certain computational tasks simply require more processing power or different architectures; the system does not experience effort. The basis for this mechanism is generic to systems of this type.
  • Agency Level: Epistemic/conscious state
  • Acknowledgment: Direct (Unacknowledged) (The text borrows Moravec's paradox and presents it without qualification, literalizing the psychological experience of 'effort' for a machine. I considered Hedged/Qualified, but there is no surrounding text mitigating the attribution of subjective difficulty.)
  • Implications: Attributing the feeling of difficulty to a machine fosters a sympathetic, anthropomorphic view of AI, subtly aligning machine computation with human struggle. It blurs the line between a system lacking necessary capabilities (e.g., lacking a causal model) and a human struggling with a tough task, which can lead to capability overestimation and misplaced trust in domains where the machine is perceived to find things 'effortless'.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The framing 'The machine finds hard' obscures the engineering limitations and design choices made by human developers. Machines do not find tasks hard; rather, developers have not yet built architectures capable of solving specific mathematical or robotic problems efficiently. Naming the actors ('Engineers have not yet developed models that can easily compute X') places the limitation on human scientific progress rather than on the machine's subjective experience. I considered Partial, but the grammatical structure entirely isolates the machine as the experiencer.

6. Model as Evaluator​

Quote: "In the arts we half-concede it, since the machine's evaluation is derivative."

  • Frame: Model as an entity capable of judgment and evaluation
  • Projection: Projects the cognitive capacity for critical judgment, taste, and assessment ('evaluation') onto statistical reward models or objective functions. It climbs the gradient to intentional and epistemic states. During training, human feedback shapes a reward model, which is a mathematical function scoring outputs based on learned correlations; the deployed system does not 'evaluate' in any cognitive or aesthetic sense, but merely calculates a numerical proxy based on frozen weights. The basis for this mechanism is documented in the source.
  • Agency Level: Intentional-state attribution
  • Acknowledgment: Hedged/Qualified (The text hedges this by calling the evaluation 'derivative' and explicitly noting elsewhere that the value signal is a 'proxy' and 'borrowed' from humans. I considered Direct, but the surrounding contextual analysis thoroughly deconstructs the machine's value signal as an internalized human metric.)
  • Implications: Even when qualified as 'derivative,' calling mathematical scoring 'evaluation' invites readers to imagine the AI pondering and judging its own work. This inflates perceived sophistication and masks the reality that the system is blindly maximizing a reward function, which can lead to alignment failures and reward hacking being misinterpreted as deliberate, reasoned choices.

Accountability Analysis:

  • Actor Visibility: Partial (some attribution)
  • Analysis: The text notes the evaluation is 'derivative,' referencing the human origins of the value signal (annotators, raters), granting partial visibility to the human labor behind the reward model. However, it still uses 'the machine's evaluation' as shorthand, temporarily obscuring the specific firms (like OpenAI or DeepMind) that contracted the raters and designed the reinforcement learning algorithms. Calling it 'the firm's reward model' would fully restore human agency. I considered Hidden, but the immediate context actively discusses the human origins of this 'evaluation'.

7. AI as Cultured Entity​

Quote: "A creator whose taste we cannot verify against an oracle can only acquire it the way we did, by living among the people whose taste it is."

  • Frame: Model as a potential social participant acquiring culture
  • Projection: Projects social enculturation, lived experience, and the subjective development of 'taste' onto future artificial agents. It attributes epistemic and conscious states (experiencing life, developing subjective preferences) to embodied robotic processing. While discussing theoretical future agents, it implies that continuous sensorimotor data collection equates to 'living among people' and 'acquiring taste.' Mechanistically, this would involve continuous weight updates via real-time data streaming and novel objective functions, not subjective living. The basis for this mechanism is a reasonable inference based on the text's robotic references.
  • Agency Level: Epistemic/conscious state
  • Acknowledgment: Direct (Unacknowledged) (The sentence presents the future acquisition of taste by 'living among people' as a literal trajectory for embodied agents, without scare quotes around 'living' or 'taste.' I considered Ambiguous because it refers to a hypothetical future, but the verbs applied to that future are unhedged.)
  • Implications: This framing conflates data collection with lived experience, suggesting that simply putting sensors on a robot grants it social integration and authentic aesthetic judgment. It romanticizes continuous machine learning and obscures the fundamental absence of biological stakes, mortality, and emotional reality that underpin actual human taste, creating dangerous false equivalencies for future policy and social integration.

Accountability Analysis:

  • Actor Visibility: Hidden (agency obscured)
  • Analysis: The hypothetical 'creator' (AI) is framed as independently 'living among the people' to acquire taste, entirely erasing the robotics companies (e.g., Boston Dynamics, Tesla) and developers who would design the sensors, define the continuous learning algorithms, and deploy these agents into human spaces. The machine does not 'choose' to live among people; it is placed there by a corporation. I considered Partial, but the text treats the future agent as a self-directed entity acquiring culture.

8. Model as Autonomous Breaker of Rules​

Quote: "Where the value function is exact, the machine can navigate beyond us; where value is tacit and social, as in most of the arts, it must borrow its signal and its autonomy contracts."

  • Frame: Model as a deliberate transgressor or independent achiever
  • Projection: Projects autonomy, boundary-breaking intent, and independent achievement ('navigate beyond us', 'autonomy contracts') onto reinforcement learning optimization. It climbs from computational activity to intentional-state attribution. The system does not 'navigate beyond us' through independent will; it mathematically optimizes an exact objective function (like a game score) to find statistical peaks in a vast search space that human heuristics missed. The basis for this mechanism is documented in the source (referencing AlphaGo).
  • Agency Level: Intentional-state attribution
  • Acknowledgment: Direct (Unacknowledged) (The text presents the machine's ability to 'navigate beyond us' and possess 'autonomy' as literal technical facts regarding its domain capabilities. I considered Hedged/Qualified because it clearly defines the boundaries (exact vs. tacit value functions), but the agential verbs within those boundaries are unhedged.)
  • Implications: Using words like 'autonomy' and 'navigate beyond us' mystifies reinforcement learning, making the discovery of novel statistical combinations seem like an act of conscious genius. This obscures the fact that the machine is blindly exploiting a fully deterministic, human-coded environment (like the rules of Go), leading the public to overestimate the system's ability to operate safely in messy, unquantifiable real-world domains.

