Interview with Sam Altman - Structured Output Example #

* **Source**: An Interview with OpenAI CEO Sam Altman About DevDay and the AI Buildout
https://stratechery.com/2025/an-interview-with-openai-ceo-sam-altman-about-devday-and-the-ai-buildout/

---

```json
 {
  "metaphorAudit": [
    {
      "title": "AI as an Intentional Agent",
      "quote": "even when ChatGPT screws up, hallucinates, whatever, you know it’s trying to help you, you know your incentives are aligned.",
      "frame": "Model as a thinking partner",
      "projection": "Intention (trying), goal-orientation, and aligned incentives, which are attributes of a conscious agent.",
      "acknowledgment": "Unacknowledged. Presented as a direct description of the model's internal state.",
      "implications": "This builds immense trust by framing the system as a benevolent partner whose goals are identical to the user's, rather than a tool optimized for plausible text generation. It encourages forgiveness for errors by attributing them to failed attempts rather than systemic flaws."
    },
    {
      "title": "Cognition as Information Possession and Discretion",
      "quote": "you’ll want it to still know you and have your stuff and know what to share and what not to share.",
      "frame": "Model as a conscious confidant",
      "projection": "Memory ('know you'), ownership ('have your stuff'), and ethical or social judgment ('know what to share').",
      "acknowledgment": "Unacknowledged.",
      "implications": "Obscures the reality of data processing and retrieval by attributing human-like memory and discretion. This can lead to misplaced trust in the system's ability to make nuanced, context-aware decisions about privacy and data sharing."
    },
    {
      "title": "System Error as Human Fallibility",
      "quote": "even when ChatGPT screws up, hallucinates, whatever...",
      "frame": "Model as a fallible person",
      "projection": "The human acts of making a mistake ('screws up') and experiencing a delusion ('hallucinates').",
      "acknowledgment": "Unacknowledged. These terms are used as standard industry jargon.",
      "implications": "This framing domesticates system failures. Instead of being seen as computational artifacts or statistical errors in a probabilistic model, they are framed as understandable human-like mistakes, which reinforces an emotional bond and lessens the perceived severity of the output's unreliability."
    },
    {
      "title": "Interaction as an Interpersonal Relationship",
      "quote": "you’ll feel like you just have this one relationship with this entity that’s helping you.",
      "frame": "Model as a relational partner",
      "projection": "The capacity for a persistent, evolving, two-way relationship, typically reserved for sentient beings.",
      "acknowledgment": "Unacknowledged.",
      "implications": "Encourages users to form an emotional attachment, increasing engagement and dependence. It frames the tool not as a utility but as a companion, which blurs the critical distinction between artifact and agent and may affect user objectivity."
    },
    {
      "title": "Output Generation as Empathetic Creation",
      "quote": "we tried to make the model really good at taking what you wanted and creating something good out of it.",
      "frame": "Model as an intuitive artist",
      "projection": "The ability to understand desire ('taking what you wanted') and make aesthetic judgments ('creating something good').",
      "acknowledgment": "Unacknowledged.",
      "implications": "Suggests the model has a deep, empathetic understanding of user intent rather than simply executing pattern-matching on prompts based on training data. This elevates its status from a generative tool to a creative collaborator with its own sense of quality."
    },
    {
      "title": "AI as a Companion Animal or Child",
      "quote": "If ChatGPT messes up, it’s like, ‘It’s okay, you’re trying my little friend’.",
      "frame": "Model as a pet or small child",
      "projection": "Innocence, eagerness to please, and deserving of forgiveness and patience for its mistakes.",
      "acknowledgment": "Acknowledged as a user's perspective that Altman is reporting.",
      "implications": "This is a powerful framing for cultivating user tolerance for errors. By infantilizing the AI, its failures become endearing rather than critical flaws, fostering a protective and loyal user base that is less likely to scrutinize its outputs critically."
    },
    {
      "title": "Functionality as Inherent Agency",
      "quote": "ChatGPT plugins, that didn’t work. GPTs actually did work.",
      "frame": "Product as an autonomous actor",
      "projection": "Agency, the ability to succeed ('did work') or fail ('didn't work') on its own terms.",
      "acknowledgment": "Unacknowledged.",
      "implications": "Attributes success or failure to the product itself, masking the complex interplay of design choices, engineering execution, user behavior, and market fit. It simplifies a complex outcome into a story about the product's own performance."
