Making Sense of AI Metaphors: A Framework for Classification and Analysis

Written by

Cinzia Pusceddu

on

Metaphors matter, perhaps never more so than when they are used to influence our thinking. And when it comes to AI, metaphors are everywhere. We and AI’s AI Metaphor Encyclopedia aims to be a resource for critical AI literacy, by using a specific framework to classify and analyze these AI metaphors. In this article, researcher Cinzia Pusceddu explains the framework we developed and use, and how it came to be.

What is a metaphor and why do AI metaphors matter?

Metaphor is usually defined as a figure of speech in which one thing is described in terms of another. For a long time, it was treated mainly as a rhetorical or stylistic device, the kind that lets Shakespeare’s Romeo declare “Juliet is the sun”, casting her as the source of light and warmth. Since the well-known 1980 paper Metaphors We Live By, however, many scholars have argued that a metaphor is also a matter of thought and action, not just of language. According to this theory, much of our ordinary conceptual system is structured by metaphorical mappings, which let us understand a less familiar or more abstract domain in terms of a more familiar or more concrete domain.

Artificial Intelligence (AI) is a complex technology that has rapidly pervaded many aspects of our lives. Unsurprisingly, many aspects of AI have been described through a large range of metaphors. We speak of AI as a tool, an assistant, a brain, a parrot; we speak of it being capable of hallucinating and bullshitting.

Such metaphors may seem like convenient ways of explaining a complex technology. However, metaphors do more than make AI easier to understand: they influence how we imagine what AI is, what it can do and how we interact with it. They are powerful tools that deeply affect our approach to AI.

Thus, awareness of how metaphors work is important not only when communicating about AI, but also for public understanding, education, policy, and debates about the social implications of AI. Understanding how metaphors shape our thinking and talking about AI, as well as our uses and misuses of AI technologies, requires a systematic way to classify them, analyse their meanings, and investigate the assumptions they carry.

This is the reason why at We and AI we started to collect and analyse AI metaphors so that policy makers, educators, researchers, the general public, and anyone interested in AI and the metaphors through which we understand it can read and learn more about them. The metaphors we collect and analyse are presented is an ‘encylopedia’, where their meanings can be explored. The library includes an observatory, where metaphors of particular interest (and related ideas) are explored in greater detail.

Developing a framework for classification and analysis of AI metaphors

To support this project, we developed a framework for the classification and analysis of AI metaphors based on metaphors studies and literature about AI metaphors. The aim is to provide a robust theoretical lens, as well as a consistent and transparent way of organising and analysing metaphorical expressions of AI.

The framework draws on established research in metaphor theory and analysis, while bringing together classification approaches developed specifically for AI. Its theoretical foundations include Conceptual Metaphor Theory and Critical Metaphor Analysis. It also incorporates two approaches specifically relevant to the study of AI metaphors: the taxonomy of AI analogies and research on the spectrum of anthropomorphism in descriptions of AI.

This approach serves different purposes. Conceptual Metaphor Theory helps explain how metaphors structure our understanding of abstract concepts. Critical Metaphor Analysis extends the investigation by examining how metaphors relate to broader social, cultural, and political contexts. Meanwhile, AI-specific taxonomies help organise the different ways in which AI is described and its level of anthropomorphism (the attribution of human traits, emotions, intentions, or physical features to non-human entities). We focus on the issue of anthropomorphism in particular, because framing AI systems as humans is pervasive, but also dangerous. Seeing AI systems as human obscures the fact that they are algorithmic machines.

The framework used in this AI metaphor encyclopedia combines their complementary strengths in a practical model for the encyclopedia, bringing together three layers of investigation: classification, contextual information, and analytical interpretation.

How the AI metaphor framework works

The framework used in this AI Metaphor Encyclopedia looks at AI metaphors through three different lenses:

  • Classification;
  • Context and metadata; and
  • Analytical notes.

In this way, the framework enables a critical analysis of AI metaphors.

1. Classification: What kind of AI metaphor is it?

The first layer of investigation organises each metaphor according to two complementary types of classification: the aspect of AI that the metaphor describes and the degree to which it attributes human-like or other agent-like qualities to AI.

A. Classifying the family of the metaphor

The first classification method organises AI analogies into five categories: essence, operation, relation, function, and impact. The categories answer five different questions:

  1. Essence: What is AI? This category covers metaphors that present AI as a particular kind of thing or entity, such as a tool, a brain, or an organism;
  2. Operation: How does AI work? This includes metaphors that explain AI’s processes or mechanisms, for example, by comparing neural networks to brains or AI systems to stochastic parrots;
  3. Relation: How do we relate to AI? This category captures metaphors that frame the relationship between people and AI, such as AI as a colleague, assistant, companion, or adversary;
  4. Function: What does AI do, or what is it used for? This includes descriptions of AI as a tutor, translator, productivity tool, or creative partner; and
  5. Impact: What consequences can AI have? This category covers metaphors that frame AI’s effects, including descriptions of AI as a threat, a disruptive force, or a transformative engine.

The categories provide a common vocabulary for organising the family to which each AI metaphor belongs. The categories are best understood as possible perspectives rather than necessarily exclusive boxes. In fact, a single metaphor can belong to more than one family.

B. Classifying the degree of anthropomorphism

The second classification method examines how far a metaphor attributes human-like characteristics, agency, or capabilities to AI. It draws on approaches to AI anthropomorphism, specifically on a seven-level scale.

