When in Doubt, Go to the Beach

Hannes Bajohr

27 June 2023

Steven Knapp and Walter Benn Michaels’s thought experiment of the wave poem aims at dissolving the inferential connection between meaning and intention, which they dub “theory.” If you discover lines in the beach that form a poem, you may infer an author behind them. Only once you realize that no human has produced the text—if after a wave washes it away, another poem appears—do you begin to doubt this relationship. You can either accept the sea as the author and become an animist or admit that natural forces have created it and that there is no author. But then there is no meaning either, Knapp and Michaels hold, as meaning and intention are one—no additional theory necessary. Assigning meaning to lines in the sand is just like seeing the face of Jesus in a piece of burnt toast: a case of pareidolia at best, at worst the mistaken belief in a communication from God.

The case against “intentionless meaning” seems at first blush highly relevant for current discussions about large language models (LLMs). The authors even apply their argument to computers: unless such machines are truly capable of intentions, any text they produce remains without meaning. But this parallelism is only superficial. In reality, Knapp and Michaels have little to say about the meaning of computerized language. This becomes obvious when looking at a recent argument against meaning in LLMs, Emily Bender and Alexander Kollers 2020 essay “Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data” (which is the basis of a section of the much more famous “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?🦜” ).

They, too, go to the beach. In their thought experiment, an intelligent octopus taps into the telegraph communication between two castaways on two desert islands. While the octopus could learn statistically likely sequences of words that appear meaningful to the two humans, they will not be grounded in any “communicative intent.” This is evident once the octopus tries to follow instructions: it must fail for it cannot refer to anything outside the symbol system. Accordingly, Bender/Koller define meaning as “relation between a linguistic form and communicative intent” via the expression M ⊆ E × I—meaning means “retrieving” an intention i from an expression e. But this retrieval is again an inferential connection. This approach, then, requires what Knapp and Michaels disapprovingly describe as “grounding meaning in intention” and what they reject as “theory.”

In fact, it is the very absence of theory—the prohibition on a hiatus between meaning and intention—that renders Knapp and Michaels’s essay much less relevant to the discussion of natural language processing than Bender and Koller’s. The former use the term simply as a binary—there is meaning or there is not—and the only way machines can produce it is for them to have intentions. That this leaves much to be conceptually desired is already clear from the fact that a wave poem and computer text result from radically different processes (from a chaotic and a deterministic system respectively). While we can produce computer text reliably and repeatedly, the wave poem may occur once in the course of a million years or never at all. Why should the same binary apply to both? And even if we grant that LLMs do not operate with meaning in the full sense, what keeps us from using the term more flexibly and ask for shades of meaning? Indeed, this is only possible if meaning is not identical with intention—in other words, if the hiatus between the terms is mediated by theory.

Bender and Koller admit such shades of meaning. Multimodal models, which connect text to images, allow for what they call a “weak” grounding of language in other data types. This may not be meaning proper, but it is not nonmeaning either. Such “dumb meaning,” as I call it, is the basis for something we now do all the time: communicating with machines as if they had intentions, for only then can we make them do what we want. Simone Natale, in his book Deceitful Media (2021), calls this phenomenon “banal deception.” It is “banal” because we are fully aware of its fictive character: my conversation with an AI assistant only works if I treat it as a communication partner while knowing that it is not. Intention, thus, can come in a fictitious modality, operating with shades of meaning, and be neither just illusory pareidolia nor the mistaken belief in sentient machines. Rather, such a fiction is a perfectly legitimate pragmatic prerequisite for successfully interacting with machines. By insisting on meaning as binary and coeval with intention, Knapp and Michaels cannot account for what is already happening; and so we need theory.

Let us end where we began, at the beach. Qui t’a faite? pensai-je—“Who made thee? I pondered,” Socrates recalls his stupefaction in Paul Valéry’s dialogue “Eupalinos,” when, during a walk on the shore, he happens upon a wholly ambiguous object. Neither clearly human-made nor natural, it does not fit his ontology. Unable to bear this ambiguity, he flings it back into the sea. Knapp and Michael, too, cannot bear the ambiguous object that is artificial text. That makes “Against Theory” irrelevant for us. For we may be faced with a future of ubiquitous LLMs in which the doubt about the origin of a text—was it written by a human or by a machine?—becomes the norm rather than the exception. Knapp and Michael’s wave poem still expresses an older standard expectation towards unknown writing, namely that in reading it “we must already have posited a speaker and hence an intention.” This standard may be crumbling as artificial texts proliferate and are at the same time not treated as meaningless. It is perfectly imaginable that we might one day find ourselves taking what I call a “post-artificial” stance toward unknown text: that we simply suspend the question of intention and read it as meaningful nevertheless—easily bearing the ambiguity about the origin of a text.

---

Hannes Bajohr is junior fellow at Zurich’s Collegium Helveticum. He is a coeditor of History, Metaphor, Fables: A Hans Blumenberg Reader (2020) as well as the author of Schreibenlassen: Texte zur Literatur im Digitalen (2022).