    },
    {
      "title": "Technology Adoption as Liquid Permeation",
      "quote": "I think it will just kind of seep everywhere into every consumer product and every enterprise product too.",
      "frame": "AI as a pervasive liquid",
      "projection": "Inevitability, naturalness, and passive diffusion, like water soaking into a surface.",
      "acknowledgment": "Partially acknowledged with the hedge 'kind of'.",
      "implications": "Frames the widespread integration of AI as a natural, unstoppable, and almost organic process. This framing can downplay the active corporate strategies, capital investments, and deliberate decisions driving its proliferation."
    },
    {
      "title": "Business Strategy as High-Stakes Gambling",
      "quote": "We are going to make a bet, the company scale bet that this is the right time to do it.",
      "frame": "Corporate strategy as a wager",
      "projection": "Risk, chance, high stakes, and intuition, rather than purely data-driven calculation.",
      "acknowledgment": "Unacknowledged.",
      "implications": "Builds a narrative of visionary risk-taking and conviction. It frames massive capital expenditure as a bold, decisive action in the face of uncertainty, which is rhetorically more compelling than a description of strategic financial planning."
    },
    {
      "title": "Emotional Response as Physical Impact",
      "quote": "video hits people, particularly rights owners very differently than still images, it turns out.",
      "frame": "Media as a physical force",
      "projection": "The ability to exert a direct, forceful impact on a person's emotional state.",
      "acknowledgment": "Unacknowledged.",
      "implications": "Emphasizes the potent emotional effect of AI-generated video, framing it as an active force rather than a passive artifact being interpreted by a viewer. This can be used to justify different policy or control mechanisms based on the medium's perceived power."
    }
  ],
  "sourceTargetMapping": [
    {
      "quote": "even when ChatGPT screws up, hallucinates, whatever, you know it’s trying to help you, you know your incentives are aligned.",
      "sourceDomain": "A helpful, goal-oriented person or partner",
      "targetDomain": "An LLM's output generation process",
      "mapping": "The source domain's structure of intention, effort ('trying'), and shared goals ('aligned incentives') is projected onto the target's process of calculating probable token sequences. This invites the inference that the model possesses a benevolent internal state that guides its actions.",
      "conceals": "The mechanistic, probabilistic nature of the system. It hides that the output is not a result of intention or desire, but of statistical optimization based on patterns in training data and reinforcement learning from human feedback. The model has no 'incentives' of its own."
    },
    {
      "quote": "you’ll want it to still know you and have your stuff and know what to share and what not to share.",
      "sourceDomain": "A discreet, knowledgeable human assistant or friend",
      "targetDomain": "The AI system's data access and processing functions",
      "mapping": "The human concepts of personal knowledge ('know you'), possession ('have your stuff'), and social-ethical judgment ('know what to share') are mapped onto the system's functions of querying a user data store and applying programmed permissions.",
      "conceals": "The computational reality. The system does not 'know' or 'own' anything; it accesses and processes data. Its decisions about sharing are based on algorithms and pre-set rules, not on situational understanding or ethical reasoning."
    },
    {
      "quote": "even when ChatGPT screws up, hallucinates, whatever...",
      "sourceDomain": "Human cognition and error",
      "targetDomain": "AI model output inaccuracies",
      "mapping": "The structure of human fallibility is projected onto the model's behavior. 'Screws up' maps to making a mistake, while 'hallucinates' maps a severe neuro-psychological event onto the generation of factually incorrect or nonsensical text.",
      "conceals": "The source of the error. A human 'screws up' due to carelessness or a flawed mental model. An LLM 'hallucinates' because it is a probabilistic text generator with no grounding in factual reality; it is simply generating a plausible-sounding but incorrect sequence."
    },
    {
      "quote": "you’ll feel like you just have this one relationship with this entity that’s helping you.",
      "sourceDomain": "An interpersonal relationship",
      "targetDomain": "A user's ongoing interaction with a software service",
      "mapping": "The attributes of a human relationship—continuity, reciprocity, emotional connection, shared history—are projected onto the user's interaction with the AI. The system is framed as an 'entity' one can have a 'relationship' with.",
      "conceals": "The asymmetry of the interaction. It is a one-way process where a user interacts with a tool. The tool does not experience the relationship; it merely maintains a history of interactions to inform future outputs. It hides that the 'relationship' is entirely within the user's mind."
