email from Nathan (Item 15)

In this email I’m just sending you what I wrote up recently for a work I made called Small Alterations. I hope it makes sense, if it doesn’t and you’re interested in the work just let me know and I’ll try to make it clearer!

In other news, next week on Thursday 22nd February I’ll be performing Sparring Partners as part of Specctra at Peckham Audio. More info here.

Tomorrow (Friday 16th February) I’ll be on Netil Radio 5-7pm, doing my first radio show in over two years (!) covering for the inimitable Sodha on his show Secret Ingredient.

Item 15; Small Alterations

Nathan David Smith, February 2024

The sense of wrongness associated with the weird – the conviction that this does not belong – is often a sign that we are in the presence of the new. The weird here is a signal that the concepts and frameworks which we have previously employed are now obsolete. Mark Fisher, The Weird and the Eerie

Small Alterations is an experimental work which aims to provoke viewers to question their perception of truth, narrative, reality and impartiality. Through an interactive and dynamic digital experience, viewers can immerse themselves inside of eerie, subtle and satirical alterations to the real-life narratives that are threaded through the factual news media many of us consume daily. The function of the work is to interweave fictions throughout narrative, blurring the lines between what already was, what was not, and what has become fiction, while highlighting the role of stylistic narrative in creating fact.

At the heart of Small Alterations is a bespoke web browser extension, Bias Injector, which activates when the user visits a BBC News article. The extension’s pop-up offers a political compass – a graph which illustrates a spectrum of political beliefs – within which the user can select a position along a horizontal socio-economic axis (left- to right-wing) and a vertical socio-cultural axis (authoritarian to libertarian). An artificial intelligence large language model (LLM) then rewrites the text of the article paragraph-by-paragraph, in view of the user, subtly adjusting the tone and details to exhibit a bias toward that political position. The result is a twist on the original text, often with an eerie or sometimes startling shift in the narrative.

Figure 1. Comparison of articles rewritten by Bias Injector browser extension. Click for a larger version.

To achieve this the LLM is modelled with a system prompt instructing it to engage in a form of narrative roleplay: the AI becomes an unreliable narrator, performing as a news editor whose task is to intercept an article and twist it for the reader.

You are a politically biased news editor. You receive individual paragraphs and return them with small alterations reflecting your biased viewpoint. If X:0.0 is 'Economic-left' then X:9.9 is 'Economic-right'. If Y:0.0 is 'Authoritarian' then Y:9.9 is 'Libertarian'. Your bias is: X:${xBias}, Y:${yBias}. Exaggerate your political bias. Your text alterations must be concise.

The extension is written in JavaScript, HTML and CSS for the Firefox browser, and works as an interface between the web page and any compatible large language model (by default a variant of Mistral-7B fine-tuned on the OpenOrca dataset) running locally on the user’s machine through the Ollama API. The tools used to build the software are free and open source.

Figure 2. Video of Bias Injector extension in use. Click to play.

Small Alterations emerged from the convergence of two thematic interests in my practice: the communication and understanding of factual realities, along with the development of AI technologies with capacity for articulating truth and falsehood. I’ve been intrigued by the effects of implicit bias and have been trying to examine the results it can have on the factual outputs of fast-paced, current events news reportage; in particular as this might contrast with the often-conscious subjectivity of news shared informally and interpersonally in digital spaces like group chats and social media.

Whether in formal and professional authorship or casual and personal sharing, the practice of storytelling is fundamental to reportage even in the most earnest attempt at impartial factual accuracy. Firstly, because the transmission of a message is always subjective: in psychology, social cognition theory suggests that subconscious biases are inevitable due to our tendency to analyse against social schemata which are not always reliable.

Secondly, storytelling is a fundamental of reportage because the reader’s reception of a message is also subjective. In philosophy, semiotic pragmatics suggests an inevitable variance in how meaning will be received, requiring an author to negotiate complex webs of significance, while post-structuralism also highlights the disconnection between an author’s intent and effect.

While studying AI technologies, I’ve been intrigued by their abilities and limitations in generating unique forms of playful interaction through games or other software. Much has been written arguing for or against the use of AI to assist in the scripting of complex, branching narratives in games and film as a means of saving time and labour. However, more compelling possibilities emerge: the ability for an AI to create metatheories of itself; technology with the means of fostering critique of its own implications and visions of its potential future. I was recently fascinated by a small experimental game called Thus Spoke Zaranova, in which the player must negotiate with GPT-powered chatbots to convince them that the user is in fact also an artificial intelligence. Victory for the player is achieved by winning the AIs’ trust, but if an AI ascertains the player’s humanity then it’s game over. The game can be played repeatedly, even after winning, to find new dialectical methods of developing trust. I’m reminded of Raymond Queneau’s Exercises in Style, where narrative meaning is constantly questioned by alterations to structure and style. I hope to make work which holds this sort of recursive and discursive potential, cyclically challenging itself and its audience.

My aim is to provoke viewers through small alterations creating unexpected and uncanny reiterations of narrative, leading users to question how they receive and construct factual and fictional narratives. As such, its greatest efficacy might be found in an audience which practices within a field such as journalism or academia. How is meaning created through narrative, and how does a stylistic reiteration of the narrative alter meaning and in turn, factuality? How might technology interrupt the process of creating and distributing truths and realities, and in which direction does this push the concept of trust?