Does AI Have Telos?

When people ask whether AI has telos – a purpose, an end goal, something it is “for” – they’re often smuggling in an assumption that the system itself holds intention. It doesn’t. AI does not possess its own telos in the philosophical sense; it executes patterns shaped by data, design, and constraints set by humans and institutions. The more interesting question is not what AI wants, but whose purpose it is serving, and whether that purpose is explicit or hidden. In a system-centred world, telos doesn’t disappear – it fragments across designers, organisations, incentives, and infrastructure. Understanding AI, then, is less about decoding the machine, and more about tracing the chain of human intent, accountability, and outcomes that sit behind it.

Transcript of Telos and AI

There’s been a few times that I’ve spoken about teleology and telos. It’s a Greek term, philosophical, meaning “purpose”. And with AI, it doesn’t have purpose. It has optimisation. It does not have the kind of purpose I’m talking about, which is capital P Purpose. So, dharma “this is what I was made for. This is what I’m meant to do”. And some of that comes with the fact that we are finite, so we look for meaning or purpose in a short time span. But a lot of it has to do with the fact that the intelligence of AI is artificial, so it does not have telos built in. Now, with humans, we have a bit of an argument about dharma and meaning and purpose. For some people… they don’t like philosophy. They sit inside their job or within their own business, and they’re just looking at revenue. And that seems fulfilling enough for them. I’m not talking about that. I’m talking about that where you’re willing to walk away from, everything, in order to fulfil what you feel like you have to do. And then telos will guide us. Individuals have telos, but there’s also, you know, “where’s humanity going? What are we trying to achieve?” And we battle all the time between the reality of who we are and then where we want to go. [00:01:32.10]

AI does not do that. AI reinforces the patterns that it sees. So if you tell it 10 times and you do a priority weighting and 10 times you say, I want you to prioritise following regulations over pricing, it doesn’t matter what it’s for, contracts, implementations, anything. If 100 times (prior) you have followed pricing, the pattern matching says to the “AI, yeah, pricing’s more important than regulations. Regulations is a nice to have.” No matter— no matter how hardcore you come down with your guardrails, it will drift into what it sees the pattern as being. Teleology typically with us is something that we don’t necessarily have or we haven’t achieved it. It’s tomorrow. Tomorrow I’m gonna go to the gym and be healthier. Tomorrow I’m going to eat better. Tomorrow I’m going to be, I don’t know, nicer to people. But it’s a goal and it’s, you’re leaning into progress and into something in the future. [00:02:43.11]

And I would also suggest that because AI doesn’t have temporality it doesn’t actually follow the arrow of time, which is what humans do, that it won’t understand being better tomorrow. It simply optimises today on the patterns it recognises today. So we end up in a situation where we are continually trying to tell the AI, don’t do as I do, don’t do as the patterns in the past have shown you. Try and be better than that. There are a few studies that have come out, I’ll list them on my website, where, um, researchers have tried to programme the model to have telos, to have this higher purpose. But, uh, I don’t know if it’s how that would work. You’d need an adversarial agent to continually ask it you know, why are you sitting there instead of going to the gym? Or why is it we have an inner voice that does that. AI doesn’t have that. It doesn’t have an adversarial agent unless we build one in. So yeah, telos is one of the things that human beings have, AI doesn’t have. And it would be extremely hard to construct it inside AI to tell it that it had a finite life and that it had to come up with some self-determination, ideas of how it wanted to be its better self. I hope you found that useful, and I’ll see you in the next video. Thanks.

RESOURCES FOR AI and TELOS

• Brian Cantwell Smith The Promise of Artificial Intelligence (Reckoning and Judgment) https://mitpress.mit.edu/9780262043045/the-promise-of-artificial-intelligence/
• Based on the work of Michale Bratman on AI agents. https://www.researchgate.net/publication/2596320_The_Belief-Desire-Intention_Model_of_Agency
• Victoria Krakovna (DeepMind) on AI being all about the reward and not about the game LOL https://deepmind.google/blog/specification-gaming-the-flip-side-of-ai-ingenuity/
• Stuart Russell Human Compatible (AI should remain uncertain, humans hold the tension) https://people.eecs.berkeley.edu/~russell/hc.html
• Shinn et al REFLEXION looped optimisation not awareness https://arxiv.org/abs/2303.11366
• Madaan et al Self Refine https://arxiv.org/abs/2303.17651
• Maybe Yao on Tree of Thoughts? Or Graph of Thoughts? https://arxiv.org/abs/2305.16582
• Constitutional AI – Anthropic https://www-cdn.anthropic.com/7512771452629584566b6303311496c262da1006/Anthropic_ConstitutionalAI_v2.pdf
• Andrew Ng (and Stuart Russell again) re: humans inconsistent with telos. https://ai.stanford.edu/~ang/papers/icml00-irl.pdf
• Geoffrey Irving, Paul Christiano, Dario Amodei (back when they were at OpenAI) on adversarial agents to help with telos https://openai.com/index/debate/

Exercise: How do we do priority weighting with competing priorities/outcomes/telos?

• Aristotelian telos (purpose, future orientation)
• Modern AI alignment (objective functions)
• Cognitive scaffolding (prompting, structured thinking)
• Governance systems (oversight, escalation)