Join us at the Dyson School of Design Engineering for a series of seminars on Responsible AI, featuring expert speakers who will share their insights on how to develop and deploy AI systems responsibly. The series is chaired by Professor Rafael Calvo.
Talk Abstract
One of the most intriguing theories in philosophy is the extended mind hypothesis: the view that our minds are so intertwined with parts of our environment that our mind itself extends beyond our skull and into the world (Clark & Chalmers, 1998). We offload maths to a calculator, memory to our notebooks – and now we offload reasoning and creativity to AI.
In this talk I will explore our relationship with AI through the extended mind lens. Is it possible that our thinking, learning, and creating become so integrated with LLMs that these systems become part of our minds? And if so, what are the practical consequences? Should we allow for our minds to merge with AIs, and in what conditions?
I will argue that
- LLM tools can meet the tests for counting as part of your cognitive system (they are instantly available and you tend to trust their outputs).
- Once these tools leave the chat box and become AI agents that decide and act for us, they extend not just our minds but our whole agency. We think, choose, and act through them.
- This means whoever owns these tools has enormous power over our minds and our agency. And software platforms have a habit of enshittifying: great while they are winning you over, worse once you are locked in (Doctorow, 2025). So watch out: if your mind is tightly linked with LLMs, and they become shit, then part of your mind gets trapped in that mess.
- To protect our mind and autonomy from enshittification, we should keep as much control as we can over our own thinking infrastructure. We should own the means of agency production.
This is a call to build. As designers and engineers, you may be the people who decide whether ordinary users can own their cognitive tools, or whether we all end up renting our minds from a handful of companies. How can we build the equivalent of a FairPhone or a Wikipedia for AI?
Small-scale local hardware will not match what giant data centres can do, so there are limits to what we can achieve locally and bottom-up. But every step is a victory. I conclude by exploring the potential, limitations, and risks of integrating our minds with local AI agents.