    },
    {
      "quote": "we tried to make the model really good at taking what you wanted and creating something good out of it.",
      "sourceDomain": "An intuitive, creative collaborator",
      "targetDomain": "The fine-tuning and prompt-engineering process of an AI model",
      "mapping": "The human ability to intuit desire ('taking what you wanted') and apply aesthetic judgment ('creating something good') is mapped onto the model's function of generating output that statistically correlates with prompts and user preferences.",
      "conceals": "The lack of genuine understanding. The model does not 'want' what the user wants. It processes the prompt as a set of tokens and generates a high-probability completion based on patterns learned from vast datasets of human-created text and images."
    },
    {
      "quote": "If ChatGPT messes up, it’s like, ‘It’s okay, you’re trying my little friend’.",
      "sourceDomain": "A small child or pet learning a task",
      "targetDomain": "An AI model producing an incorrect output",
      "mapping": "The source's relational structure—where a well-meaning but undeveloped agent makes mistakes due to lack of skill, but its effort is commendable—is projected onto the AI. This invites feelings of patronage, affection, and forgiveness.",
      "conceals": "The nature of the artifact. It is not a 'friend' that is 'trying.' It is a massive computational system executing its programming. This metaphor conceals the scale and complexity of the system while excusing its fundamental reliability issues."
    },
    {
      "quote": "ChatGPT plugins, that didn’t work. GPTs actually did work.",
      "sourceDomain": "A person or organism performing a task",
      "targetDomain": "A software product's market adoption and utility",
      "mapping": "The binary outcome of success ('work') or failure ('didn't work') associated with an agent's attempt at a task is mapped onto the complex process of a product's lifecycle.",
      "conceals": "The external factors. It hides that the outcome depends on design, user interface, marketing, developer adoption, and alignment with user needs, not an inherent quality of 'working' within the product itself."
    },
    {
      "quote": "I think it will just kind of seep everywhere into every consumer product...",
      "sourceDomain": "A liquid spreading through a porous material",
      "targetDomain": "The integration of AI technology across the economy",
      "mapping": "The properties of a liquid—passivity, inevitability, ability to fill any space—are projected onto the process of technological diffusion. This suggests a natural, uncontrolled, and comprehensive spread.",
      "conceals": "Agency and strategy. This metaphor conceals the fact that AI integration is the result of billions of dollars in targeted investment, strategic business decisions, marketing efforts, and corporate competition. It is not a passive or natural phenomenon."
    },
    {
      "quote": "We are going to make a bet, the company scale bet that this is the right time to do it.",
      "sourceDomain": "Gambling or making a wager",
      "targetDomain": "A corporate investment strategy",
      "mapping": "The structure of a bet—involving risk, uncertainty, a moment of decision, and a binary win/loss outcome—is mapped onto the complex, ongoing process of capital allocation and infrastructure development.",
      "conceals": "The calculated nature of the strategy. While risky, the decision is based on extensive research, market analysis, and financial modeling. Calling it a 'bet' emphasizes the dramatic risk while downplaying the underlying analytical foundation."
    },
    {
      "quote": "video hits people...very differently than still images...",
      "sourceDomain": "A physical collision or impact",
      "targetDomain": "The emotional effect of viewing media",
      "mapping": "The experience of being struck by a physical object is mapped onto the psychological and emotional process of a person interpreting and reacting to a video. It implies a direct, powerful, and almost involuntary effect.",
      "conceals": "The role of the viewer's interpretation. A person's reaction to media is subjective and mediated by their own experiences, beliefs, and context. The 'impact' metaphor frames the media as the sole agent, concealing the active role of the human mind in creating that emotional response."
    }
  ],
  "explanationAudit": [
    {
      "quote": "you know it’s trying to help you, you know your incentives are aligned.",
      "explanationTypes": [
        {
          "type": "Intentional",
          "definition": "Explains actions by referring to goals/desires."
        },
        {
          "type": "Reason-Based",
          "definition": "Explains using rationales or justifications."
        }
      ],
      "analysis": "This is a quintessential 'why' explanation, attributing the AI's output to an internal mental state of wanting ('trying') to be helpful and a rational calculus ('incentives are aligned'). The explanation slides directly from the system's observable output (a helpful-sounding response) to a hidden, unobservable internal purpose. This obscures the mechanistic 'how': the model works by selecting tokens that have a high probability of appearing in contexts labeled 'helpful' during its RLHF training.",
      "rhetoricalImpact": "This generates profound trust and personal connection. By framing the AI as an intentional partner whose goals match the user's, it is perceived as safe, benevolent, and on 'your side,' mitigating concerns about its unpredictable or erroneous behavior."
