Metaphors present one thing in terms of another. In doing so, they enable us to understand an unfamiliar or complex subject by viewing it through the lens of something more familiar. They are cognitive tools that show us how to think about a subject, going beyond anything we could merely say about it.
When we call a large language model “fancy autocomplete”, a “blurry JPEG of the web”, or a “digital brain that hallucinates”, are we merely decorating our speech? Or are these metaphors actively shaping the way researchers, policymakers, and ordinary users think about, interact with, and regulate artificial systems?
Key Points:
- Metaphors are thinking tools, not mere decoration. A metaphor works as a frame that packages a way of seeing a subject into a compelling image or phrase.
- AI metaphors highlight and hide. Framing an LLM as “fancy autocomplete” highlights mechanical prediction and hides the capacity for novel outputs; framing it as an “alien” highlights inscrutability and hides the mundane statistical mechanics underneath.
- The central question is whether a metaphor is apt, that is whether it illuminates features of the target in a way that provides significant value. Good design, policy and communication should involve care and reflection on the costs and benefits of the use of AI metaphors.
1. The power of metaphor
Consider describing an LLM as “fancy autocomplete”. It evokes a familiar image of a candidate for the next word popping up in response to what you have typed. That image engenders a way of thinking about LLMs: a tendency to notice certain features of LLMs, the mechanical nature of their prediction, their passivity and automaticity. Such elements become central, while the system’s capacity to produce extended, structurally complex, and semantically nuanced text recedes from view.
In these respects, metaphors do more than state facts. They serve as frames that activate perspectives, which affect how we characterise that which we think about under the guise of the metaphor. When someone says “ChatGPT is a blurry JPEG of the web”, the literal content is false, yet it frames the subject in a way that highlights something true: that LLMs perform lossy compression. This in turn evokes a more general perspective on LLM outputs as degraded copies. It casts a feeling of imperfection and untrustworthiness over everything the model produces. The “blurriness” makes the gap between input and output prominent, while the possibility that lossy compression might itself be a form of genuine learning drops out of sight.
2. Varieties of metaphor
AI metaphors cluster into recognisable families.
Mimicry metaphors (“stochastic parrot”, “blurry JPEG”) highlight the derivative quality of LLM outputs.
Similarly, tool metaphors (“glorified auto-complete”, “calculator for words”) deflate AI toward the mechanical. This is useful as a corrective to anthropomorphism, but hides the fact that LLMs are typically non-deterministic in operation and can generate text that departs substantially from the prompt.
By contrast, agential metaphors (“co-pilot”) imbue artificial systems with social roles and quasi-human capacities, trading on root metaphors like “intelligence” and “learning” so embedded in AI discourse as to be nearly invisible.
Inscrutability metaphors (“alien”, “delphic oracle”, “superintelligence”) foreground power and mystery and obscure the mundane engineering and its commonalities with other technologies which are not similarly revered.
And famously, metaphors make good insults. Along these lines are the more derogatory metaphors such as pollution metaphors (“AI slop”) that frame artificial systems through their harms to our information environment, and power metaphors (“colonising loudspeaker”) that foreground the inequity inherent in outputs reproducing dominant cultural perspectives while marginalising others.
3. Apt and inapt metaphors
An apt metaphor is one whose perspective tracks something real and important about the subject. Its way of making features prominent and central corresponds, at least roughly, to the way those features actually matter for some purpose. An inapt metaphor, by contrast, invites a perspective that distorts: it makes the wrong things salient, buries what matters, or generates inferences that lead us astray.
Consider the “blurry JPEG” metaphor. The perspective this metaphor induces may be quite appropriate for many uses. Clearly, sometimes LLM outputs should be treated as degraded rather than authoritative. To that extent, the metaphor invites useful caution. But the same metaphor also makes invisible the possibility that lossy compression might produce something genuinely new. The “blurriness” might sometimes be more like abstraction or generalisation than degradation. In that respect, the metaphor is inapt.
Now consider the “assistant” metaphor. Its perspective may be apt enough when it comes to describing the user interface and the typical interaction pattern in which the system simulates a helpful, cooperative partner. But it makes invisible the absence of any capacity for commitment or loyalty on the system’s part. Users who internalise the assistant frame may over-trust the system, delegate inappropriately, or fail to check outputs, and suffer harm as a consequence.
The point here is that the value of metaphorical framing turns on aptness. A metaphor that makes you see an artificial system in a way that corresponds to its actual structure, capabilities, and limitations is doing something valuable. And this is the case even if its perspectival effects are hard to undo. A metaphor that makes you see it in a way that systematically distorts those things is doing something harmful, precisely because its effects are hard to undo.
4. Metaphor and interface design
Metaphors don’t just describe AI systems — they shape how those systems are built and encountered. Design choices are saturated with metaphorical framing, often so deeply embedded that the framing goes unnoticed. The chat interface, the conversational tone, the blinking cursor waiting for your reply: all of these encode the “assistant” metaphor into the product itself, long before any user consciously adopts that frame.
This matters because the way a system is presented to users constitutes their expectations of how it works, what outputs to trust, and what tasks to delegate to it. When those expectations align with the system’s actual capabilities and limitations, users make better decisions. When they don’t, the costs are real: over-reliance, misuse, disappointment, and erosion of trust — outcomes that harm users and undermine the product alike.
This is hardly unique to AI. Video game players have long understood that “playing against the computer” does not mean playing against another person. The lesson generalises: if metaphorical framing can be apt or inapt, then design choices that embed such framing can be evaluated for accuracy. In principle, we should be able to measure whether an interface’s implicit metaphors lead users toward or away from a realistic understanding of the system they are using.