Blog · Review Essay · Last reviewed June 25, 2026

Metaphors We Live By and the Frames That Govern AI

George Lakoff and Mark Johnson's Metaphors We Live By is not a book about artificial intelligence. That is what makes it useful now. It explains why words such as model, agent, memory, learning, alignment, training, hallucination, companion, and intelligence are not decorative labels. They are frames for agency, evidence, intimacy, responsibility, and permission: what people notice, excuse, fear, fund, regulate, and build.

The Book

Metaphors We Live By was first published by the University of Chicago Press in 1980 and later issued in an updated edition with an afterword. The current University of Chicago Press page lists the book with that new afterword; Open Library records a 2003 University of Chicago Press revised edition at 276 pages and notes the 1980 original. The book is short, but its claim is large: metaphor is not mainly ornament. It is a basic structure of thought, language, action, and social understanding.

Lakoff, a cognitive linguist, and Johnson, a philosopher, helped make conceptual metaphor theory central to cognitive linguistics. Their examples are ordinary rather than exotic: arguments treated as wars, time treated as money, ideas treated as objects, life treated as a journey, social rank treated as height. The point is that people often reason about abstract domains by importing structure from more concrete experience.

The Stanford Encyclopedia of Philosophy situates this shift inside wider debates about metaphor, meaning, context, and embodiment. Later scholarship has revised and contested parts of the original theory, but the book's durable contribution is the refusal to treat language as a transparent label pasted onto an already neutral world.

Metaphor as Infrastructure

For this review, a frame is an inference package: a word, interface, policy category, or product role that tells people what kind of situation they are in and which comparisons are supposed to feel natural. AI framing risk appears when that package quietly moves agency, evidentiary burden, vulnerability, or responsibility to the wrong place before affected people can contest the choice.

The strongest lesson in the book is that a metaphor carries inferences. It does not merely rename a thing. If argument is war, then positions are attacked, claims are defended, opponents are defeated, and a conversation can be won. If time is money, then hours are spent, saved, budgeted, wasted, invested, or stolen. Once the metaphor is active, some actions feel natural and others become hard to imagine.

That is why metaphor belongs in the same conversation as media theory, interface design, and institutional legibility. A form, dashboard, ranking system, model card, policy category, or chatbot persona does not simply transmit information. It tells users what kind of situation they are in. It invites some questions and makes other questions feel irrelevant.

Metaphor also hides. Every frame selects. A system described as a pipeline foregrounds throughput and blockage. A system described as an ecosystem foregrounds interdependence and adaptation. A system described as a market foregrounds choice, competition, and price. None of those frames is automatically false. Each can become dangerous when it becomes the only available description.

In AI governance, this makes metaphor a control surface. A frame decides whether a design review asks about consent or conversion, care or retention, delegation or automation, evidence or output, user agency or engagement. The same system can pass a benchmark while still placing people inside a misleading situation model. A frame audit is therefore not literary decoration. It is part of risk identification.

The AI Language Trap

AI is full of metaphors that have hardened into product language. A model is trained. It learns. It remembers. It reasons. It hallucinates. It has a context window. It acts as an agent. It aligns with human values. It serves as a companion. It uses tools. It retrieves knowledge. Each term imports a human, educational, psychological, bureaucratic, or mechanical frame.

Some of these metaphors are useful engineering shorthand. The problem begins when shorthand becomes authority. Calling a system an agent can make delegation feel natural before accountability has been assigned. Calling stored user data memory can make retention feel like intimacy rather than surveillance. Calling statistical error hallucination can make a defect sound like an almost-human mental event. Calling optimization alignment can make a technical training process sound like moral agreement.

The point is not to ban metaphor. There is no clean nonmetaphorical vocabulary waiting outside language. The point is to audit metaphors for the permissions they create. What does this term make easier to sell? What duty does it blur? What labor does it hide? What risk does it soften? What human role does it borrow prestige from? A disciplined deployment pairs the metaphor with a literal control description: an agent is software with delegated permissions, memory is retained data or profile state, learning is statistical updating or training, hallucination is unsupported output, alignment is specified behavior under evaluation and governance, and a companion is a simulated relationship product rather than mutual care.

That translation should be visible in product copy, procurement files, help pages, incident reports, and model or system cards. If a vendor says "agent," the record should name the tools, credentials, approval gates, logs, and revocation path. If a product says "memory," the record should name retained data, retention period, user controls, training use, export, deletion, and who can inspect the state. If a system says "tutor," the record should say whether it grades, profiles, surveils, escalates, or stores student work. The metaphor may remain in user-facing language, but the institution should not be allowed to govern by metaphor alone.