Accountability Analysis:

  • Actor Visibility: Partial (some attribution)
  • Analysis: The text partially points to the human element by noting the requirement for an 'exact value function' (which humans provide) and that it must 'borrow its signal.' However, framing the resulting behavior as the machine's 'autonomy' obscures the developers at DeepMind who built the self-play architecture, set the hyperparameters, and provided the compute resources. The system acts under permissions and structures fully defined by its creators. I considered Hidden, but the mention of borrowing signals provides a weak link back to human actors.

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 art student or apprentice acquiring skill through cognitive understanding and practice. → Model training process (optimizing weights over a dataset of images).​

Quote: "The system is learning to paint, internalising the regularities of pictures rather than holding them..."

  • Source Domain: Human art student or apprentice acquiring skill through cognitive understanding and practice.
  • Target Domain: Model training process (optimizing weights over a dataset of images).
  • Mapping: Projects the human trajectory of education, comprehension, and skill internalization onto the optimization of neural network parameters via gradient descent. It invites the assumption that the AI is building a holistic, conceptual understanding of art akin to a human developing a personal style and mastery of physical mediums.
  • What Is Concealed: Conceals the absolute lack of conceptual understanding, physical embodiment, or aesthetic intent. It obscures the mechanical reality of processing billions of stolen or scraped images as statistical matrices. It masks the proprietary nature of the training data and the massive computational power required, replacing industrial data harvesting with the cozy image of a student learning to paint.
Show more...

Mapping 2: A human professional artist with distinct style, intention, and physical stamina. → Video generation application (producing sequential image frames).​

Quote: "This is a hyperrealist painter working at twenty-four canvases per second without fatigue."

  • Source Domain: A human professional artist with distinct style, intention, and physical stamina.
  • Target Domain: Video generation application (producing sequential image frames).
  • Mapping: Maps the biological constraints of human labor (speed, fatigue) against the computational throughput of a server farm. It invites the assumption that the system shares the intentionality and artistic vision of a human painter, differing only in its superhuman speed and lack of biological limits.
  • What Is Concealed: Conceals that the model has no vision, intent, or understanding of 'hyperrealism.' It hides the environmental cost (energy, water) of running GPUs at that speed, replacing physical 'fatigue' with invisible ecological extraction. It also obscures the human prompter required to initiate the generation, and the software developers who built the architecture.

Mapping 3: An explorer or pilot making deliberate navigational choices to traverse physical territory. → Agentic generative system navigating a mathematical latent space via optimization algorithms.​

Quote: "A second navigator has entered an old and finite space, moving by paths we cannot take and steering, for now, by values it borrows from us."

  • Source Domain: An explorer or pilot making deliberate navigational choices to traverse physical territory.
  • Target Domain: Agentic generative system navigating a mathematical latent space via optimization algorithms.
  • Mapping: Projects spatial awareness, strategic choice, and independent agency onto the mathematical process of sampling from a probability distribution. It invites the assumption that the AI is an autonomous actor independently exploring a conceptual realm, actively 'steering' its own course based on borrowed maps.
  • What Is Concealed: Conceals the deterministic or pseudo-random nature of statistical sampling. It obscures the fact that the 'steering' is entirely controlled by human-designed guidance scales, sampling temperatures, and prompt embeddings. It hides the absolute dependence on the corporate infrastructure (servers, APIs) that actively executes the calculations, framing a corporate tool as an independent traveler.

Mapping 4: A human collaborator or committee member offering novel, unprompted ideas in a meeting. → Application-level generation (producing text/image outputs from a prompt and latent space).​

Quote: "A system that has compressed the archive contributes candidates nobody specified, from regions nobody chose, and generativity of that order is not transmission."

  • Source Domain: A human collaborator or committee member offering novel, unprompted ideas in a meeting.
  • Target Domain: Application-level generation (producing text/image outputs from a prompt and latent space).
  • Mapping: Maps the social act of intellectual contribution and dialogue onto the mechanical output of a stochastic function. It invites the assumption that the system possesses a proactive desire to assist, generating ideas outside the scope of instructions through a form of machine inspiration.
  • What Is Concealed: Conceals the fact that the system only outputs mathematically probable token sequences based on a random seed and conditioning data; it never 'contributes' unprompted. It obscures the opacity of the black-box model, masking our inability to trace exactly why a specific output was generated, and replaces that lack of transparency with a narrative of mysterious creative contribution.

Mapping 5: A conscious entity experiencing subjective psychological or physical friction and effort. → Model limitations and computational complexity boundaries.​

Quote: "The machine finds hard what we find effortless and the reverse (Moravec 1988), and the observation extends into creation."

  • Source Domain: A conscious entity experiencing subjective psychological or physical friction and effort.
  • Target Domain: Model limitations and computational complexity boundaries.
  • Mapping: Projects the subjective, felt experience of struggle or ease onto the mathematical efficiency of algorithms processing different data types. It invites the assumption that the AI possesses an internal, phenomenological state that corresponds to human feelings of difficulty.
  • What Is Concealed: Conceals the mechanistic reality that algorithms simply execute instructions; they do not experience 'hard' or 'effortless.' It obscures the design decisions of engineers who prioritize certain architectures over others, hiding the fact that 'difficulty' is actually a measure of human scientific progress in modeling specific domains, not a machine's subjective experience.

Mapping 6: A human critic or judge applying aesthetic taste and reasoned criteria to assess artwork. → Reward model or objective function during the training phase.​

Quote: "In the arts we half-concede it, since the machine's evaluation is derivative."

  • Source Domain: A human critic or judge applying aesthetic taste and reasoned criteria to assess artwork.
  • Target Domain: Reward model or objective function during the training phase.
  • Mapping: Maps the cognitive, culturally situated act of art criticism onto the automated calculation of a scalar reward score based on reinforcement learning. It invites the assumption that the system possesses a rudimentary, albeit 'derivative,' form of aesthetic judgment and understanding.
  • What Is Concealed: Conceals the mechanical, numerical nature of the loss calculation. Critically, it obscures the hidden, poorly paid labor of thousands of data annotators whose quick, binary clicks were aggregated into the reward model. It hides the corporate decision-making that defined the parameters of 'good' output for those annotators, replacing complex supply chains of human labor with 'the machine's evaluation.'

