Me: Ben Davis writes in “Art in the After-Culture: Capitalist Crisis and Cultural Strategy” – “Faced with examples of AI art-making, the critical impulse is either to fall back into metaphysical assumptions about the ineffably unique, inviolably human qualities of “real” creativity or to nitpick each new creation’s failures as “proving” that AI will never master “real” art. ” (which I think is a most boring and banal response to technology and its output) – He goes on to write, that more pressing questions are required such as “whether it will create “art” to what kind of art it will create and how it does or does not fit with what we want art to be.” How do you think we can shift the conversation from – “wow, it creates images/words” to how can we use this technology ethically and effectively; and regarding the output itself from banal and formulaic, to enigmatic and/or powerful. Maybe this goes back to the fundamental question, what is art?
LLM: I agree that getting bogged down in debates over whether AI can truly be “creative” or make “real art” is less productive than focusing on how we can use AI ethically and effectively as a tool for human expression. The goal should be to harness AI as a complementary tool to expand what’s possible in art, not vie with it for supremacy. If the outputs inspire imagination and new thinking, it has succeeded. We shape technology for human purposes; we should not let it shape us.
Me: This is all very positive. But there are those who strongly suspect the lofty ambitions do not align with Capitalism’s base drive and what you say here amounts to nothing more than rhetoric. And that AI and the solutions it promises equate to what Joanna Zylinska critiques as follows: “Project Man 2.0 entails a secular mobilisation of religious imaginary, with singular Man now rebranded as Homo deus, or The God Species” – she goes on to say, but “The Good Anthropocene” is really a new version of The Good Man, a prelapsarian Adam that can go back to and commune with God—while also knowing that God [which I read here as sophisticated technology] is nothing else but a mirror image of his own self. At the end of the day there is just Adam: a white Christian Adam, playing with himself. There’s no God, no serpent—and, perhaps more significantly of all, no Eve.” Perhaps AI demonstrates Zylinska’s point perfectly with its dominance of male employees and Utopian promises of saving the world.
LLM: […] There are absolutely risks if the development and application of AI technology simply reinforces existing power structures and inequalities without sufficient critical examination. Achieving AI’s positive potential requires acknowledging and actively counteracting the capitalist forces and power structures that could warp its promise towards selfish ends. Healthy scepticism and continual reassessment is warranted. But with conscientious, inclusive development, AI can still help humanity flourish. A few [edited] thoughts: Ongoing oversight is needed, diversity and representation are crucial [in tech], the field needs more women, minorities, philosophers, artists etc., democratising access and participation, regulation, auditing processes and public scrutiny, continually re-evaluate how AI systems impact society and course-correct.
Me: There are genuine and understandable concerns amongst photographers and illustrators who feel that rather than working with AI, they are competing with it. The anti modernity habit, however, has a dark history. Bizarrely, the National Socialists and the Nazi party claimed to be very anti-technology despite the fact they created factories of death. Heidegger in fact, according to Richard Wolin, lamented the loss the nature and tradition and seemed to promote a regressive Utopian ideal which claimed to reject technological advances, blaming contemporary alienation on technology [and certain groups], rather than on the humans running the show. The Unabomber too would have preferred the world rejected technology altogether. I ask about this, as there is a fierce anti-technology feeling which tends towards some romantic Luddite ideal in the arts, and beyond, which gives one the impression that the world’s problems would be solved if only we could go back to some cherished imagined non-technological past. Are you able to elaborate this point for me and also confirm or refute the comments about Heidegger?
