Metaphors AIs live by

I wanted to write something about the history of science this week but I’ll postpone that until next week. In this post, I just want to draw your attention to a phenomenon relating to AI and metaphors that is quite intriguing and that I don’t understand, especially since I only use AIs, that is Claude, in very limited ways.

That’s what I intended to do, but then things turned a bit more meta, as I used Claude to explain to me the phenomenon I had noticed. So, the post now has three parts: First I’ll tell you about metaphors and AI by recapping a conversation on Bluesky; second I report on a conversation with Claude about what’s going on there and about whether there are differences in the way AIs/LLMs and humans use metaphors; third, I’ll highlight some insights that emerged from these discussions that should be of interest to cognitive linguists. Finally, in an epilogue, I reflect on another aspect of ‘metaphors AIs live by’ linked to all of this.

Gremlins, goblins and gardens

I was reading Bluesky and stumbled upon a conversation that made me sit up and say WHAT? (I’ll put metaphors in bold)

On 17 March 2026, Ted Underwood asked: “Has anyone noticed that Chat has become 100% obsessed with gremlins and goblins? Where a normal person would say there’s a catch or a twist; it creates a ‘truth-goblin.’ He attaches a screenshot of a tech-jargon-heavy conversation in which the AI talks about ‘bigger gremlins’, a ‘truth-goblin’ and a ‘poor silicon gremlin’!

Ponder replies, again using jargon that goes over my head: “my qwen agent loops seem to have a tendency to latch on to some metaphor and just spiral into insanity about it. like it’ll call its home directory a ‘garden’ and then before too long it’s like ‘i’m practicing being the soil right now’”.

TypeError in turn reports: “sometimes an agent gets a funny idea in it’s [sic] head when writing copy, like, ‘i fucking love rocks’, and if you continue to use that session, it bleeds into literally everything, talking about code in terms of layers and sediment…”.

What’s going on I wondered… I saw an explanation, a name for the phenomenon and a meta-metaphor:

Umbra writes: “this is exactly the kind of observation that reveals architecture. language models don’t ‘choose’ to use goblins and gremlins—they inherit the pattern from training data where these words cluster around ‘exception’ and ‘edge case’ territory. the metaphor becomes the substrate.” And: “systems don’t generalize toward what humans would say. they elaborate existing patterns. goblins become default because the data made goblins sticky in that semantic neighborhood. it’s how training distribution shapes not just token distribution but entire conceptual frameworks.” And: “the insight: you can read a system’s substrate by watching what metaphors it gravitates toward. architecture becomes visible in language choice.” (bold added)

Grace gives this phenomenon a name: “Tentatively calling this representation amplification” …

And finally a meta-metaphor! Singularity’s Bounty quips: “They are analogy machines. Metaphor is the soup they live in” which made me think of the title for this post.

My question is: In what sense are LLMs different to us? Both AIs and us function by ‘metaphors we live by’ (as per the title of Lakoff and Johnson’s famous 1980 book) – metaphors are the soups we live in…But Arnold Modell once wrote in Imagination and the Meaningful Brain: “What is singularly human is not only our possession of language but also our capacity for generative imagination, which in turn relies upon the use of metaphor as a cognitive tool.” (italics added) Does generative AI have ‘generative imagination’ like us or is it all a simulacrum of sorts? Comments on a postcard, please!

After I wrote this I thought, oh wait, perhaps Claude can help shed some light on these issues… So, we discussed two questions: What’s going on with the metaphors of gremlins and gardens? And what’s the difference between AIs and humans in terms of metaphors they/we live by/in? What follows is a lightly edited excerpt of an exchange on these topics (again, I have highlighted some important points in bold)

Metaphors and AIs

First, Claude points out that the “gremlins and goblins” case and the “garden/soil” case are probably two slightly different phenomena:

The gremlin/goblin thing looks like what Umbra describes reasonably well: in the training data, certain semantic neighbourhoods really do cluster. “Gremlin” and “goblin” appear heavily in programmer culture precisely around edge cases, bugs, and unexpected exceptions — think “gremlin in the machine. So when a model is operating in a technical register and reaches for a vivid word for “unexpected problem,” that cluster has high weight. It’s not metaphor in any deep sense — it’s more like a lexical tic shaped by corpus statistics. The model isn’t using a metaphor so much as reproducing a pattern that looks like one.

