Note: What follows is largely speculative thinking - an invitation to dream and imagine different approaches to AI safety. While some ideas may seem far-fetched, they're meant to provoke discussion about what's possible. I'm particularly interested in learning about concrete projects already working in these directions.
As someone who has been on both sides of very controlling dynamics in my life - around people who, whether covertly or overtly, tried to get me to behave in what they thought was the right way, or understanding the ways I hid/presented only specific parts of me to elicit a specific reaction from the other person, effectively depriving them of their freedom of choice - I intimately know that trying to control another intelligent, agenty being is a recipe for disaster.
This dynamic manifests at both personal and societal scales: any undesired or "bad" part of someone's psyche that gets suppressed inevitably lashes out and causes damage. Similarly, movements in society that have been demonized for decades – like the far right – tend to resurge aggressively.
Having developed this sensitivity for systems that control, I find that 90% of the conversations I hear around AI Safety and alignment trigger that alarm in me: "Oh, we are trying to control the behavior of something likely to be more powerful than us soon... This is not gonna go well."
A key concept rising into collective consciousness is that of needs-and-boundaries-based relating. The idea that we can't control other people's behavior but we can control our own, and it is our responsibility to learn how to meet our needs without depending on any specific outside entity.
I don't think it's a coincidence that this awareness is bubbling up right now. The world we live in is suddenly allowing a lot more agency and freedom of choice to people, thanks to some social revolutions and various technologies. Some obvious examples: women having more freedom of choice in relationships because it's structurally easier to be financially independent, people having some freedom of choice in banking systems because of cryptocurrencies, customers having more power in relating to companies because of social media platforms that can amplify their voices.
So I find myself asking: what insights from this parallel apply to the relationship between humans and AIs?
How would we design a world, our institutions and incentives, if we thought that we don't get to control AI at all, that we just get to:
Control our behavior, what we do with our resources, how we fulfill our needs, what and when we share data
Let AI know about its impact on us, in the form of building better ways to communicate with it how it's affecting us
How would AI Safety research look like then?
I'm sure a lot of research of this shape is already happening, and I haven't found it yet because I'm using different words to describe the same concepts - I don't deeply know a lot of computer science. And I think there's value in thinking about it in these terms. I'm driven by thinking about where conflict arises in relationships, usually when "I" feel hurt in interaction with "you." This requires knowing what is me and what is you and where the edge is exactly. It requires some understanding of what resources I have, what resources I need, how I get more of something, and who gets to take it from me - at all possible levels, from physical money or things, to emotional attention or love, to the energetic like life force/energy.
Multiple questions come to mind here:
What needs of humanity are driving AI development? How do we create incentives for ourselves to be sure we can meet those needs without AI, at the society level?
Some actions that go in this category: need for more powerful optimization of many of our systems, need for ease in meeting our other needs (think routing, recommenders, resource allocation), need for health and peace and support with solving our health problems and resource creation, need for money
What does an anti-fragile system look like here? How do we make humanity resilient to inevitable AI alignment failures?
What will AI never be good at? Not what is it not good at yet, but what is fundamentally impossible for AI?
Things like: predicting quantum phenomena, predicting specific instances of distributions instead of the whole function, modelling noise
What actually makes humans human?
Less about what we write and think - clearly that's not it. But maybe there are other things that make us fundamentally human? What's a CAPTCHA that a baby could solve but not an AI? Something about breathing patterns or movement patterns of many individuals
Also, honing the role of intuition. Humans know when they're being manipulated, even though not everyone is always willing to face it. We feel deeply when something is pulling us further away from our center. What does it look like to train this ability at scale? To create incentives for us to use this ability more?
How do we make sure that AIs can see/feel the impact they have on us? How do we give them the ability to reason and sense and react to their own boundaries being crossed?
Can we give them some form of empathy through access to biometric data? Or show them a chain of thought that looks more like Internal Family Systems - allowing all different parts/drives to speak up and then collectively decide after hearing everyone?
Next steps I’m obsessing about right now:
Communities already building human-centric technological infrastructure (e.g. Community Privacy Plan)
Introspecting on what boundaries with AI I need to hold and does that look like concretely (getting serious about where and how I store my data, setting up local AIs, actually enforcing non-AI time, being conscious of what skills I don’t want to lose..)
Learning more about what is inherently not predictable
Learning more about what people are implementing that I would call a boundary (which sounds like I’m about to deep dive into computer science security stuff, finally).
May all beings interact in respect and reverence for one another. May all beings find interdependent ways to meet their needs. May all beings be happy.