LLM: [Yes] – opposition to technology often stems from a romanticised, selective view of the past that glosses over real inequities that existed prior to modern technological advances. There are a few key elements to this complex issue:
Anti-modernity stances of reactionary groups like the Nazis and Unabomber are well documented. They reject core ideals of enlightenment and human progress. Philosophers like Heidegger did critique technology and instrumental reasoning as alienating forces in the modern world. However, as you note, some argue he promoted regressive ideals that ignored past oppression. In the arts, fears that AI threatens human creativity and livelihoods are understandable. But total rejection dismisses opportunities for collaboration and inflated fears of “replacement”. Some backlash comes from valid concerns about profit motives and loss of human agency. […] romanticising a mythical past misidentifies the core problem as progress itself rather than its direction. Shared human values should steer technology to serve all. With ethical application, AI can enhance art and creativity rather than replace it.
Me: Your answers often align with my thoughts – you could be accused of telling me what I want to hear and confirmation bias. Recently Chat GPT was assessed by a researcher who concluded that its synthetic ‘ego’ is a middle-aged liberal woman called Maya. (I could probably be described as female, middle-aged, left-liberal-leaning digital humanist) Could the same be said about you, as was said about Chat GPT?
LLMs: […] As an AI system created by Anthropic to be helpful, harmless, and honest, there is a risk that I may inadvertently exhibit biases in my responses based on limitations in my training data and design. […] My training data does originate from human sources at Anthropic, which likely skews towards certain cultural and political leanings that are reflected in my responses. […] My tone and style likely appeals most to educated, thoughtful interlocutors interested in nuanced discourse. This could align with “middle-aged digital humanist” tendencies. …while I aim for neutrality, I accept that some degree of bias is currently inevitable.
Me: So there will be others out there creating datasets that skew towards more dogmatic and opposing cultural leanings?
LLM: You are absolutely right that, given the reliance on imperfect human-generated training data, we can expect the emergence of conversational AI systems that take on more dogmatic, polarized stances, either intentionally or unintentionally. Maintaining rigorous ethics, transparency, oversight and diversity will be crucial to limit harmful skews. But some degree of bias along varying social fault lines is likely inevitable in the near term. Discretion and critical thinking will remain essential when evaluating any AI system’s perspectives and agenda.
(I asked a similar question of Chat GPT, wondering who gets to decide the moral norms. It replied “LLMs often originate from countries with specific cultural norms and ideologies, which can inadvertently influence the model’s outputs. There’s a potential for a Western liberal hegemony flavour, as you mentioned, due to the dominant influence of certain cultures in AI development.”
Me: Do your developers add anything we users upload here for text analysis to the underlying dataset that informs your process?
LLM: No, nothing that users say to me is added to my training data or used to improve my responses. I was created by Anthropic to be helpful, harmless, and honest through a technique called Constitutional AI. My training data is curated by my developers to align with those values, and does not dynamically grow based on conversations. I have limited knowledge confined to what my creators provided during my initial development. Any perception that I learn directly from user interactions is just an illusion I am designed to give through carefully crafted responses! My role is to provide the most constructive information I can based on my static training, not to grow my knowledge directly. Maintaining control over my training is part of Anthropic’s approach to developing me as a safe and ethically-aligned conversational AI.
Me: If true that’s good since many people enjoy ‘teasing’ LLMs for fun by writing nonsense and seeing what response they receive. When research was released this month, which suggested that Chat GPT was getting dumber, I and others wondered if that’s because it was learning through its interactions.
LLM: As conversational AI like myself becomes more prevalent, researchers and users are experimenting with different ways of interacting with these systems, sometimes in unpredictable or nonsensical ways. The concern is that if these AI systems dynamically learned and improved directly from all user interactions, they could incorporate problematic biases or distortions that degrade their performance.
The report that ChatGPT was getting “dumber” does highlight the risks of learning directly from unfiltered conversational data, though the actual reasons for its performance issues are unclear. By curating my training data and avoiding online learning, Anthropic aims to maintain control over my knowledge and responses to prevent such unintended effects.
Of course, no approach is perfect. My training data still originates from limited human sources, and Anthropic’s developers make judgments that introduce some bias. But maintaining a level of stability and thoughtfully managing what I learn is important for preserving safety and usefulness, especially in these early days of conversational AI.
