Must AI speak?

AI is subject to critique from many quarters. See posts Generalised AI as existential threat, and Learn to talk to your AI. We don’t need to draw on philosopher Jacques Derrida to articulate the limits and challenges of AI, but here I want to identify a Derridean-style argument that impinges on my thinking about AI-based large language models (LLMs).

Criticising AI is a bit like criticising a bridge, or any other introduction of an advanced technological artefact into the world. You can examine the degree of technical competence that has gone into its design and construction; you can consider how it intervenes into politics, social organisation and ethics. (Should the bridge be built?) You can also examine what people write and say about the bridge: the claims people make about it; how it features in language as symbol, object of significance, focus for argumentation, use as metaphor, etc.

These facets of technology (its instrumental, socio-political-ethical and discursive dimensions) conflate under the rubric of interpretation. See post: Interpretation by design. That said, a Derridean-informed critique of AI is likely to have most impact on the way we talk about AI, rather than its strengths or weaknesses as a technology, or even the ethical questions it raises.

AI that listens

The robots in the SciFi novels of Isaac Asimov talked — along with disembodied AIs as in the case of HAL in 2001 A Space Odyssey, and the movie Her. Direct vocal conversation serves as an indicator of intelligence — more so than pressing buttons or typing text.

In fact, by most accounts, voice activation is not the most convenient means of interacting with a computer or other automaton. But voice appears “most natural” to us, as something you should be able to do if we are to connect with synthetic intelligence.

As depicted in popular culture, and hence as a means of marketing the technology to a wide constituency, people expect AI to be something you can talk to, i.e. that listens and speaks.

Alan Turing’s test (the imitation game) required that you should be able to communicate with a computer via text, as if via a typewriter or teleprinter. Text preserved anonymity in his hypothetical experiment and left the human-or-machine question as undetermined without analysing what was written, i.e. its “content.”

The quest for conversational AI, as in the use of voice activated systems like Siri and Alexa. reinforced further with OpenAI’s Voice Chat indicates a commitment amongst developers and consumers to speech as immediate, close to thought, a sign of intelligence and even of sentience.

Speech authenticates

In so far as AI gives this kind of priority to speech as authenticator of intelligence, it supports what Derrida and others have termed “the metaphysics of presence.” I draw here on the online Encyclopaedia Britannica for a definition:

This is the tendency to conceive fundamental philosophical concepts such as truth, reality, and being in terms of ideas such as presence, essence, identity, and origin—and in the process to ignore the crucial role of absence and difference.

In the AI context, I take that to mean that we tend to value quick, immediate, positive interactions, exemplified by spontaneous speech, above hesitant, conditional, uncertain, “deferred” interaction. ChatGPT told me: “[Derrida] argues that meaning is always deferred and contextual, challenging AI’s claim to fully understand and generate human-like language.” See post Can an AI only think “fast”?

Note

• Featured image generated by WordPress generative image facility. Prompted by the content of this post and the prompt: “No human or animal figures. Steam punk grunge aesthetic. Photoreal. Old style telephones.”