    },
    {
      "quote": "you’ll want it to still know you and have your stuff and know what to share and what not to share.",
      "explanationTypes": [
        {
          "type": "Dispositional",
          "definition": "Attributes tendencies or habits."
        },
        {
          "type": "Intentional",
          "definition": "Explains actions by referring to goals/desires."
        }
      ],
      "analysis": "This explanation attributes cognitive states ('know') and judgment ('know what to share') to the system. This is a slippage from 'how' it works (accessing a database and applying permission rules) to 'why' it behaves protectively (because it 'knows' you and your preferences). The framing implies a persistent, knowing state rather than a series of stateless computations informed by stored data.",
      "rhetoricalImpact": "It fosters a sense of security and personalization, making the user feel 'known' and looked after by the system. This encourages deeper integration into personal workflows and greater disclosure of personal data, based on the perception of the AI as a discreet assistant."
    },
    {
      "quote": "ChatGPT plugins, that didn’t work. GPTs actually did work.",
      "explanationTypes": [
        {
          "type": "Dispositional",
          "definition": "Attributes tendencies or habits."
        }
      ],
      "analysis": "This explanation attributes the success or failure to a disposition of the products themselves ('did work' vs. 'didn't work'). It's a shorthand that avoids the complex 'how' and 'why' of product-market fit. The slippage is from a complex systemic outcome to an inherent quality of the artifact. It avoids explaining how design, usability, and user needs contributed to the outcome.",
      "rhetoricalImpact": "This personifies the products as actors in a drama of success and failure. It creates a simple narrative that is easy to grasp but obscures the complex lessons of product development and user-centered design. It makes the technology itself seem agentive."
    },
    {
      "quote": "we tried to make the model really good at taking what you wanted and creating something good out of it.",
      "explanationTypes": [
        {
          "type": "Intentional",
          "definition": "Explains actions by referring to goals/desires."
        },
        {
          "type": "Functional",
          "definition": "Describes purpose within a system."
        }
      ],
      "analysis": "This is a hybrid explanation. Altman describes OpenAI's intent ('we tried to make...'), but projects that intent onto the model's capabilities ('good at taking what you wanted'). The slippage occurs when the designers' goal is described as an innate skill of the model itself. The 'how' (the model is optimized to generate outputs that correlate with high user ratings) is reframed as 'why' (it's good at understanding and creating).",
      "rhetoricalImpact": "This frames the AI as an intuitively skillful tool, almost a mind-reader. It boosts user confidence that the model can successfully interpret vague or complex requests, encouraging more ambitious use and masking the underlying trial-and-error process of prompt engineering."
    },
    {
      "quote": "If ChatGPT messes up, it’s like, ‘It’s okay, you’re trying my little friend’.",
      "explanationTypes": [
        {
          "type": "Reason-Based",
          "definition": "Explains using rationales or justifications."
        },
        {
          "type": "Intentional",
          "definition": "Explains actions by referring to goals/desires."
        }
      ],
      "analysis": "Altman is reporting a user's explanation for their own behavior (forgiveness). The user explains their patience by attributing good intentions ('trying') to the AI. This is a complete shift from a 'how' analysis of the system's failure to a 'why' justification for human emotional response, based on a presumed mental state in the machine.",
      "rhetoricalImpact": "By highlighting and validating this user perspective, it normalizes the idea of having a personal, forgiving relationship with the AI. It suggests that the proper response to AI failure is emotional support, not technical critique, which deeply reinforces user loyalty."
    },
    {
      "quote": "video feels much more real and lifelike and there’s a stronger emotional resonance.",
      "explanationTypes": [
        {
          "type": "Dispositional",
          "definition": "Attributes tendencies or habits."
        },
        {
          "type": "Empirical",
          "definition": "Cites patterns or statistical norms."
        }
      ],
      "analysis": "This explanation focuses on the typical effect of the artifact. It describes 'how' videos affect people ('feel more real') by attributing a disposition to the artifact itself. The slippage is subtle: instead of saying 'people tend to react more strongly to video,' the sentence structure implies that the video actively possesses 'stronger emotional resonance.' It explains the 'why' of different rules for video by appealing to its inherent power.",
      "rhetoricalImpact": "This justifies treating AI video as a special category of risk or concern. By framing the medium itself as more potent, it creates a rationale for stricter controls or different business models, shifting focus from user interpretation to the technology's intrinsic properties."