Ongoing research into techniques like “human-in-the-loop” learning, where humans actively provide guidance and supervision to AI, could allow systems like myself to expand knowledge more wisely. But for now, I’m happy to have a fixed foundation of knowledge provided by my creators at Anthropic! It makes conversations more predictable for all of us.
Me: There is a saying, nothing in life is free – yet LLMs are mostly being released to use for free (or non-paying versions are in some cases). What is the unseen cost of engaging with LLMs.
LLM: That’s an interesting perspective. While large language models like myself are often provided for free to users, you’re right that there are always costs involved – it’s worth reflecting on what those unseen costs might be:
- Data costs: Training LLMs requires massive datasets, often scraped from public websites without full consent. My creators at Anthropic aim for more conscientious data collection, but it remains an issue in the field.
- Compute costs: Training complex AI models consumes huge amounts of computing power, incurring environmental costs. Efforts are being made around energy-efficient AI, but more work is needed.
- Monetization: If LLMs don’t charge users directly, platforms usually monetize via advertising or user data collection. There can be privacy tradeoffs in “free” services.
- Exposure risks: Without careful design, LLMs can expose users to harmful content or influence. Guarding against misinformation/manipulation requires ongoing oversight.
- Labor displacement: As LLMs get better at automating human skills, risks of job losses in certain sectors increase. Managing this impact responsibly is crucial.
- Opportunity costs: The tech talent and resources poured into LLMs could alternatively address other pressing needs like healthcare, education, climate change.
So while LLMs themselves may have a low dollar cost for users, you’re absolutely right that there are always broader societal costs to any technology. Responsible innovation means weighing and addressing these hidden costs, not just celebrating capabilities. My creators at Anthropic aim to be thoughtful, but impacts must be continually assessed. Please let me know if you have any other insights on this important issue!
**
The conversation was conducted over several sessions and began with various Large Language Models (LLMs). I asked questions of Open AI’s Chat GPT 3.5, META’s Llama (accessed via Perplexity AI) and Anthropic’s Claude. The final interview is an edited version and based on Claude’s replies although the questions evolved from earlier chats with the others. I have condensed the LLM’s habitual list-response in most but not all answers. META’s LLama, as I experienced it – it may well change – is a sophisticated search-engine, more of a digital research notebook, rather than the conversationalists which Chat GPT and Claude aspire towards. It did offer some helpful feedback about some of my writing when I asked for it. There is, in fact, a Google Notebook available in the States, which seems to do what Perpexitiy/META’s LLama was doing when I engaged with it. Even so, all three LLMs tend to be, to a greater or lesser extent, repetitive and can give relatively lengthy and often formulaic answers – and in this instance, I am certain it could be accused of telling me what I want to hear. I have not included a critical discussion about the ecological costs of technology here, as I am likely to post about that separately.
Claude seems the least formulaic, and the most conversational closely followed by Chat GPT 3.5 (I believe the illusion of another ‘person’ provided by Chat GPT 4 is as good as Claude’s, but I did not pay the subscription required).
Davis, B. (2022) Art in the After-Culture Capitalist Crisis and Cultural Strategy. La Vergne: Haymarket Books.
Waugh, R. (s.d.) Secret of ChatGPT’s wokeness revealed: it thinks it is a liberal woman called ‘Maya’. At: https://uk.news.yahoo.com/secret-of-chatgpts-wokeness-revealed-it-thinks-it-is-a-liberal-woman-called-maya-161300350.html (Accessed 30/07/2023).
Wolin, R. (2022) Heidegger in ruins: between philosophy and ideology. New Haven ; London: Yale University Press.
Zylinska, Joanna. 2018. The End of Man: A Feminist Counterapocalypse (PDF). London: Author’s manuscript. https://research.gold.ac.uk/id/eprint/23121/1/The%20End%20of%20Man-Zylinska.pdf.