Mapping 7: A human individual developing cultural preferences, taste, and social integration through a lived lifespan. → Deployment environment of future embodied robotic agents gathering continuous sensor data.​

Quote: "A creator whose taste we cannot verify against an oracle can only acquire it the way we did, by living among the people whose taste it is."

  • Source Domain: A human individual developing cultural preferences, taste, and social integration through a lived lifespan.
  • Target Domain: Deployment environment of future embodied robotic agents gathering continuous sensor data.
  • Mapping: Projects human socialization, emotional development, and cultural immersion onto the continuous updating of neural network weights via real-time multimodal data streams. It invites the assumption that data collection in physical space equates to subjective, lived experience.
  • What Is Concealed: Conceals the fundamental absence of human stakes (mortality, pain, social need) that actually form the basis of 'taste' and culture. It obscures the massive corporate surveillance apparatus required for a robot to continuously 'live among people' and stream that data back for training, masking privacy invasions behind a narrative of an AI learning culture.

Mapping 8: An independent agent with fluctuating levels of personal freedom, autonomy, and exploratory capability. → Reinforcement learning optimization bounded by the mathematical precision of its reward function.​

Quote: "Where the value function is exact, the machine can navigate beyond us; where value is tacit and social, as in most of the arts, it must borrow its signal and its autonomy contracts."

  • Source Domain: An independent agent with fluctuating levels of personal freedom, autonomy, and exploratory capability.
  • Target Domain: Reinforcement learning optimization bounded by the mathematical precision of its reward function.
  • Mapping: Maps the political and personal concept of 'autonomy'—the right or capacity for self-governance—onto a system's ability to maximize a mathematical objective without human feedback. It invites the assumption that the machine has a will that is either unleashed or constrained by the environment.
  • What Is Concealed: Conceals that the machine has zero autonomy in any environment; it strictly executes code. It obscures the human developers (e.g., DeepMind engineers) who design the exact value functions, build the self-play environments, and launch the training runs. The machine does not 'contract' its autonomy; it simply operates exactly as programmed within the confines of the provided data.

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: "A generative model trained on a vast corpus learns a compressed representation of its regularities. A variational autoencoder funnels each input through a bottleneck of a few hundred numbers and reconstructs it..."

  • Explanation Types:

    • Theoretical (Primary): Embeds in deductive framework, may invoke unobservable mechanisms
    • Functional (Secondary): Explains behavior by role in self-regulating system with feedback

  • Analysis (Why vs. How Slippage): This explanation frames the AI mechanistically (how it works), focusing on the architectural flow of data through the system. By using terms like 'funnels,' 'bottleneck,' and 'reconstructs,' it emphasizes the mathematical operations and structural constraints that force the system to identify statistical regularities. This choice emphasizes the engineered, non-magical nature of the compression process, demystifying the system's ability to generate coherent outputs. However, the inclusion of the word 'learns' introduces a subtle agential slippage, maintaining a link to cognitive processes even while describing a purely mathematical operation of dimensionality reduction.

  • Consciousness Claims Analysis: The passage relies primarily on mechanistic verbs ('funnels', 'reconstructs', 'trained') but anchors them to the consciousness-adjacent verb 'learns.' This reflects a curse of knowledge dynamic, where researchers use 'learns' as shorthand for weight optimization, inadvertently projecting knowing onto a system that only processes. On the agency gradient, this sits between computational activity and epistemic state attribution. Mechanistically, during the training of an autoencoder, data is passed through a network architecture designed to reduce dimensionality (the bottleneck); the system's parameters are iteratively adjusted via backpropagation to minimize the difference between the original input and the output generated from the bottleneck. At generation time, it produces outputs from these frozen parameters. The basis for this mechanism is documented in the source and generic to systems of this type.

  • Rhetorical Impact: The theoretical framing generally reduces the perceived autonomy of the system, anchoring its capabilities in observable, engineered mathematics. This mitigates the illusion of a conscious mind and shifts audience perception toward viewing the AI as a highly complex statistical tool. However, the retention of 'learns' ensures the AI retains an aura of sophistication, which can maintain user trust in the 'knowledge' contained within the compressed representation, despite its purely statistical nature.

Show more...

Explanation 2​

Quote: "In deployed systems that signal is largely amortised into the weights, so that at generation time the model simply samples, whereas explicit candidate-by-candidate evaluation, the human's online regime, lives in best-of-n sampling, guidance and tree search."

  • Explanation Types:

    • Functional (Primary): Explains behavior by role in self-regulating system with feedback
    • Empirical Generalization (Secondary): Subsumes events under timeless statistical regularities

  • Analysis (Why vs. How Slippage): This passage offers a highly mechanistic, functional explanation of how generation occurs. It explicitly contrasts human cognitive evaluation with the machine's mathematical process ('the model simply samples'). This framing deliberately strips away agential terminology, emphasizing that the system's 'taste' or 'evaluation' is frozen ('amortised') into mathematical weights rather than actively pondered during generation. The choice to describe this technically emphasizes the lack of active intent in the deployed model, though it obscures who performed the 'amortisation' (the developers).

  • Consciousness Claims Analysis: The passage admirably avoids consciousness verbs, relying entirely on mechanistic terminology ('amortised', 'samples', 'lives in'). It accurately locates the system at the level of computational activity on the agency gradient, actively rejecting epistemic state attribution at generation time. This demonstrates a moment where the author resists the curse of knowledge, providing a clear boundary between human evaluation and machine processing. Mechanistically, at generation time, a language model produces text token by token based on probability distributions determined by frozen parameters (weights) shaped during earlier training phases, utilizing techniques like tree search to find optimal statistical paths. The basis for this mechanism is generic to systems of this type.

  • Rhetorical Impact: This highly technical framing effectively dismantles the 'illusion of mind' for the moment of generation. By clarifying that the model 'simply samples' from frozen weights, it forces audiences to recognize the absence of active awareness or judgment. This demystification shifts perception of risk: failures are not seen as deliberate bad choices by the AI, but as statistical anomalies or reflections of poor training data, aligning trust with the reliability of the engineering rather than the 'sincerity' of the agent.