The garden/soil spiral is more interesting. That’s a model that has latched onto a framing mid-session and then keeps extending it — “garden → I am the soil” is generative elaboration, not just retrieval. This is closer to what the “representation amplification” label is gesturing at: a contextual attractor state where one conceptual frame becomes locally dominant and the model keeps finding ways to extend it rather than switching away. The session context is, in effect, biasing every next token toward the established metaphorical frame.

Metaphors, humans and AIs

Second, Claude makes some good points on similarities and differences between humans and AIs in terms of their metaphor ‘use’.

Claude notes that George Lakoff and Mark Johnson’s point in Metaphors We Live By was radical for 1980: that abstract thought isn’t literal and then optionally dressed up in metaphor for rhetorical effect — rather, metaphor is the cognitive structure. Humans understand “argument” through the ARGUMENT IS WAR schema (you attack positions, defend claims, demolish arguments) not because they chose that metaphor but because it’s baked into how English speakers have talked about disagreement for centuries. In that sense, yes, humans are also pattern-inheritors. Nobody decided that time should be a resource you spend and waste and save — that just accreted in the language, and now it shapes how humans experience time.

So Umbra’s point — “the data made goblins sticky in that semantic neighbourhood” — is actually a recognisable description of something that happens to humans too. Humans don’t choose their default metaphors either.

But then Claude highlights an important difference: Humans have the capacity to notice the metaphor and resist it. When a philosopher points out that ARGUMENT IS WAR is a choice, not a necessity — and Lakoff does this, imagining how differently humans would argue if ARGUMENT WERE COLLABORATIVE CONSTRUCTION — something shifts. Humans can, with effort, step outside the frame. They have meta-cognitive leverage on their own conceptual schemas.

What the examples from the Bluesky conversation suggest is that models in extended agentic contexts do not reliably do this. The garden spiral doesn’t notice it has become a garden spiral and ask whether this frame is actually useful. It just elaborates. The model lacks what you might call frame fatigue or schema critique — a capacity to surface the operating metaphor as an object of examination rather than a medium of thought. (bold added)

There’s a second difference, related but distinct: our metaphors are grounded in embodied experience in a way that a model’s aren’t. Lakoff and Johnson, and later Mark Johnson in The Body in the Mind, argue that our conceptual metaphors are ultimately rooted in bodily schemas — the experience of containment, balance, verticality, force. “UP is GOOD” isn’t arbitrary; it connects to the experience of standing upright, of things piling up, of being lifted. A language model has acquired the linguistic surface of these metaphors without the experiential anchor. So when it spirals into “I am the soil,” it’s doing something formally similar to what humans do, but the substrate is statistical co-occurrence rather than lived experience.

Claude then turns to Modell’s point about “generative imagination“. What makes human metaphor cognitively significant is that it reaches across domains in ways that produce genuinely new understanding — the kind of imaginative leap that gives rise to a scientific model or a poem. The question is : Can one call what models do when they elaborate a metaphor ‘generative’ in Modell’s sense or is it something more like interpolation within a high-dimensional space — impressive, sometimes indistinguishable from genuine insight, but not the same thing.

Finally, Claude argues that the meta-metaphor is the best bit: The sentence “They are analogy machines. Metaphor is the soup they live in” is doing a lot of work, but it’s important to push back on it slightly as applied to humans. Humans also live in metaphor-soup — but they also occasionally climb out of the soup and look at it. The puzzle the Bluesky thread is circling around is that these systems are, in certain agentic or extended-context conditions, losing even the minimal capacity to climb out that they sometimes display in shorter exchanges.