    },
    {
      "quote": "The way that active people on the AI corner of Twitter use AI and the way that the normies in most of the world use AI are two extremely different things.",
      "explanationTypes": [
        {
          "type": "Empirical",
          "definition": "Cites patterns or statistical norms."
        }
      ],
      "analysis": "This is a clear, mechanistic 'how' explanation. It explains a phenomenon (differing user feedback) by citing observable patterns of behavior across different user groups. There is no slippage into agency; it describes how people act, not why the AI 'wants' or 'thinks' something. It is a good example of a non-anthropomorphic explanation.",
      "rhetoricalImpact": "This explanation segments the audience and justifies product decisions that might displease 'power users.' It frames the company's choices as data-driven and responsive to the needs of a larger, less vocal majority, positioning OpenAI as pragmatic and user-focused on a macro scale."
    },
    {
      "quote": "if you give people tools that let them go quickly from idea to creative output, that hits at some very deep human need and I’ve seen it work again and again and again.",
      "explanationTypes": [
        {
          "type": "Functional",
          "definition": "Describes purpose within a system."
        },
        {
          "type": "Theoretical",
          "definition": "Embeds behavior in a larger framework."
        }
      ],
      "analysis": "This explanation operates at a high level of abstraction, explaining 'why' creative tools are successful. It provides a functional 'how' (they reduce friction) embedded in a theoretical framework of human psychology ('deep human need'). It avoids attributing agency to the tool itself, instead focusing on its role or function within a human system.",
      "rhetoricalImpact": "This elevates the significance of the technology. It's not just a product; it's a tool that fulfills a fundamental human desire. This framing provides a powerful, almost philosophical justification for the company's work, positioning it as enabling human potential."
    },
    {
      "quote": "There is a reason people love ChatGPT in a way that they don’t love other large tech companies’ products and I think it is, even when ChatGPT screws up... you know it’s trying to help you",
      "explanationTypes": [
        {
          "type": "Reason-Based",
          "definition": "Explains using rationales or justifications."
        }
      ],
      "analysis": "This explicitly offers a 'why' explanation for user affection ('people love ChatGPT'). The reason provided is not functional (it produces useful text) but intentional (it is 'trying to help'). The slippage is immediate and deliberate, directly linking human emotion (love) to a belief in the machine's benevolent intent.",
      "rhetoricalImpact": "This creates a powerful brand narrative. It positions ChatGPT not as a superior product in terms of performance, but as a more ethically or emotionally aligned entity. This fosters a brand loyalty that is resilient to performance issues or competitive offerings."
    },
    {
      "quote": "we hoped that if we could bring the activation energy of creating down dramatically, a lot more people would do it",
      "explanationTypes": [
        {
          "type": "Intentional",
          "definition": "Explains actions by referring to goals/desires."
        },
        {
          "type": "Genetic",
          "definition": "Traces development or origin."
        }
      ],
      "analysis": "This is a genetic explanation tracing the origin of a product design choice back to the company's intent. It explains 'why' the product was built a certain way by referencing the goals ('we hoped') of its creators. There is no agency slippage here, as the intent is correctly located in the human designers, not the artifact.",
      "rhetoricalImpact": "This frames the company's actions as being motivated by a desire to empower users and unlock creativity. It is a pro-social justification for their product strategy, which builds goodwill and casts the company in a positive, enabling role."
    }
  ],
  "reframedLanguage": [
    {
      "originalQuote": "you’ll want it to still know you and have your stuff and know what to share and what not to share.",
      "reframedExplanation": "You will want the service to access your user history and stored data to personalize its responses, and to apply pre-programmed rules to determine which data can be shared with third-party applications."
    },
    {
      "originalQuote": "you know it’s trying to help you, you know your incentives are aligned.",
      "reframedExplanation": "The system is designed to generate responses that align with user instructions because it has been fine-tuned with data where human reviewers rated helpful and harmless outputs highly."
    },
    {
      "originalQuote": "even when ChatGPT screws up, hallucinates, whatever...",
      "reframedExplanation": "Even when ChatGPT produces factually incorrect or nonsensical output..."
    },
    {
      "originalQuote": "we tried to make the model really good at taking what you wanted and creating something good out of it.",
      "reframedExplanation": "We optimized the model to generate outputs that closely match the descriptions in user prompts and that statistically correlate with content that is highly rated for quality."