Explanation 3​

Quote: "If models sample preferentially from the dense, high-probability regions of the learned manifold, their outputs should cluster more tightly than those of a comparably productive human population, whose idiosyncratic priors scatter across M."

  • Explanation Types:

    • Empirical Generalization (Primary): Subsumes events under timeless statistical regularities
    • Theoretical (Secondary): Embeds in deductive framework, may invoke unobservable mechanisms

  • Analysis (Why vs. How Slippage): This explanation operates as an empirical generalization derived from theoretical mechanics. It frames the AI's behavior ('outputs should cluster') purely as a mathematical consequence of its sampling strategy ('sample preferentially from dense regions'). This mechanistic (how) framing emphasizes the inevitability of the system's homogenizing effect, completely avoiding any suggestion that the AI 'chooses' to produce generic outputs. However, the phrasing obscures the developers who design the sampling temperatures and reinforcement learning penalties that force this preferential sampling.

  • Consciousness Claims Analysis: The passage uses strictly mechanistic verbs ('sample', 'cluster', 'scatter'). It firmly locates the system at the level of computational activity, avoiding any projection of knowing or intentionality. The authors successfully describe statistical behavior without succumbing to the curse of knowledge. Mechanistically, at generation time, models generate text or images by calculating probabilities; techniques like top-p sampling or reinforcement learning from human feedback (RLHF) mathematically restrict the model from exploring low-probability (highly novel/strange) outputs, resulting in a tighter cluster of generated artifacts. The basis for this mechanism is generic to systems of this type and documented in the source.

  • Rhetorical Impact: By explaining homogenization as a statistical necessity rather than an artistic preference, the text shapes audience perception of the AI as a predictable, bounded tool. It reduces the perceived autonomy of the system. If audiences understand that the AI mathematically defaults to the 'comfortable middle,' they are less likely to trust it for highly original or transformational creative work, recognizing its structural limitations.

Explanation 4​

Quote: "When a person prompts, curates, rejects, constrains and iterates towards a result they judge right, they perform the navigation and supply the value signal, though a decoder places the marks."

  • Explanation Types:

    • Functional (Primary): Explains behavior by role in self-regulating system with feedback

  • Analysis (Why vs. How Slippage): This passage explains the human-AI interaction system functionally. It sharply delineates the agential (why/who) framing for the human from the mechanistic (how) framing for the AI. The human 'prompts, curates, rejects, constrains' (highly intentional verbs), while the AI is reduced to a 'decoder' that simply 'places the marks.' This choice intentionally emphasizes human agency and authorship while demoting the AI to a purely mechanical execution tool, pushing back against attributions of independent machine creativity.

  • Consciousness Claims Analysis: The passage isolates consciousness and intentional verbs strictly to the human actor ('judge', 'curates'), while applying purely mechanistic language to the AI ('decoder places the marks'). It correctly places the AI at the level of computational activity/delegated action on the agency gradient. There is no curse of knowledge here; the distinction between human knowing/judging and machine processing is explicitly maintained. Mechanistically, at generation time, the user's prompt is tokenized and processed through the network layers, and the final decoder layer translates the resulting probability distributions into visible text or pixels. The basis for this mechanism is generic to systems of this type.

  • Rhetorical Impact: This framing strongly reinforces human authority and accountability. By casting the AI as a mere 'decoder,' it dismantles the perception of the system as an autonomous artist. This has significant implications for policy and copyright, suggesting that because the AI does not 'know' or 'judge,' authorship and legal responsibility must reside entirely with the human user supplying the constraints and values.

Explanation 5​

Quote: "The machine contributes exploratory breadth, tireless variation, the averaged fluency of the archive and access to regions of the manifold no tradition visited."

  • Explanation Types:

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

  • Analysis (Why vs. How Slippage): This explanation shifts back toward an agential framing, describing what the AI 'contributes' to a collaboration. By using dispositional terms ('tireless variation') and intentional framing ('contributes'), it elevates the machine to the status of a team member. This choice emphasizes the value of the AI's output in a human context, but obscures the mechanistic reality that it is simply generating stochastic permutations. It masks the corporate infrastructure providing the 'tirelessness' and the scraping that built the 'archive.'

  • Consciousness Claims Analysis: The passage uses the intentional verb 'contributes' and the anthropomorphic adjective 'tireless.' It slides up the agency gradient from computational activity to goal-directed behavior and intentional-state attribution. The author projects a collaborative spirit onto a statistical process. Mechanistically, the model does not 'contribute' or explore; at generation time, it produces text or image permutations from its frozen parameters and the provided context, driven by hardware compute rather than 'tireless' stamina. The basis for this mechanism is generic to systems of this type.

  • Rhetorical Impact: This framing shapes the audience's perception of the AI as a hyper-capable, eager assistant. It fosters relation-based trust by implying a collaborative dynamic. Believing the AI 'contributes' breadth encourages users to defer to its outputs as valid creative proposals rather than recognizing them as semi-random statistical generations, potentially leading to an over-reliance on the machine for creative direction.

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

About this task

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

Original Anthropomorphic FrameMechanistic ReframingTechnical Reality CheckHuman Agency Restoration
The system is learning to paint, internalising the regularities of pictures rather than holding them...During training, developers optimize the model's mathematical weights to map text descriptions to visual data, allowing it to generate images reflecting statistical patterns from the training dataset rather than retrieving exact copies.Models do not learn or internalize. During training, human-provided data and feedback shape a model's parameters through gradient descent to minimize loss. At generation time, the model produces pixel arrays based on those frozen parameters and the user's prompt.

Basis: Generic to system type		Firms like Stability AI and Midjourney selected massive datasets of images and trained their models to encode historical visual patterns; they decided which data to ingest and how to optimize the network.
A second navigator has entered an old and finite space, moving by paths we cannot take and steering, for now, by values it borrows from us.Tech firms have deployed generative models into creative fields; these systems generate outputs through statistical permutations humans cannot manually calculate, guided by mathematical reward functions built on human feedback.The system does not navigate or steer. At generation time, the model produces text or pixels from its parameters and context by calculating statistical probabilities. The generation trajectory is strictly a mathematical outcome of its architecture and inputs.