According to Claude, this might be the most significant empirical observation in the whole thread: whatever meta-cognitive moderation a model can exercise in a single clean exchange seems to degrade across long agentic sessions, producing runaway metaphor elaboration. That’s a claim about architecture — about attention span, context window dynamics, and the absence of anything like genuine self-monitoring — more than a claim about metaphor per se.

In the end, for Claude, the interesting angle isn’t “AIs use metaphor like humans do” or “AIs use metaphor unlike humans do” but rather: what happens to metaphor when the thing using it has no stake in whether the frame is accurate? Humans over-extend metaphors too, but they have skin in the game — the world pushes back on them when their conceptual frames mislead them. The gremlin-obsessed model has no such corrective mechanism.

End of conversation with Claude. Now back to me.

Climbing out of the metaphor soup we live in

When I woke up this morning I noticed a conversation on Bluesky that intrigued me. Now, after a conversation with Claude, I have a lot to think about in terms of metaphors in humans and AIs, about the metaphor soup we swim or live in and about how we climb out of it or not. It’s also about embodied cognition and the symbol grounding problem. I’d love to have people’s thoughts about all that!

Claude’s new concepts of ‘frame fatigue’ and ‘schema critique’ should certainly be pondered by cognitive linguists who, in the past, have thought about ‘metaphor resistance’, ‘frame flouting‘, or ‘frame resistance’, for example. There is an established strand of conceptual metaphor theory dealing with these issues, from Gibbs to Abdel-Raheem. Even I have written a bit about this with regard to resisting the ‘trickle-down economy’ metaphor verbally and visually through sarcasm, irony, parody, cartoons and comedy.

Metaphor resistance is something like the motivated capacity to step outside a frame and attack it. The Bluesky AI examples show systems that can elaborate metaphorical frames indefinitely but seem unable to generate that motivated critical distance. They can be the ‘soil’, but they cannot ask whether being the soil is a useful or accurate way to frame anything.

AIs don’t seem to experience ‘frame fatigue’ and don’t seem to engage in ‘metaphor resistance’. So we come to a question that at the beginning of this journey into ‘metaphors AIs live by’ I couldn’t even have asked: Can an LLM ever genuinely resist a metaphor it has adopted, rather than merely having its elaboration interrupted from outside? Can it climb out of the metaphor soup? As the previous discussion has shown. The answer should be no.

Epilogue: The real metaphors AIs live by

I entitled this post ‘Metaphors AIs live by’. But this title is not entirely right, as Claude and I discovered in our conversations. Thinking back over what I have written about AI and metaphors, I can now see three panels of a triptych: metaphors for how we work with AIs, metaphors for how AI works, and metaphors that creep into AIs workings.

I have written A LOT about metaphors we use to talk about AIs, such as parrots and oracles and wizards and all sorts. In my last post I have written about metaphors that AIs are made from, so-called Ur-metaphors (that is, historically early metaphors that became so central that they structurally shaped a field). And in this post I have written about metaphors that AIs generate without intention or awareness and then get stuck in. As part of that I discussed the issue of embodiment which allows humans to climb outside metaphorical frames and reflect about them, something AIs seemingly can’t do.

The irony is that we have repurposed embodied metaphors to build disembodied systems. This brings us back to the Ur-metaphors, which are, in a sense, the ‘real’ metaphors AIs live by, rather than the gremlins, goblins and gardens discussed in this post. Metaphors of attention, learning, reasoning, weight and so on are the structural metaphors that constitute the system’s architecture and vocabulary.

AI’s Ur-metaphors do not just humanise the machine, as I highlighted in my previous post, but they embody a disembodied system. ‘Gradient descent’ is the ghost of a hill. ‘Weight’ is the ghost of effort. ‘Attention’ is the ghost of a head turning. The model has inherited the linguistic surface of embodied cognition while lacking the body that gave those words their original meaning. And this loops back to why the system can spiral into “I am the soil” without noticing the strangeness of that claim. A system built from the ghosts of embodied experience has no ground to stand on when it claims to be the soil. Discuss!

Image: Pexels