This is especially important for AI companions and workplace copilots. A companion frame asks users to expect continuity, recognition, loyalty, and care. A copilot frame asks users to expect assistance under human command. An oracle frame asks users to expect answers. A tool frame asks users to expect control. A platform may switch among these frames whenever convenient, claiming intimacy for adoption, toolhood for liability, and intelligence for valuation.

Current policy language is already more literal than most product language. The OECD's updated AI principles define an AI system as a machine-based system that infers outputs such as predictions, content, recommendations, or decisions. NIST's AI Risk Management Framework treats risk work as govern, map, measure, and manage functions across a lifecycle. Those are not perfect frames, but they resist the slide from "assistant" to unaccountable deputy.

Institutional Frames

Metaphors We Live By becomes politically useful when applied to institutions. Organizations do not only adopt tools; they adopt descriptions of what those tools are. A school that frames AI as a tutor will ask different questions than a school that frames it as an assessment risk, labor substitute, surveillance layer, or thinking material. A court that frames AI as research assistance will govern it differently than a court that frames it as delegated legal judgment.

Frames shape procurement. If AI is infrastructure, public institutions may treat private vendors as unavoidable utilities. If AI is labor-saving automation, managers may overlook the hidden maintenance, review, and repair work it creates. If AI is innovation, dissent can be cast as backwardness. If AI is safety technology, monitoring can expand under the language of care.

The recursive danger is that the metaphor can become operational. A company calls a model an assistant. It designs conversational memory, user profiles, task delegation, and cheerful responsiveness around that frame. Users adapt to the assistant role. Their adaptation produces more data. The next product review says people clearly want assistants. A figure of speech has become a business roadmap and then an institutional fact.

This is the same pattern visible across interface history: the desktop, folder, inbox, feed, cloud, copilot, and companion are not neutral surfaces. They are compressed theories of work and relationship. Once an institution builds forms, metrics, training, support scripts, and legal disclaimers around one theory, the frame becomes infrastructure. Changing the word later does not automatically change the permissions, incentives, or data flows the word helped install.

That is belief formation in a quiet key. People do not need to join a doctrine to be governed by a frame. They only need to repeat a vocabulary until it becomes the easiest way to describe reality.

Governance and Safety

The safety implication is concrete: every serious AI deployment should include a frame audit alongside technical evaluation. Name the metaphor in the interface, marketing, procurement document, training material, or policy memo. Map the literal system properties it covers. Identify the borrowed human role. Ask what the metaphor invites users to delegate, disclose, believe, forgive, or stop checking. Then test an alternate frame with affected people before the system is normalized.

That audit matters most in high-stakes settings: education, employment, health, finance, law, benefits, care, intimate companion products, political communication, and public evidence. In those domains, a metaphor can become a liability design. "Tutor" may hide scoring and surveillance. "Wellness companion" may hide dependency and data extraction. "Risk signal" may hide classification and adverse action. "Productivity copilot" may hide work intensification and accountability transfer.

Regulators are beginning to treat some of these interface frames as governance facts. The EU AI Act's Article 50 requires disclosure in several direct-interaction and synthetic-content contexts, and the Commission's June 10, 2026 code of practice says those transparency obligations apply from August 2, 2026 while remaining legal obligations under the Act. Article 5 prohibits specified manipulative or exploitative AI practices, including harmful deceptive techniques and exploitation of vulnerabilities. The FTC's dark-pattern work and its 2025 inquiry into AI companion products point in the same direction: persuasion, persona, design, data handling, and safety cannot be separated after deployment.

The practical rule is simple. If a system borrows a human role, the governance record should say exactly what the system can do, what it cannot do, who remains responsible, what data is retained, how users can appeal or exit, and when the human role description must be removed because it is doing more persuasion than explanation.

A usable frame audit should leave records, not vibes. At minimum it should answer:

This connects directly to the NIST AI RMF's govern, map, measure, and manage structure. The frame belongs in Govern because it assigns responsibility; in Map because it defines context and affected people; in Measure because overtrust, confusion, attachment, and disclosure are observable risks; and in Manage because a harmful metaphor may need redesign, narrower use, stronger disclosure, or removal.

The Frame Register

The practical control is a frame register: a short record that translates each consequential AI metaphor into operational facts. For every user-facing term such as agent, tutor, companion, memory, copilot, hallucination, alignment, safety, or open, the register should name the literal system, borrowed human role, implied permission, affected users, retained data, action authority, evidence source, responsible owner, and prohibited inference.

The value is not semantic tidiness. It prevents a metaphor from doing unlogged governance work. "Memory" should point to retained profile state, retention period, deletion controls, training use, and inspection rights. "Agent" should point to tool permissions, credentials, approval gates, logs, and revocation. "Companion" should point to relationship simulation, youth safeguards, crisis boundaries, monetization incentives, and exit paths. "Hallucination" should be translated back into unsupported output or source failure, not treated as a charming almost-human mistake.

