Evolving Minds,
Evolving Language

The rise of anthropomorphic language

• AnthroScore (Cheng et al. 2024): anthropomorphic language in CL & CS papers has risen steadily over 15 years

A field divided

• When asked whether AI understands, AI researchers were roughly split 50/50 on this issue according to a 2022 survey
• Inflationists are happy to say that these systems really do understand and reason
• Deflationists take the cognitive vocabulary for an illegitimate anthropomorphism

Reframe how we think about
anthropomorphic locutions

Look at the rise of anthropomorphic language through the lens of metaphor. There is a transfer going on — our ways of thinking about biological minds are being creatively extended to artificial minds.

These metaphors illustrate something fundamental about how language evolves.

Metaphor as conceptual adaptation

• not merely playful or poetic
• a community adjusting its expressive resources
• the words we talk with and the concepts we think with interact: as we apply words in new ways, we gradually adjust our conceptual repertoire

Three questions that will structure this talk

1. 1What gets transferred when a term is metaphorically extended to a new domain?
2. 2What guides the transfer — what makes it non-arbitrary?
3. 3Does it track anything real — does the world sustain the transfer?

Seeing a dog as a dog

• Every dog presents a different sensory profile — yet you see each one as a dog
• What bridges the gap between the sensory particular and the general concept?

DOG

Imagination

• not fantasising; the capacity to bring a concept to bear on a particular
• Kant: imagination is “a necessary ingredient of perception itself”

Perception is saturated with thought

• P.F. Strawson: “past and possible perceptions are alive in present perception”
• Wittgenstein: “the echo of a thought in sight”

From routine to creative concept application

• If even routine concept application involves imagination, then creative concept application involves the same faculty more inventively deployed
• “dog runs”→“cat runs”→“water runs”
• Extension not covered by rules, yet humanly graspable
• Goes against tradition seeking to exhaustively explain language in terms of general rules

Creativity precisely comes in where the rules give out.

The continuity thesis

Seeing a neural network as understanding

Metaphors produce a framing effect

Metaphorically extending a term to X induces us to see or consider X in a new light (Davidson, Moran)

Don’t look for special metaphorical meanings of words at the level of language; focus on how metaphorical usage channels our attention directed at the world

But Davidson leaves the framing effect as a black box — an unstructured “noticing of similarities.”

The bowtie structure of concepts

Metaphor opens the bowtie

When we extend a concept to a new domain, we do not transfer this bowtie structure wholesale. The metaphor invites us to explore which inferential connections hold in the new domain and which do not.

Adaptation as selective inferential transfer

• Referent changes. What gets transferred are inferential connections
• Conceptual adaptation is the work of selectively preserving the inferential connections that fit, severing those that don’t, and potentially forging new ones

The shape of metaphor processing

Metaphors undergo a career

Sleeping metaphors

Cornelia Müller (2008): metaphors don't die; they oscillate between states of waking and sleeping. Sleeping metaphors can be reawakened (Goldstein, Arzouan, and Faust 2012).

Diagnosing the debate

Neither side recognises the creative metaphorical extension.

The bind

Our existing concept of understanding carries implications that do not fit AI—of consciousness or responsibility, for example.

This means that our cognitive vocabulary is simultaneously

People resist calling it “reasoning” because it’s not human reasoning. But I think there is no way we can not call it reasoning.

The reintroduction argument

Meeting conceptual needs

Conceptual needs: two roles

1. I. Explain metaphorical usage:
   Researchers reach for cognitive vocabulary not out of carelessness, but in answer to real practical pressures
2. II. Guide the extension:
   The conceptual adaptation realized by the metaphorical usage should be responsive to our conceptual needs

Grokking

Robert Heinlein’s Stranger in a Strange Land describes Martians as capable of achieving an alien mode of deep understanding

• The source concept being drawn on to make sense of this new phenomenon is that expressed by our term “understanding”
• Performs a lexical marking of the strangeness of the understanding involved
• The lexical marking keeps the metaphor awake: holds open the tension between the understanding-frame and the alienness-frame

Modular addition: addition with a ceiling

• Like adding days of the week: after the seventh day, we start over
• More precisely: if the modulus is 113, then 55 + 71 is not 126 but 13 — because 126 overshoots 113 by 13
• We do it in two steps: we first add the numbers and then divide by the modulus. The remainder is our answer

The model was never told this method. It was just shown thousands of solved examples.

What gets transferred?

A selective subset of the concept’s inferential connections. Metaphors open up the bowtie structure of concepts and let us explore which connections hold in the new domain and which do not.

What guides the transfer?

Conceptual needs: our cognitive concepts perform work that we need to see performed in human-AI interactions – for example, the concept of understanding guides how we place our trust.

But our cognitive concepts also carry commitments that don’t map onto machines. This puts us in a bind calling for conceptual adaptation. Our conceptual needs drive and discipline that adaptation.

Does it track anything real?

The case of grokking showed that it can. When a model groks, it ceases to rely on rote memorisation and forms a compact computational circuit encoding a general principle.

But this is a deeply different form of understanding – understanding without consciousness

Conclusion

What looks like a terminological dispute is really a struggle over which inferential connections to endorse — a struggle whose consequences reach into hospitals, courtrooms, and regulatory agencies.

Metaphorical extension is not ornamental. It is a vital mechanism by which a linguistic community reshapes its expressive resources for a world that keeps outrunning them.