AI and deconstruction

I recall the early days of the world-wide web when scholars recruited the philosophy of Jacques Derrida to show that the endlessly interlinked web expressed how language actually works — like hypertext. I’m thinking of the writing of George Landow (“Hypertext as collage writing”) and Jay Bolter (“Topographic writing: Hypertext and the electronic writing space”) in the 1990s.

I followed a similar tactic in my previous posts recruiting Derrida to help explain the apparent success of large language models (LLMs). Vast repositories of written texts (i.e. digital texts) processed through well designed neural network models are sufficient to form the basis of conversational exchanges between humans and machines. It appears that human sociability and cognition can rely on text after all. In the sense that Derrida meant it, writing precedes speech.

I want eventually to critique the use of (radical) philosophy to vindicate the development and application of technologies. But that’s for a later discussion.

In the mean time I will expand on claims about the positive ways that AI and the philosophy of Derrida seem to converge. ChatGPT helped me recall these aspects of Derrida’s philosophy, but the words are my mostly own.

Intertextuality

As mentioned in relation to hypertext, Derrida developed the idea intertextuality—the notion that texts are interconnected and derive meaning from their relationships with other texts. That resonates with how LLMs function. These models learn from vast corpora of interconnected texts, generating responses based on patterns and contexts within this web of texts.

As indicated in previous posts, processed text (written typed, digital, etc) persists longer than speech (i.e. simply talking). But a reading of Derrida emphasises that writing is not just a record of speech; it’s a fundamental component of how we structure and understand the world. This reflects the implication rom Derrida that writing itself is imbued with the complexities of human life.

Différance

During a conversational session, ChatGPT presented a linkage I had previously not thought much about.

Derrida’s concept of “différance”—the idea that meaning is always deferred and differs depending on context—can be seen in how LLMs process and generate text. The models do not produce fixed meanings but generate responses that vary based on the input context, reflecting the fluid and dynamic nature of language that Derrida described.

As I have discussed elsewhere, LLMs typically deploy “semantic embeddings,” multidimensional numerical vectors attached to words and tokens. These vectors are derived from the contexts of words and tokens in the corpus, but are inscrutable as repositories of definitions and meanings. The instantiations of words and tokens as meaningful entities only emerges in the contexts of use as the model generates its sentences, word by word. Meanings are on hold until the words and tokens start to appear in the generated text. Other meanings emerge depending on contexts. LLMs make explicit the temporal deferral of meaning in language in general.

Authorship

Derrida’s emphasis on interlinked texts challenges traditional notions of authorship and authority. Any author builds on the writings of others. In the company of other Structuralists and Poststrucuralists, Derrida opened up the contingent nature of originality, creativity, authenticity, and authority in relation to the claims and responsibilities of authors.

With their dependence on preexisting and multiply-sourced training texts, LLMs renew the questions of authorship, amplified through recent high profile ethical and legal questions and challenges.

Active reading

Derrida emphasized the active role of the reader in this textual play. Similarly, the interaction between users and LLMs involves active interpretation and engagement. Users guide the models through prompts and questions, influencing the generated responses and co-creating the conversational interaction.

Readers share in the role of authoring the meaning of the texts they read. Again, LLMs make this shared role explicit in the way they operate.

The processes by which Derrida analyses authoritative texts and theories in order to unsettle their claims is often referred to as “deconstruction.” See post Deconstruct that!

Bibliography

• Bolter, Jay David. “Topographic writing: Hypertext and the electronic writing space.” In Hypermedia and Literary Studies, edited by Paul Delany, and George P. Landow, 105-118. Cambridge, MA: MIT Press, 1994.
• Kuchtová, Alžbeta. “The Incalculability of the Generated Text.” Philosophy & Technology (2024) 37:25 37, no. 25 (2024): 1-20. 
• Landow, George P. “Hypertext as collage-writing.” In Hypermedia and Literary Studies, edited by Paul Delany, and George P. Landow, 150-170. Cambridge, MA: MIT Press, 1994.
• Landow, George P., and Paul Delany. “Hypertext, hypermedia and literary studies: the state of the art.” In Hypermedia and Literary Studies, edited by Paul Delany, and George P. Landow, 3-50. Cambridge, MA: MIT Press, 1994.
• Lunefeld, Peter (ed.). The Digital Dialectic: New Essays on New Media. Cambridge Mass.: MIT Press, 1999.