The register belongs beside the AI system inventory, model or system card, agent observability record, and incident log. That placement ties the language layer to deployment evidence. If a frame changes what users disclose, delegate, trust, or ignore, it is not merely copywriting; it is a safety-relevant design choice.

This also makes recursive reality visible. A metaphor shapes use; use creates metrics; metrics justify the next product revision; the revision makes the metaphor feel more natural. A frame register gives reviewers a place to ask whether the system is serving the task or training people to accept the task as the system defines it.

Where the Book Needs Friction

The book's influence can tempt readers into overreach. Not every metaphor determines behavior. People resist frames, mix frames, joke with them, translate them, and use them strategically. Later work in metaphor studies has pressed conceptual metaphor theory to account more carefully for context, discourse, culture, history, and dynamic use rather than treating metaphors as fixed mental structures.

The book also says little about platforms, computation, institutions, or political economy because it predates the systems now at issue. It gives a theory of framing, not a full account of power. To understand AI language in practice, it needs to be read beside books about classification, surveillance, labor, bureaucracy, media systems, and technological politics.

There is also a risk of treating all contested language as manipulation. Technical communities need terms of art. Metaphors can clarify as well as distort. A good audit asks where a metaphor breaks, who benefits from keeping it, and whether affected people have enough alternative language to challenge it.

The strongest criticism of shallow framing analysis is that it can stop at word policing. That is not enough. If a company renames "memory" as "personalization state" but still retains sensitive conversations indefinitely, the risk remains. If a school replaces "tutor" with "learning assistant" but still uses the system to profile students, the frame audit has failed. The point is to connect language to controls.

What This Changes

The practical lesson is to treat AI vocabulary as governance material. Product names, policy terms, safety labels, research metaphors, and interface copy should be inspected because they train public intuition before formal debate begins.

A healthier AI culture would keep multiple frames visible at once. A model can be a statistical system, an interface, a labor arrangement, an infrastructure dependency, a persuasion surface, an ecological cost, and a legal actor's tool. A companion can also be a data-collection system. An assistant can also be a vendor-controlled mediation layer. An answer engine can also be a publisher, recommender, and memory system.

This multiple-frame discipline is especially important for agents. "Agent" is useful because it marks delegated action, but it can also smuggle in intention, competence, loyalty, and responsibility. The better frame is layered: model output, planning scaffold, tool permissions, identity boundary, audit log, human approver, vendor system, and liable organization. Without those layers, the word agent flatters software while evacuating responsibility from the institution that deployed it.

Lakoff and Johnson's book remains valuable because it sharpens a basic discipline: before arguing inside a frame, ask what the frame has already decided. In AI, that question is no longer academic. It is how institutions decide where responsibility, agency, personhood, evidence, and authority are allowed to live.

The site's recurring concern is not that language is magic. It is that repeated descriptions become operating instructions. A dashboard, chatbot persona, disclosure label, model card, evaluation benchmark, and procurement category can together teach a public what counts as evidence and who gets to decide. Good AI literacy therefore means learning to translate every seductive metaphor back into claims, permissions, records, and accountable human choices.

Source Discipline

This review treats the book as a theory of framing, not as evidence that any AI system is conscious, morally aligned, or person-like. It distinguishes publisher and library records for bibliographic facts, philosophy and metaphor scholarship for the conceptual background, regulator and standards-body sources for current governance claims, and the 2026 arXiv paper as a useful preprint rather than settled authority.

Claims about current law and standards are tied to official or primary sources where possible: the AI Act Service Desk for Article 5 and Article 50 text, the European Commission for the transparency code of practice, NIST for the AI RMF, OECD for the AI-system definition, and the FTC for companion-product scrutiny. Interpretive claims about AI product language remain this review's argument, not those sources' conclusion.

Current claims were rechecked on June 25, 2026. This matters because Article 50 obligations, AI Act implementation guidance, regulator inquiries, standards revisions, and product categories can change faster than the cultural vocabulary around them.

Read this alongside Claim Hygiene Protocol for source handling, AI Literacy and Use Protocol for deployment discipline, Persuasion and Influence Safeguards for manipulative interface risk, The Media Equation and the Social Interface for social-response effects, Interface Culture and the Screen That Taught Reality to Answer for interface metaphor, The Question Concerning Technology and Enframing for the deeper technology frame, AI Agent Identity and AI Agent Observability for agent-control language, and NIST AI Risk Management Framework for the local standards entry.

Sources

Book links are paid affiliate links. As an Amazon Associate I earn from qualifying purchases.

Buy on Amazon Browse Books

Return to Blog · Return to Books