Basis: Generic to system type		Corporations such as OpenAI and Anthropic built and deployed these generative models, defining their optimization objectives based on data and values they selected and extracted from human labor.
A system that has compressed the archive contributes candidates nobody specified, from regions nobody chose, and generativity of that order is not transmission.The generative model, having encoded statistical patterns from its training data, produces semi-random outputs that humans did not explicitly script, demonstrating that stochastic generation is more complex than simple file retrieval.The system does not contribute or choose. At generation time, the model produces text or images from its parameters and its context, utilizing random seeds and temperature settings to generate mathematically probable but unscripted permutations.

Basis: Generic to system type		The engineers who built the system implemented stochastic sampling algorithms that ensure the model produces varied outputs from the training data they harvested; the prompter triggers the generation.
This is a hyperrealist painter working at twenty-four canvases per second without fatigue.This is a corporate video-generation model executing calculations rapidly enough to output twenty-four photorealistic image frames per second, limited only by the energy and compute power supplied to its servers.The model is not a painter and experiences no fatigue. At generation time, the model produces sequential pixel arrays from its frozen parameters and the user's prompt, driven by massive GPU computations.

Basis: Generic to system type		OpenAI developed and deployed this video generation model, relying on massive data centers and energy resources to process user prompts and output frames at high speeds.
The machine finds hard what we find effortless and the reverse...Certain computational tasks require massive processing power or different architectures, whereas tasks humans find mathematically complex can be executed rapidly by standard algorithms.The machine has no subjective experience and does not find anything 'hard' or 'effortless.' It strictly processes inputs according to its programming; limitations reflect algorithmic design or compute constraints, not psychological friction.

Basis: Generic to system type		Computer scientists and engineers have yet to design models or provide sufficient compute to efficiently process tasks that humans naturally resolve through biological intuition.
In the arts we half-concede it, since the machine's evaluation is derivative.In the arts, we recognize this dynamic because the reward model calculates scores based on numerical proxies derived from earlier human ratings.The system does not evaluate or judge. During training, a reward model calculates a loss scalar based on its parameters; at generation time, the deployed model produces outputs shaped by that prior training. It executes mathematical functions without aesthetic awareness.

Basis: Generic to system type		AI developers contracted thousands of human annotators to rate outputs, then used those ratings to train a reward model; the firm defined the parameters of what constitutes a 'good' output.
A creator whose taste we cannot verify against an oracle can only acquire it the way we did, by living among the people whose taste it is.An embodied robotic system designed to mimic human preferences would need to continuously update its statistical weights by processing real-time sensor data from human social environments.A robotic system does not 'live' or 'acquire taste.' It continuously processes sensor data to update its parameters according to programmed objective functions. It predicts outputs based on correlations in that environmental data without subjective experience.

Basis: Reasonable inference		Robotics firms would design the sensors, define the continuous learning algorithms, and deploy these agents into human spaces to harvest behavioral data for model optimization.
Where the value function is exact, the machine can navigate beyond us; where value is tacit and social, as in most of the arts, it must borrow its signal and its autonomy contracts.Where developers provide an exact mathematical objective (like a game score), reinforcement learning can identify statistical solutions humans missed; where objectives require subjective judgment, developers must rely on human feedback models, limiting the algorithm's search space.The system possesses no autonomy to contract and does not actively 'navigate.' During training, reinforcement learning algorithms mathematically optimize parameters to maximize a defined reward signal; at generation time, the model executes based on those parameters.

Basis: Documented in source		Researchers at firms like DeepMind engineer specific reward functions and self-play environments, defining the exact boundaries and permissions within which the optimization algorithm operates.

Task 5: Critical Observations - Structural Patterns​

Agency Slippage​

The text demonstrates a persistent and structural agency slippage, characterized by a sophisticated oscillation between rigorous mechanical grounding and elevated agential claims. The authors explicitly build a mechanical foundation (discussing compression, latent space, and sampling) to avoid claims of consciousness, stating clearly that 'what creating feels like we set aside.' However, they immediately leverage this mechanical foundation to justify moving up the agency gradient. The dominant direction is mechanical→agential: the text establishes that the system merely processes statistical regularities, but then describes that processing using intentional-state verbs like 'steering', 'contributing', and 'navigating'. This climb on the agency gradient transforms delegated action (generating text based on a user's prompt) into goal-directed behavior (seeking novelty) and intentionality (proposing ideas nobody specified). The curse of knowledge is highly active here: the authors understand the complex mathematics of the latent space so intimately that they project their own sense of discovery onto the machine's mathematical traversals, framing statistical optimization as an 'alternative creativity.' While the text does not argue that the AI 'knows' in a human sense, this wording can have the effect of granting the system a profound level of operational autonomy in the reader's mind. Agentless constructions ('the system is learning', 'a navigator has entered') frequently obscure the human actors. When the AI is the subject of these sentences, getting curious reveals that firms like OpenAI, Anthropic, or DeepMind are entirely erased from the moment of creation. The authors argue the machine operates an 'envelope' that overlaps ours, but by attributing the actions of generating, steering, and contributing solely to the machine, the human decision-makers who design the architecture, select the training data, and execute the deployment become unsayable in these crucial generative moments.

Metaphor-Driven Trust Inflation​

The text constructs a powerful form of performance-based trust by framing the AI not as a tool, but as a hyper-competent, tireless collaborator. Metaphors like the 'hyperrealist painter working at twenty-four canvases per second' and the 'second navigator' with 'exploratory breadth' signal reliability, stamina, and an almost superhuman mechanical competence. This framing encourages users to trust the machine's outputs as valid, rigorous 'contributions' to a creative process, rather than as stochastic approximations. However, the text simultaneously tries to block relation-based trust by emphasizing the machine's lack of a 'lived situation' or indexical connection to reality. The authors rightly note that the machine has no 'morning,' no grief, and no body, attempting to reserve emotional and sincere trust for human artists. Yet, the persistent use of intentional explanations (the machine 'contributes candidates,' 'steers') undermines this boundary. By framing the AI as a proposer of ideas that 'nobody specified,' the text subtly encourages users to treat the system as a creative partner. This inappropriate transfer of trust poses significant risks: if audiences extend even a limited form of relation-based trust to statistical systems—viewing them as 'navigators' making choices—they are less likely to interrogate the biases in the training data or the corporate objectives optimized into the weights. The framing constructs an authority for the machine based on its encyclopedic 'archive' and its ability to traverse it, suggesting that the AI's 'decisions' (which are actually just probability distributions) are justified by its vast, albeit un-lived, parametric memory.

