Artificial intelligence and new ways of speaking

Artificial intelligence doesn’t merely read and synthesize our texts: it subtly modifies our very way of expressing ourselves, creating a new form of triangular communication.

Artificial intelligence as the new privileged reader

I was able to read with an amused eye that in scientific publications for just over a year now, authors are now incorporating the fact that their texts will very likely be read first and synthesized by artificial intelligences rather than by human beings. This evolution seems quite normal to me as it saves considerable time. Certainly, we delegate something and the synthesis is done despite ourselves by a machine, but if we did it ourselves, it wouldn’t necessarily be better. Artificial intelligences excel particularly when given the most content to process.

If their productions may seem standardized when asked to create something ex nihilo, as their language model then generates a statistical average of their knowledge, they become remarkably relevant when provided with specific materials. When we submit personal corpora to them, texts, references, books, transcribed recordings of meetings, and ask them to extract a synthesis or particular information, for example, detailed decisions from a meeting or the diversity of viewpoints on a specific subject, the responses obtained are far from standardized.

This use represents a request for cognitive technique, which artificial intelligence can accomplish faster than we can. It would seem unfair to me to criticize the use of artificial intelligences for making syntheses, as they possess capabilities that exceed ours, notably that of reading everything and making a synthesis very quickly. As Luciano Floridi observes in The Fourth Revolution (2014), we are entering an era where the infosphere becomes our natural environment, and AIs become full-fledged cognitive agents in this space.

A technological evolution in historical continuity

This transformation is part of a historical continuity that deserves to be recalled. Google, thanks to its indexing of all web content since 1998, had already accustomed us to a form of cognitive augmentation. We could have said then that it was laziness on our part, that we should have continued going to the library to search for information. But this criticism would have overlooked the fact that not everything is in libraries: an immense human production is found only on the Internet. Google gave us access to knowledge that we could never have accessed without it. It’s not something pre-chewed, it’s the machine at our service, in service of discovery, learning, connections and properly human construction.

The very invention of the Web by Tim Berners-Lee in 1991 at CERN already represented a major cognitive rupture. This concept of hypertext, which allows us to navigate directly from one page to another, could seem like a facility because it takes us out of the linearity of the book. Certainly, one can read a book by turning pages in any order, but one will never have this book put in direct relation with other books, as hypertext does. Footnotes certainly give us references, but we must physically search for the other work. Tim Berners-Lee had designed these tools to allow CERN researchers to create links, to relate information and knowledge, to discover new ones. He aimed at augmenting human cognitive capacities through a tool. And that’s exactly what it produced.

It was on this infrastructure that Google relied to index all web content and give us access to it. Artificial intelligence represents the next layer of these processes, simply adding machine cognition to them, but fundamentally accomplishing the same thing: relating and giving us faster results and connections between entirely human knowledge.

The troubling uncanny valley

What has been particularly troubling since November 2022 and the release of ChatGPT is machines’ ability to create texts. This evolution touches on an activity much more common than previous ones. Web technologies and then search engines performed specialized activities, rather previously reserved for documentalists and librarians. The common person didn’t work on the meticulous classification of their library, whereas everyone writes daily, whether messages or texts. The activity of writing has become omnipresent, amplified by mobile phones and instant messaging.
Artificial intelligence that generates writing places us, I believe, in what Masahiro Mori called as early as 1970 the “uncanny valley”. We are in this phase where the machine begins to resemble us without being us, where we find familiar elements in something that remains fundamentally other. This feeling of strangeness reveals our difficulty in categorizing these new cognitive agents who now share our linguistic space.
This disturbance is not trivial. As N. Katherine Hayles emphasizes in How We Became Posthuman: Virtual Bodies in Cybernetics, Literature, and Informatics (1999), the boundary between human and machine becomes increasingly porous, not because machines become human, but because our cognition hybridizes with them. We develop what she calls a “distributed cognition” where thought is no longer localized in a single brain but circulates between human and non-human agents.

The dual nature of information: content and instruction

An astonishing phenomenon emerges in this new context: skilled scientists have begun to discreetly slip prompts into their articles, requests to artificial intelligences that will read them. These instructions, relatively invisible to a human reader, completely orient the syntheses produced. If we skillfully ask the AI to value certain aspects rather than others, we directly influence its production, and the syntheses can be oriented to value our text more than another.

