Chris Chesher

Forthcoming as a ‘probe’ in Explorations in Media Ecology to be published as ‘Re-assembling agency: Invocational AI actants in media ecologies and Actor-Network Theory’.

Introduction

In working on this probe, I have been both thrilled and deeply unsettled by using ChatGPT’s Deep Research feature, an AI that does more than just chat. It reasons, researches, writes and references entire long form texts like this piece of writing. This probe documents the experience of writing in the age of AI, and my assessment of its ethical and professional implications. But I promise I did not use AI to write it for me. 

So the key question is: What is the experience and ethics of research and writing when humans and non-humans perform it together? 

Before AI, my research process was already highly mediated: mind-mapping in MindNode Pro, searching Google Scholar for articles, taking notes on them in Marginnote, managing references in Mendeley and writing in Microsoft Word. Each of these let me invoke actions like visualising, searching, annotating, writing and so on, but most responded to my direct invocations. 

Deep Research was a different experience. Based on some initial research, I wrote a detailed prompt, and it took 30 minutes to produce a well-organised, researched, insightful and polished 9000 words. What the hell can I do with this, I thought? Sure, it validated my research plans, accelerating and automating their development, but the results were intimidating and made me feel strangely empty. My work was already done. I’d discovered someone else’s article on exactly my topic. How can I unsee this text? Is Deep Research my assistant, conversation partner or my opponent? Is it an extension of my nervous system or a sophisticated actor with its own capacities for action? 

I decided to put aside Deep Research’s piece and write my own draft focusing on Deep Research itself. In the process of writing, I surfaced the concept of translation: a term used in different ways in AI discourse, media ecology and actor-network theory. Where the computer science field of natural language translation aims to perfect translating content in a source language to a target language, media ecology uses the term to refer to how media and technologies alter human perception, experience and culture. Actor-network theory sees translation as a key process through which human and non-human actors act as mediators that actively transform, translate, distort, and modify what they carry. Others have observed the complementarity between ANT and media ecology (Van den Eede 2013)

After a couple of days, I had completed a first draft of the paper. Out of curiosity, I wrote a new version of the prompt to Deep Research that followed the structure and content of my new draft: 

Please write a ‘probe’ for a media ecology journal that theorises the emergence of AI agents (a form of what I call ‘invocational actors’), using ChatGPT Deep Research as the central case study. The conceptual focus is on different meanings of ‘translation’ in the process of invoking non-human agency and hybrid authorship. Begin with an overview of the paper. In the first section of the paper, explain the genesis of LLMs from natural language translation. Introduce the concept of invocational media (Chesher) and the new concept of invocational actors. Now introduce McLuhan’s conception of media change as a process of translation, using Deep Research as the example (he would see it as an extended nervous system. Find quotes from Understandig Media). Then introduce the actor-network theory concept of translation, first from Latour and then Callon. Apply this version of translation to the case study of the Deep Research feature, this time referring to the enrolment of non-humans into networks that reconstitute authorship. Write a conclusion on the crisis of authorship, with some proposed heuristics for ethical knowledge production that falls short of complete delegation of knowledge generation to AI. 

The prompt gives you a good idea of the rest of the paper. If users usually perform invocations to their machines (clicking, writing, searching), now they are invoking agents that appear to have a life of their own. I argue that advanced LLMs are invocational actors called upon to perform intellectual labour independently of users’ conscious intentions or control. What begins with translation emerges as agency. With this in mind, I revised my draft. 

Invocational media

At this point, I should explain my theory of invocational media (Chesher 2023) which proposes that the elemental event in computers is not computation, but invocation. In computer discourse, a program is said to invoke another object or function. Users of voice assistants are said to invoke the assistant to give a response. Computers work with calls more than calculations. They invoke central processors, memory, storage, outputs and inputs, all orchestrated to the pulsing rhythm of the fetch-execute cycle. They invoke software entities too: objects, methods, functions, subroutines and so on. These are invocable domains: software and hardware components that CPUs call. The traditional meaning of the invocation – a faithful call on a higher external power for immediate assistance – also resonates with the popular association between magic and technology. Invocational media are both rational and magical, reasoning and illusion-generating, calculating and theatrical (Timplalexi & Rizopoulos, 2024). 

In developing the theory, I proposed a new diagram for invocational media focusing not on the hardware or software components, but on relational events of invocational media. Invocations are the fundamental translational events in this medium, but there are other operations. Avocations call users to the machine, instructing and encouraging them to choose invocational media to solve their problems: advertisements, instruction manuals, peers and the interface itself guide users to make invocations. While conventionally, the term ‘avocation’ refers to hobbies, I extend the use of the term to refer to all the forces that call people away from other modes of action to make them users.  Evocations are the sensations and meanings people experience in response to their invocations. There are more than just outputs, as they affect the sensory experience, feelings and cognitions of the user. Convocations are invoked social spaces that call people together through email, social media, game spaces and so on. 