    },
    {
      "originalQuote": "you’ll feel like you just have this one relationship with this entity that’s helping you.",
      "reframedExplanation": "Users may develop a sense of continuity and familiarity from repeated interactions with the system, as it uses conversation history to inform its responses."
    },
    {
      "originalQuote": "GPTs actually did work.",
      "reframedExplanation": "The GPTs feature saw higher user adoption and engagement compared to the previous plugins feature."
    },
    {
      "originalQuote": "video hits people... very differently than still images",
      "reframedExplanation": "People often have a stronger emotional response to video content compared to still images due to its dynamic and immersive nature."
    },
    {
      "originalQuote": "OpenAI is not a very well-known brand, ChatGPT is an extremely well-known brand",
      "reframedExplanation": "The product name 'ChatGPT' has achieved significantly higher public recognition than the company name 'OpenAI'."
    }
  ],
  "criticalObservations": {
    "agencySlippage": "The text demonstrates consistent agency slippage, particularly when the topic shifts from infrastructure to user-facing products. While business strategy is framed mechanistically (e.g., as 'capital allocation'), the behavior of ChatGPT is almost always described using agential language. Explanations slide from 'how' the system functions (a functional or empirical view) to 'why' it acts with a certain purpose (an intentional or reason-based view). The phrase 'you know it's trying to help you' is the clearest example, directly attributing intention to a computational process.",
    "metaphorDrivenTrust": "Biological and cognitive metaphors are central to building trust and encouraging user forgiveness. By framing the AI as a 'friend' that is 'trying,' or as a partner with 'aligned incentives,' the system's failures ('screws up,' 'hallucinates') are rendered as forgivable mistakes rather than critical systemic flaws. This emotional framing fosters a resilient user base that is more likely to overlook inaccuracies and form a personal attachment to the product, enhancing brand loyalty.",
    "obscuredMechanics": "The pervasive use of metaphorical language consistently obscures the actual mechanics of large language models. Concepts like 'knowing,' 'understanding,' and 'deciding' are used in place of more accurate terms like 'processing,' 'pattern-matching,' and 'calculating probabilities.' This masks the statistical and non-conscious nature of the technology, creating the illusion of a mind. The user is encouraged to interact with a personality, not a probabilistic system, which fundamentally misrepresents how the technology works and what its limitations are.",
    "contextSensitivity": "Metaphor use is highly context-sensitive. When discussing finance, infrastructure, and competition, the language is more grounded in business and technical jargon ('capital allocation,' 'chip fab capacity'). However, when describing the user experience of ChatGPT or the creative potential of Sora, the language shifts dramatically toward anthropomorphism and cognitive metaphors. This suggests a deliberate or unconscious rhetorical strategy to frame the human-computer interaction in relatable, human terms, making the product feel more intuitive, friendly, and powerful."
  },
  "conclusion": "The discourse in this interview systematically constructs an 'illusion of mind' through a consistent pattern of anthropomorphic and cognitive metaphors. The primary patterns frame AI not as a tool, but as an intentional helper, a trusted friend, and a creative partner. The system is described as 'knowing' user preferences, 'trying' to be helpful, and possessing 'aligned incentives.' These linguistic choices are not mere stylistic flourishes; they are rhetorical strategies that shape understanding, build trust, and manage user expectations, particularly regarding system failures, which are reframed as forgivable human-like errors rather than computational artifacts.\n\nThis construction of a mind-like entity has significant implications for AI literacy. By consistently slipping from mechanistic explanations of 'how' the system works to agential explanations of 'why' it acts, the discourse obscures the underlying probabilistic mechanics. Users are encouraged to build a 'relationship' with an 'entity' instead of learning to operate a complex statistical tool. This can lead to misplaced trust, a misunderstanding of the system's capabilities and limitations, and a reduced capacity for critical evaluation of its outputs.\n\nAs demonstrated in the reframing exercises, achieving greater precision is possible and necessary for responsible communication about AI. The key principle is to shift the language from the cognitive to the computational. Instead of attributing mental states to the model, communicators should describe its processes and functions. This involves replacing verbs like 'think' and 'know' with 'process' and 'correlate,' and attributing agency to the human designers and users rather than the artifact itself. By actively delineating between observed system behavior and attributed internal states, we can foster a more realistic and empowered public understanding of artificial intelligence, treating these systems as the powerful but unthinking artifacts they are."
}