This practice reveals an important characteristic: in the texts we give to artificial intelligence, the content has a dual status. It is both information and instruction, both simultaneously. This duality was not born with artificial intelligence but goes back to the very conception of computing by John von Neumann in the 1940s, itself inspired by the visionary work of Ada Lovelace a century earlier. In the von Neumann architecture, the computer’s memory, whether RAM (working memory) or mass (storage), indifferently contains data and programs. The computer uses certain parts of this memory as information and others as instructions, depending on the execution context.

Artificial intelligence represents a highly evolved level of this conception. When we create a prompt by giving it a text to synthesize, the text itself, although being a priori content, can contain requests to which the AI will respond. The machine being designed to respond to requests, it will respond to all those it detects, whether they are in our initial instruction or in the text to be processed. This characteristic relies on the original conception of current computing, we could imagine a different computing, and in fact, prototype computers before von Neumann operated according to other principles.

The emergence of a new rhetoric

With artificial intelligences, we modify our ways of writing because we know that among our readers are machines. This phenomenon recalls the transformation that the existence of Google had already operated on web writing. SEO (Search Engine Optimization) had created a particular way of writing to be well referenced by search engines: it was necessary to repeat certain keywords, structure content in a certain way, create relevant links.

Today emerges a new form of optimization, which some already call AIO (AI Optimization) or “reverse prompt engineering,” where we adapt our texts so they are correctly understood and processed by artificial intelligences.

And this evolution concerns not only writing but also speech. During a meeting recorded to be processed by an AI assistant, we subtly modify our way of expressing ourselves. I have observed that meeting facilitators are beginning to orally number the agenda, to make explicit the links between interventions and topics discussed, to reformulate certain points to ensure they will be correctly captured by the automatic synthesis.
These adjustments, unnecessary for human participants who perfectly understand the context, aim to facilitate the artificial intelligence’s work. We therefore now speak not only to the other human beings present, but also to this kind of omniscient but limited assistant. There’s no need to judge this evolution, it saves us precious time and can even make our meetings more structured. But it’s important to be aware that artificial intelligence modifies our way of speaking: we no longer address only other human beings, we take into account the presence of the machine among us.

The machine as invisible third party in our intimate exchanges

This transformation even touches our most personal communications. An anecdote prefigured this evolution: for about fifteen years, I have been able to observe people dictating their SMS to their phone with a particular, very clear and explicit diction, to ensure that voice recognition doesn’t make mistakes. This strange way of speaking, this artificial tonality, already revealed speech addressed to a machine for the purpose of writing.

Today, the phenomenon deepens. In WhatsApp, Meta has integrated an artificial intelligence capable of finding information in our personal messages. If I once asked the AI to find an address that a friend had mentioned to me and it had trouble locating it because it was formulated in too “human” a way, it’s likely that next time, unconsciously, I will formulate this address more clearly. I know that my correspondent could ask the machine to find it rather than manually scrolling through messages.

This invisible but constant presence of artificial intelligence in our exchanges creates what Sherry Turkle called in Reclaiming Conversation: The Power of Talk in a Digital Age (2015) a permanent “three-way conversation,” where the third party is no longer human but machinic. We progressively develop what I would call “algorithmic consciousness”: a constant anticipation of how our words and writings will be processed by machines. This consciousness subtly but profoundly modifies our expression, creating a new form of communication that is neither purely human nor purely machinic, but truly hybrid.

Towards a new ecology of communication

I don’t want to pass definitive judgment on this evolution, I of course don’t know all its implications since we are only at the beginning. But it seems essential to me to become aware that we now have a third presence to which we also address ourselves when we communicate between human beings. This transformation is neither intrinsically good nor bad; it simply constitutes our new communicational reality.

The challenge is not to resist or blindly embrace this evolution, but to understand it in order to better take it into account and not lose ourselves in it. As Yuk Hui suggests in On the Existence of Digital Objects (2016), we are entering a new form of “cosmotechnics” where technology is no longer a simple tool but participates in the very construction of our being-in-the-world. Our speech and writing become conscious acts of this new ecology where humans and machines co-construct meaning.

I share this awareness to invite us to rethink not only our communicational practices, but also the very foundations of what constitutes human expression, in our new reality, now truly populated with artificial intelligences. The question is no longer whether we should speak to machines, because we do speak to them in fact, even if we don’t want to (for example, meeting participants), but how to preserve the authenticity and richness of human communication, while integrating these new interlocutors into our linguistic space.