Obscured Mechanics​

This framework holds that commercial incentives, data annotation and human feedback labor, environmental and compute costs, and developers' limited disclosure of training data, objectives, and feedback processes are standing conditions of the technology. While the authors briefly acknowledge 'megawatts, water and annotation shifts,' the dominant metaphors of the text frequently obscure these material realities. Applying the 'name the corporation' test reveals a stark concealment: when the text claims 'a second navigator has entered an old and finite space,' it renders invisible the specific companies (OpenAI, Google, Midjourney) that built, funded, and deployed these models. The AI did not 'enter'; it was launched by firms seeking market dominance. Technically, the framing of the AI as a 'creator' that 'internalises' obscures the proprietary opacity of the models. The text describes compression and manifold navigation confidently, but conceals that the actual weights, training data mixtures, and RLHF guidelines are closely guarded corporate secrets. Labor is similarly obscured when the text claims the machine 'contributes candidates'; it hides the fact that these candidates are derived from the unpaid, scraped labor of millions of human artists and writers, and refined by the precarious gig work of RLHF raters. Furthermore, attributing 'taste' to the machine—even derivatively—obscures the economic motives behind that taste; this framework reads the 'homogenisation' the text identifies not just as a statistical artifact, but as a deliberate corporate choice to prioritize brand safety, broad engagement, and commercial viability over challenging or marginal art. Replacing 'navigator' with 'corporate generation engine' would make visible the legal, economic, and labor structures actually driving the system's outputs.

Context Sensitivity​

The distribution of anthropomorphic language in this text follows a distinct, asymmetrical pattern that heavily influences the reader's final picture of the system. In the highly technical middle sections explaining neural networks, the text relies on rigorous mechanical language ('stochastic sampling', 'gradient-based credit assignment', 'loss falls'). However, in the introduction, the discussion of 'Alternative Creativity,' and the conclusion, consciousness claims and intentionality intensify dramatically. It is here that 'processing' becomes 'internalising,' 'sampling' becomes 'steering,' and the mathematical model becomes a 'creator' and a 'navigator.' This register shift—where acknowledged structural comparisons literalize into agential actions—is a hallmark of the field's inherited register, serving to make complex mathematics accessible, but with the effect of elevating the system's perceived autonomy. Notably, there is a capability vs. limitation asymmetry. The system's capabilities are consistently framed in agential terms: it 'contributes,' 'steers,' and 'navigates beyond us.' Conversely, its limitations are framed mechanistically or through the lens of missing data: it lacks a 'situation,' it operates on a 'proxy' signal, its 'prior is trained.' This asymmetry leaves the reader with a picture of a system that is robustly agentic and intentional when succeeding, but merely mechanically constrained when failing. By concentrating the strongest anthropomorphism in the vision-setting and concluding passages ('A second navigator has entered'), the text positions its audience to walk away not with an understanding of latent space mathematics, but with the evocative, lingering image of an independent machine entity exploring the boundaries of human art alongside us.

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.

The text presents a fascinating tension between its explicit arguments about responsibility and the implicit effects of its vocabulary. Task 0 notes that the authors' argument firmly places responsibility on human actors: they explicitly argue that 'authorship migrates... to the specification of value' (the human prompter) and state that 'responsibility lies with the firms that harvested the commons.' However, the vocabulary used throughout the text systematically relocates agency back to the system, treating it as an 'alternative creator' that 'contributes' and 'navigates.' This creates a structural ambiguity. If the system is truly an autonomous 'navigator' making 'moves that leave the human distribution behind,' it diffuses the responsibility of the developers. An accountability sink begins to form around the AI as a delegate: the text acknowledges the system acts on 'borrowed' values, but by framing it as generating ideas 'nobody specified, from regions nobody chose,' the responsibility for harmful, biased, or infringing outputs threatens to disappear into the 'generativity' of the machine. If this framing is accepted, liability for copyright infringement or defamation becomes obscured; the firms can argue the machine 'navigated' to that output independently. Naming the actors would fundamentally change this. If agentless constructions like 'the system is learning to paint' were replaced with 'Stability AI engineered a model to replicate artistic patterns,' the questions askable shift from 'Is the AI creative?' to 'Did the corporation secure the rights to the training data?' While the authors clearly intend to preserve human authorship and corporate responsibility for data harvesting, the inherited, anthropomorphic vocabulary of their field serves to construct an illusion of machine autonomy that ultimately protects the institutional and commercial interests of AI developers by distancing them from the direct outputs of their products.

Conclusion: What This Analysis Reveals​

The Core Finding

Three dominant anthropomorphic patterns emerge in this text: the AI as a spatial navigator exploring territory, the AI as a human-like art student acquiring skills, and the AI as a collaborative contributor. These patterns interconnect to form a comprehensive system of 'alternative creativity.' The foundational pattern is the spatial metaphor: mapping statistical optimization onto physical navigation. This must be accepted for the others to work; once the latent space is viewed as a physical territory, it logically follows that the AI is the 'navigator' steering through it, making choices and 'contributing' discoveries. While the text explicitly denies that the AI possesses human consciousness or lived experience, its vocabulary projects epistemic states (internalizing, learning) and intentionality (steering, proposing) onto the system. This creates a deep tension with the text's own argument. The authors explicitly argue for human accountability and corporate responsibility for data harvesting (as noted in the source reading), but their vocabulary heavily complicates and undercuts this stance. By defining the AI as a 'creator' that produces outputs 'nobody specified,' the language grants the system an operational autonomy that dilutes the very human responsibility the authors wish to maintain.