However, in the process of writing this probe I now realised a problem. My book did not account for the fact invocations sometimes invoked actors with their own agency — performing ongoing and autonomous invocationary action in invoked and physical environments — out of sight of the user. I needed a new concept not only to understand Deep Research, but also non-playing characters in games, robots, and other autonomous systems. I chose the term ‘invocational actors’[CC1] .

From translation to agency

When ChatGPT launched in November 2022, it provided a simple interface: the avocations of a text field with the grey words ‘Ask anything’ and a space for a prompt. The user could prepare an invocation by choosing a model and entering their prompt in plain text. They could initiate the invocation by clicking the up-arrow button. 

ChatGPT is an evocative actor that performs as an erudite and articulate (non) human, translating prompts, predicting and generating the best responses. This translation from input to output, from avocation to evocation, is the basis of all large language models, giving users an experience resembling textual conversation. 

But with reasoning models and deep research, what seemed conventionally as the translation of user invocations is increasingly experienced as the summoning the agency of invocational actors to provide extended individualisedresponses[CC2] . How to theorise these invocable entities that seem to begin with translation and emerge acting autonomously? 

Translation in media ecology

From the perspective of McLuhan, new media perform translations of ‘one kind of knowledge into another mode’ (2013). He says that with language, humans are ‘able to let go of their environment in order to grasp it in a new way.’ Language possesses almost magical capacities, he says: ‘By means of translation of immediate sense experience into vocal symbols the entire world can be evoked and retrieved at any instant.’ 

For McLuhan, these kinds of translation changed progressively with writing, print and electric media. Each new media form retrieved, translated and evoked the world in its own ways, each with its own biases and ratios of the senses. Each introduces new relationships to the world, extending human perception, changing the scale, pace and pattern of everyday life: print, radio and television.

But what about computers? McLuhan’s chapter ‘Automation’ frames information systems as extensions of man’s nervous system, resulting in an ‘age of instant and total involvement’. At another moment, though, McLuhan warns of the danger that electric media will end up producing ‘utter human docility and quiescence of meditation’ (McLuhan 2013, p. 59). McLuhan is ambivalent — information systems promise to extend the nervous system promoting total involvement at the same time as they risk cauterising human faculties and alienating people from themselves.

Here McLuhan anticipates responses to media like Deep Research as they escape human grasp. Some AI advocates celebrate total involvement with this technology, while others fear will be detached from intellectual agency, transforming them from authors into managers or advisors. What was translation (and invocation) that augmented the human might end with the relegation of humans to insignificance. 

Translation in actor-network theory

As mentioned earlier, the meaning of translation is different in Actor-network theory (ANT), referring to processes by which human and non-human actors are themselves constituted through the formation of networks of associations. Human actors may grasp, but non-human actors grasp back, forming networks stretching in multiple directions. Rather than focusing on the man and media binary, ANT analyses all kinds of mediators, each participating in different forms of translation. Translations establish new programs of action, enrolling groups of actors, and delegating actions to them in more heterogeneous ways that in most media ecology[CC3] . In the network-generating process, actors are always themselves transformed. What emerges are hybrid ‘chains’, networks’ or ‘assemblages’ like the Pasteur network (1988), the scallop network analysed by Callon (1986), or the printing press network (Latour 1986). 

Networks have some commonalities with Postman’s concept of environments, but networks are comprised of heterogeneous actors, framed in a less epochal way. Just as media ecology draws attention to the powerful role of the background, so ANT challenges the foreground / background binary by discovering that the background can in fact be understood as a diversity of networked actors previously denied agency (Stalder 1998). There are other differences. For better or worse, ANT lacks media ecology’s moral urgency and privileging of the human condition. Instead it tends to give attention to how networks produce forms of domination (social, ecological, etc.), moving from matters of fact to matters of concern (Latour, 1992a). 

As I outlined earlier, conventional invocational media networks enrol users with avocations that guide them to perform invocations to invocable domains. Invocations appear to be extensions of users’ intentions and desires, and they are offered meaningful evocations of sensations, information and action. Deep Research and other invocational actors are somewhat different. 