The scale ranges from descriptions that present AI as an ordinary object or instrument to those that portray it as possessing extraordinary, godlike, or mythical qualities:

  • Tool: AI is presented as an instrument, such as a calculator:
  • Infrastructure: AI is compared to a system or resource, such as a library or database:
  • Biological: AI is described through biological concepts, such as neural networks:
  • Animal: AI is framed through animal characteristics, as in the expression “stochastic parrot”:
  • Human-like: AI is presented as an assistant or helper:
  • Enhanced capabilities: AI is portrayed as having exceptional abilities, as in the metaphor of a mentor or genius: and
  • Godlike or mythical: AI is associated with superhuman, mythical, or potentially overwhelming powers, as in the Terminator or golem figures.

The value of this scale is that it distinguishes between metaphors that imply very different relations with AI and different levels of intensity in the personification of AI. Describing AI as a calculator does not carry the same implications and impact on people’s perception of AI as describing it as a genius or a godlike intelligence. This scale helps to identify the potential risks and implications of anthropomorphic AI metaphors.

The two classification layers work well together. While the taxonomy used identifies what aspect of AI the metaphor addresses, the anthropomorphism scale identifies how AI is represented in terms of human-like qualities and agency.

2. Context and metadata: Where does the metaphor come from and what do we know about it?

In addition to classification, it is important to understand the context in which the metaphor appears and all the relevant metadata we have and the details we know about it. This is why the second layer in the framework records contextual information. This includes the source in which the metaphor appears, the sector or setting in which it is used, its historical origins, and relevant citations or quotations.

For example, a metaphor may appear in a company’s description of its AI product, a government policy document, an educational resource, a news article or a research publication. These contexts matter because metaphors are used for different purposes in each context: to explain a technology, promote a product, establish a policy position, or help people make sense of an unfamiliar system. Therefore, recording a metaphor’s context makes it possible to trace a metaphor back to its source and understand the circumstances in which it is used. 

This layer of investigation in the framework is important for research transparency as it allows readers to check the evidence behind an interpretation.

3. Analytical notes: What does the metaphor emphasise and hide? Whose interests does it serve?

The third layer of investigation aims to examine the implications of each metaphor. Drawing on Conceptual Metaphor Theory and Critical Metaphor Analysis, it asks three central questions:

What does the metaphor highlight?

Metaphors direct attention towards particular characteristics of AI. The metaphor of AI as a tool, for example, emphasises instrumentality, practical utility, and the possibility of human control. The metaphor of AI as an assistant highlights support, collaboration, and interaction. Identifying these highlighted characteristics helps explain why a metaphor may be appealing or widely used in a particular context.

What does the metaphor hide or minimise?

Every metaphor foregrounds some features of what we want to explain while leaving others in the background. Describing AI as a tool, for instance, may understate the complexity of the systems involved, their dependence on data and infrastructure, or the organisational decisions that shape how they are developed and deployed. Depending on the context, a metaphor may also make issues such as surveillance, corporate power, or accountability less visible.

Whose interests does the metaphor serve?

Critical Metaphor Analysis focuses on investigating metaphors in relation to the social, cultural, and political contexts in which they circulate. In the case of AI, this means asking who uses a metaphor, what purpose it serves, and how it frames responsibility, authority, and control.

For example, a company that describes AI as a productivity tool may be emphasising efficiency and practical benefits. A critical analysis could investigate whether that framing also gives sufficient attention to the people affected by the technology, the conditions under which it is deployed, and the organisation’s responsibilities for its outcomes.

Asking whose interests a metaphor serves means treating the metaphor as part of a wider discourse and highlighting its effects, assumptions, and possible implications.

Analytical questions like these help make the encyclopedia more than a collection of AI metaphors, turning it into a resource for critical inquiry. They help readers understand not only what a metaphor is and what it means, but also how it frames AI and what may be at stake in that framing.

From metaphor identification to critical interpretation

The framework recognises the importance of distinguishing between metaphor identification and metaphor interpretation. Using both Conceptual Metaphor Theory and Critical Metaphor Analysis as theoretical lenses, it supports the act of interpreting the meanings of AI metaphors and explaining their significance within their social and cultural contexts. Interpreting AI metaphors in this way also includes addressing fundamental questions about the risks of AI use, such as cognitive surrender, cognitive offloading and AI psychosis, as well as the implications of hype and the lack of critical awareness in AI approaches.

The AI Metaphor Encyclopedia: a resource for understanding AI narratives

The purpose of the framework is to make the AI Metaphor Encyclopedia a catalogue of metaphors on AI through which readers can explore the language used to describe all aspects of AI, compare different representations, trace them to their sources, and investigate their wider implications.

Combining established metaphor research with AI-specific classification approaches, the framework provides a rigorous structure that, at the same time, aims to be practical and transparent. The classification supports consistency across the library; the contextual information provides the evidence needed for further investigation; and the analytical notes encourage critical engagement with the ways AI is represented.

Ultimately, the AI Metaphor Encyclopedia aims to help readers recognise that the language used to describe AI is never neutral. Metaphors offer ways of making AI understandable, but they also dramatically influence the questions we ask about it, the capabilities we attribute to it, and the responsibilities we assign to the people and institutions involved. By making these patterns visible, the framework supports a more informed, nuanced, and critically grounded understanding of AI.

References

Image by Harriet Humfress

RELATED METAPHORS