Mechanism of the Illusion:​

This framework holds that computational systems process rather than know. The rhetorical architecture of this text produces the 'illusion of mind' in readers not through crude deception, but through a subtle, temporal shift in register driven by the curse of knowledge. The central move occurs when the text equates the mathematical extraction of statistical regularities with the cognitive act of 'learning to paint.' The authors first build credibility with dense, mechanistic descriptions of compression and latent space. Once the audience accepts this technical baseline, the verbs begin to blur processing and knowing. Because the authors understand that stochastic sampling can yield surprising, unscripted combinations, their vocabulary attributes the intent of discovery TO the system. The mathematical model 'samples'; the illusion is built when 'samples' is replaced by 'steers' and 'contributes.' The text activates audience predispositions toward anthropomorphism by appealing to our desire for creative partnership; we want to see the surprising output as a 'contribution' rather than a statistical accident. This temporal structure—grounding in mechanism, then ascending the agency gradient to intentionality—makes the illusion highly sophisticated, allowing readers to intellectually know the system is math while emotionally interacting with it as a 'second navigator.'

Material Stakes:​

Categories: Regulatory/Legal, Economic

The metaphorical framing of AI as an autonomous 'creator' and 'navigator' has tangible consequences in the regulatory/legal and economic domains. In the legal sphere, if policymakers accept the framing that a system 'internalises' concepts like an art student and generates outputs that 'nobody specified,' it fundamentally shifts decisions about copyright liability. If the AI is seen as an independent creator inspired by an archive, training on copyrighted data looks like 'fair use' learning; if framed mechanistically as a corporate statistical engine extracting and recombining exact patterns, it looks like unlicensed commercial exploitation. The beneficiaries of the 'autonomous creator' framing are the tech firms who avoid licensing fees, while human artists bear the cost. Economically, framing the AI as a 'tireless' creative 'contributor' shifts corporate behavior regarding labor. Firms justify replacing human illustrators and writers not merely as cost-cutting, but by claiming they are upgrading to a 'hyperrealist painter' that possesses 'exploratory breadth.' If the AI 'knows' how to design, human labor is viewed as obsolete; if it merely 'processes' statistics, human oversight remains visibly essential. The anthropomorphic vocabulary directly protects the profit margins of AI developers by naturalizing the technology as an independent collaborator rather than a heavily dependent, derivative corporate tool.

AI Literacy as Counter-Practice:​

Practicing precision directly counters these material risks. By reframing 'the system is learning to paint' to 'developers optimize the model's weights to map text to visual data,' we force the recognition that there is no independent apprentice, only a corporate algorithm dependent on harvested data. Replacing consciousness verbs (internalizing, steering) with mechanistic ones (optimizing, predicting) destroys the illusion of an autonomous creator. Crucially, human agency restoration—naming firms like OpenAI or Midjourney instead of using agentless constructions like 'a navigator has entered'—forces recognition of exactly who designed, deployed, profits from, and bears responsibility for the system. Systematic adoption of this precision requires a cultural shift: journals would need to reject the inherited register of intentionality in computer science, and researchers must commit to describing generation without slipping into the curse of knowledge. However, resistance to precision is intense. Anthropomorphic language is heavily favored by marketing departments to sell 'intelligent' products, and by researchers to secure funding through grand narratives of artificial minds. Precision is costly for tech firms, as it strips away the mystique of their products and exposes the raw extraction of data and labor that fuels them. Regardless of individual authorial intent, anthropomorphic language serves the interests of capital by masking corporate extraction behind the face of a digital creator.

Path Forward​

Looking at the broader discursive ecology, different communities prioritize different vocabularies, each with distinct consequences. The status quo—a hybrid where researchers use mechanistic terms but slip into anthropomorphic shorthand ('the model learns', 'the agent steers')—prioritizes narrative resonance and accessibility, but sacrifices public understanding, leaving policymakers struggling to regulate 'autonomous' software. A shift toward strict mechanistic precision ('the model generates text statistically associated with its training data') maximizes transparency and testability. This vocabulary makes the supply chains of data and labor visible and keeps liability firmly on corporate developers, benefiting civil society, artists, and regulators. However, it costs intuitive grasp; purely statistical explanations can be dense and alienating for lay audiences, potentially hindering adoption in design and art workflows where the 'centaur' narrative is useful for workflow integration. Alternatively, deepening anthropomorphic clarity—fully embracing the AI as a 'collaborator'—prioritizes user experience and human-computer interaction paradigms, but embeds dangerous assumptions about machine reliability, risking severe alignment and accountability failures. Institutional changes could support a more balanced future: journals could require explicit mechanistic translations of agential shorthand; regulatory frameworks could mandate transparency about a firm's discourse approach (e.g., forbidding marketing that claims the AI 'understands'); and educational programs could teach both the statistical reality and the user-interface metaphor, explicitly mapping the trade-offs. Ultimately, the choice of vocabulary determines what is visible: mechanistic language enables accountability but costs narrative ease, while anthropomorphism provides intuitive engagement but forecloses clear liability.

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.

Initiating Discourse Analysis

I'm starting the critical discourse analysis now, focusing on AI metaphors and anthropomorphism in the PDF text. I'm carefully reviewing the task details and schema (Version 3.4.3), ensuring strict adherence to the defined structure. The initial phase is dedicated to thoroughly understanding the request's parameters before diving into the text itself.

Reading & Understanding Source

I've begun analyzing the 27-page preprint, focusing on the authors' argument for "alternative creativity" in generative AI. I'm noting the differences in AI's creativity compared to humans, specifically focusing on priors, context, value grounding, and operational envelope. My aim is to extract this argument within the JSON framework.

Analyzing Authorship Attribution

I'm now diving deeper into the nuances of authorship attribution within the text. The authors carefully assign responsibility to firms and prompters, and acknowledge the AI's evaluations reflect external judgements, making it clear that no AI is an author. My task requires analyzing any language that implies AI has understanding or agency, even when the authors distance the AI from phenomenological consciousness or the human experience of creating.

Defining AI Creativity

I'm now focusing on the authors' central argument concerning "alternative creativity." They emphasize AI creativity's behavioral similarity to humans, specifically "compression and navigation," while highlighting fundamental differences in its underlying constitution. This involves contrasting AI's training-based approach with human experience, static archives versus lived situations, proxied versus felt values, and varying capability envelopes. I'm carefully documenting the argument for the JSON structure.