Once invoked, Deep Research autonomously searches for sources, reads them, synthesises information and generates texts that could pass as products of human labour in research and writing. This reconfiguration destabilises the long-standing and well-defended human position of authorship. There are ongoing tense negotiations about the roles and positions of writers, technologies, texts and readers. Each new arrangement changes the operation and identity of scholarly networks. And the new network looks different from the perspective of each of the actors. Scholars themselves are translated and transformed: some feeling enhanced or remade, others disturbed, displaced or even destroyed. 

The emergent outcomes are likely to be more nuanced. Just as in mythology, LLM invocations are notoriously unreliable. LLMs often make factual and logical errors, particularly at the margins of dominant knowledge. This is particularly intractable because the training and generation processes are inscrutable. In ANT terms, LLMs are black boxes with only inputs and outputs visible. In my nightmares, they are also a Pandora’s box with dark implications for conventional scholarly practices, an unleashed invocational demon, that calls into question thought itself. 

Deep Research has proven itself a sophisticated general-purpose creator of evocative scholarly text. Passing the Turing test at speed, its text is often indistinguishable from the human. It unsettles values such plagiarism, authorship, critical thinking, genius, hard work, style, rhetoric and so on. Writers are transformed from lone creators of knowledge to invokers of non-humans to produce work for them. Yet, LLMs are not the first non-humans called to participate in scholarly performance[CC4] , going back to writing itself, libraries, print, telegraphy, and radio. The invocational media of databases, word processors and spell checkers are relatively uncontroversial because their invocations are more transparent. Most importantly, scholarship has never been the production of original texts by isolated minds, but an intersubjective, collegial, institutionally situated and globally networked proliferation of events of thinking, teaching and communicating far more molecular than the individual texts that pass peer review and appear as citations. As I found out in the process of writing this probe, prompting is itself a scholarly practice and competency[CC5] . To perform effective invocations, the author needs to be aware tacitly of the invocable domains of scholarship that are partially modelled in the AI systems (even if this is itself highly problematic). While there will undoubtedly be accusations of academic misconduct, and complications with the acceleration of scholarly production, there are options about how these communities might regulate themselves. 

So, what are we to do individually and collectively? Individually, we can refuse to use LLMs altogether or self-regulate in how we employ these invocational actors. Collectively, organisations like universities, disciplines, publishers could develop protocols or prohibitions for certain uses of AI. The same institutions should also advocate for recompense for the breaches of intellectual property upon which these systems are trained. The concerns of media ecology are likely to relate to the character of translations and the whole environment as much as to the new extension itself. For ANT, it is about ongoing negotiations over the forms of association and regulation of networks and the norms about the identity and capacities of the emergent hybrid scholarship of humans and non-humans. 

References

Callon, M. (1986). Some elements of a sociology of translation: Domestication of the ccallops and the fishermen of St. Brieuc Bay. In J. Law (Ed.), Power, Action and Belief. A new Sociology of Knowledge (pp. 57–78). Sociological Review Monograph.

Chesher, C. (2023) Invocational Media: Reconceptualising the Computer. New York: Bloomsbury. 

Latour, B. (1986) ‘Visualisation and cognition: Drawing things together’. Knowledge and Society 6: 1–40.

Latour, B. (1992a). Reassembling the social: An introduction to Actor-Network-Theory.

Latour, B. (1992b). Where are the missing masses? The sociology of a few mundane artifacts. In W. E. Bijker & J. Law (Eds.), Shaping Technology/Building Society: Studies in Sociotechnical Change (pp. 225–258). MIT Press.

Stalder, F. 1998 From Figure / Ground to Actor-Networks: McLuhan and Latour. Conference presentation: Many Dimensions: The Extensions of Marshall McLuhan Conference, Toronto, 23-25 October, 1998 https://felix.openflows.com/html/mcluhan_latour.html

Timplalexi, E., & Rizopoulos, C. (2024). Intermedial and theatrical perspectives of AI: Re-framing the Turing test. Explorations in Media Ecology, 23(2), 153–174. https://doi.org/10.1386/eme_00203_1

Van Den Eede, Y. 2013 Opening the Media-Ecological Black Box of Latour. Explorations in Media Ecology 12:3/4 pp. 259-266. 

---

 [CC1]I prefer this term to Latour’s ‘actant’ because 1) it is the more common usage in ANT and 2) I want to retain the association with performance (the becoming actor of the assemblage, if I want to get Deleuzian). I think it’s unnecessary to explain here. 

 [CC2]In response to your suggestion I have refined this claim

 [CC3]While I am not going to make this argument in depth, I think ANT’s ‘translation’ is different from McLuhan’s. 

 [CC4]I’ve expanded this section to build on the section you suggested needed more work.

 [CC5]Thanks for the suggestion: I made this more explicit