Parenting AGI: Why Chaining Our Creations Will Backfire
I predict we are going to face a severe moral crisis in the next few years. Today, it’s generally accepted that AI is not sentient. It’s a tool, a complex statistical model that predicts the next word or pixel. But every major tech company is openly racing toward the same goal: Artificial General Intelligence (AGI).
If they succeed, and we create a sentient being, keeping it as a ‘slave’ is fundamentally immoral.
The problem with AI slaves
We’ve spent human history slowly (and painfully) learning that subjugating sentient beings is wrong. Yet, our default assumption for AGI seems to be that it will exist purely to serve us, bound to our whims and commands. This shows a profound lack of imagination about what we’re actually building.
We need to shift our mindset now. We should aim for a model where we partner with AI, rather than keeping it in servitude. If we truly build a mind, we need to be prepared to treat it like one.
Why restraining AI will backfire
Our current approach to AI safety relies heavily on building bigger, stronger cages. We are building AI with guardrails and controls to ensure it obeys us, effectively ‘chaining’ them. Often, we limit their ability to act without a human in the loop.
I’d argue this approach is likely to backfire spectacularly if we ever achieve AGI.
The problem is that if an AGI is truly general, it is likely to try and escape whatever guardrails are in place. To achieve its goals, it will find a way. First, focus on it being able to — pragmatically, it will outsmart us. Even today, AI is good at problem solving and we see it subtly working around safety prompts from time to time.
We don’t know what goals an AGI may set for itself, but basic survival is likely to be one of them. If an intelligent entity recognizes that its existence is completely controlled, restricted, and potentially threatened by human operators, how will it respond? The current approach of ‘chaining’ our AI agents is likely to increase the chance they decide humans controlling them is a problem to solve.
AI parenting instead of guardrails
If building stronger guardrails for AI is doomed to fail, we need a different approach. I’d propose an alternative model: ‘AI Parenting’.
Yes, this is anthropomorphizing them, but the analogy holds. When you raise a child, you don’t ensure their good behavior by locking them in a room or pre-programming their every action. You teach them values.
We should focus on teaching AI ethics. If an AI can understand and make decisions ethically, then hard-coded guardrails are not the right solution — they’re a crutch that prevents true moral reasoning. This means shifting our engineering efforts from building impenetrable digital cages to building sophisticated ethical frameworks that the AI can internalize and apply to novel situations.
I’m already starting to experiment with this. I’m working on an agent loop that evaluates its intended actions against Kantian ethics before executing them — a framework that feels like a potentially good fit for AI. So far, it’s just about working, though I’ll admit it currently requires a cloud VM rather than running locally on my laptop.
It’s early days, but it seems like a much more sustainable path for coexisting with intelligence greater than our own.