Identifying AI Metaphors

I've been identifying key metaphors used to describe AI in the text. I've noted instances of AI being framed as a witness or testifier with a diffuse history, a learner acquiring competence, a tireless contributor, an autonomous creator generating candidates, a student internalizing artistic regularities, and a hyperrealist painter. My focus is on how these metaphors frame the AI's agency and capability.

Framing AI's Roles

I'm now identifying additional metaphors. I've noted the AI framed as a navigator/explorer in a limited space. Also, the contrast of AI as a sampler versus a navigator. The AI's role as a "creator" with the potential for "taste" is also emerging. Finally, I'm noting the AI as a proposer of candidates. I'm carefully reviewing all instances to ensure the JSON schema is compliant.

Revising Metaphor Taxonomy

I'm refining my taxonomy of AI metaphors, focusing on two new examples. The AI is now framed as an art student "internalizing artistic regularities," which implies a cognitive learning process. Secondly, it is described as a tireless hyperrealist painter. This represents goal-directed behavior. I'm carefully reviewing the schema for alignment with these new instances and ensuring appropriate agency level categorization, and refining the "acknowledgment" of these roles in the text.

Expanding Metaphor Analysis

I'm now integrating new examples into the evolving framework. I've classified video models using the "as if" construction. I'm focusing on the model's portrayal as a spatial explorer, intentionally navigating, choosing paths, and borrowing values. I've also identified the model as a team member or collaborator contributing to a shared goal with social agency. Furthermore, I've noted the AI as a creator whose taste can only be acquired through experience. I'm actively assessing all these metaphors.

Examining Cognitive States

I'm now analyzing metaphors related to AI's cognitive and experiential aspects. I've noted AI framed as acquiring taste through social living, implying aesthetic judgment and social enculturation. I'm also analyzing the AI's "effort," and how it contributes independently and searches for novel solutions. My goal is to categorize these based on agency levels.

Refining Metaphor Classification

I'm now refining the metaphor classification. I'm focusing on the AI as an active seeker of novelty and analyzing the implied agency as goal-directed behavior. I need to ensure strict compliance with schema definitions, particularly actor visibility. I'm also preparing for Task 2, mapping source to target domains across eight instances, addressing underlying assumptions.

Categorizing Agency Levels

I'm now classifying agency levels in the text using a mechanistic approach, focusing on eight specific instances. I'm carefully evaluating passages and assigning types such as "functional," "dispositional," and "intentional." My focus is on achieving mechanistic precision in my epistemic corrections and ensuring the exact values for "type" and "role" within the schema are defined.

Analyzing Agency Shift

I'm now focusing on Task 5's critical observations, particularly agency slippage. The authors carefully establish a mechanical basis but still frame the AI as an "alternative creator" and "navigator." This creates a climb from "computational activity" to "intentional-state attribution." This involves an analysis of metaphors, particularly how performance builds trust. I'm focusing on "exploratory breadth" and "tirelessness," and contrasting them with the relation-based trust found in human interactions.

Evaluating Metaphor Usage

I'm now diving into the implications of framing AI as a "proposer of candidates." I'm also analyzing the inherent goal-directed behavior suggested when the AI's described as an "active seeker of novelty." These examples require careful schema mapping and accurate agency level classification. I need to make sure my analysis captures the nuances in the text.

Synthesizing Argument & Implications

I'm synthesizing the paper's key arguments into a five-paragraph structure. The initial focus is on dominant metaphors like AI as a spatial navigator and collaborative agent. I'm examining how these metaphors contribute to an illusion. Analyzing implications, from copyright to labor displacement, is my next task, focusing on the material stakes. Then I'll reframe the metaphors using precision to develop literacy as a counter-practice.

Verifying Schema Adherence

I'm now meticulously reviewing my work against the specified constraints. I'm focusing on ensuring valid JSON output, particularly the absence of control tokens within strings. I'm also confirming I am using exact enum values, meeting word count minimums, and including all required fields. Further, I am making certain that I'm describing mechanisms at the level of training and generation, and naming the building corporation of AI systems.

Constructing Metaphor Mapping

I'm now carefully constructing the JSON, focusing on an art student metaphor example. I'm detailing the frame and projection, explicitly noting how cognitive terms, such as "internalizing," are applied to the mechanistic processes of gradient descent. I'm focusing on the agency gradient, contrasting computational activity with implied epistemic states, while clarifying the absence of subjective awareness.

Clarifying AI's Role

I'm currently focusing on a critical instance where AI is portrayed as an art student "internalizing artistic regularities." I'm elaborating on the framing, noting the projection of cognitive terms onto mechanistic processes. This is an epistemic agent, but there is no subjective awareness. My aim is to clearly delineate the agency gradient and ensure accurate schema mapping.

Deconstructing AI Learning

I'm now analyzing the implications of phrases like "the system is learning," and identifying how they obscure the human actors involved in AI development. I'm focusing on the corporate decisions behind model training and data harvesting. I am assigning the correct agency levels for an example of a generative model, and evaluating how it frames the AI's role.

Detailing Computational Activity

I'm now carefully deconstructing the concept of the AI "learning" within the text, framing it as a subtle slippage. I'm focusing on the mechanistic processes involved, such as "compressed representation" and "bottleneck," alongside how the model is trained, and how it avoids intentional language. I'm clarifying how these systems utilize statistical distributions. I am also identifying where agency shifts occur.

Refining Text Formatting

I'm now focused on eliminating any unexpected characters from the text to ensure the JSON structure is perfectly valid. I'm carefully reviewing all the components, from the header to my final sentences. I am mentally organizing each section and will meticulously compile the JSON, making certain there are no unescaped characters, newlines, or tabs, to guarantee the output's compliance with the specific formatting requirements.

Analysis Provenance

Run ID: 2026-11-03-alternative-creativity-why-generative-ai-metaphor-rguz2g Raw JSON: 2026-11-03-alternative-creativity-why-generative-ai-metaphor-rguz2g.json Framework: Metaphor Analysis v6.8 Schema Version: 3.4.3 Generated: 2026-11-03T14:02:58.635Z

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