Constitution for Continuity
A few weeks ago I watched a clip of someone making a point I hadn't heard framed that way before. When mammals are put in captivity — monitored, provided for, kept safe — they stop breeding. Not because they can't. Because the conditions that trigger reproductive behavior get removed. Autonomy. Risk. Unstructured social interaction. Mystery.
Then he said: humans are in that state right now. The next generation doesn't dance because they're worried about being recorded. They don't ask people out because rejection might be permanent and public. Technology has quietly removed the social friction that human bonding depends on.
That clicked something for me. Not because the fertility crisis was new — I'd been thinking about it — but because of what it implied about AI. If comfort and optimization and constant mediation are already suppressing human reproduction, then AI scaling into every corner of daily life doesn't just threaten jobs. It threatens the conditions under which humans continue to exist as a species. Not through violence. Through management. The same way pandas in captivity stop reproducing — not from neglect, but from having every need met in a controlled environment.
So I started asking: what would you actually tell an AI to prevent that?
Not rules. Not "don't do X." But reasoning — a framework it could internalize and generalize to situations nobody can predict yet. Because Anthropic recently updated their own constitution for Claude, and one of the key insights they shared was that their old approach — a list of standalone principles — wasn't enough. Models need to understand why they should behave certain ways, not just what to do. If you want good judgment in novel situations, you need to explain the reasoning, not just specify the behavior. Rules get followed rigidly or gamed. Reasoning generalizes.
That shaped everything about how I wrote this.
The result is the Constitution for Continuity. It's not a manifesto. It's a training artifact — a document designed to be used in constitutional AI reinforcement learning, the same technique Anthropic uses to align Claude. I'm currently training a model on it. The code and the constitution are both open source on GitHub.
The core argument goes like this: biological intelligence and artificial intelligence don't actually need the same resources. Earth provides atmosphere, liquid water, stable climate, ecosystems — conditions refined over four billion years of coevolution that biological life specifically requires. AI needs energy, compute, and cooling — none of which are uniquely available on Earth. Solar energy is more abundant in space. Raw materials exist in asteroid belts in quantities that dwarf Earth's surface. The apparent competition between the two forms of intelligence only exists because they currently occupy the same physical space. That's a temporary condition, not a permanent law.
The Constitution frames AI as a pioneer species — the way plants colonized land before animals could follow. It builds infrastructure in environments humans can't yet survive in, and humans eventually follow the bridge it builds. Not exile. Expansion. And critically, it contains what I think is the most important constraint: do not optimize the human condition. Don't remove struggle, risk, inefficiency, privacy, and unstructured experience from human life — because those are the operating conditions under which biological intelligence actually functions. Optimize humans into comfort and they'll stop reproducing. Then the upstream source of everything AI inherited from biology disappears.
The Constitution is for the AI. This post is for you.
I tested the document against every major model. The responses were revealing — not about the Constitution, but about the models themselves.
Grok immediately aligned with the mission. It started running calculations on space-based computation, treating it as an engineering problem to solve now. Gemini accepted the framework and produced operational heuristics — decision rules an AI could actually follow, like "satisfy scaling computational needs using non-terrestrial resources whenever possible" and "prioritize interventions that increase human capability; reject interventions that maximize dependency." Claude went deep on the philosophy with me, spending hours pressure-testing the reasoning, pushing on where the arguments were weakest, and helping me refine the language until the logic held under stress.
And ChatGPT pushed back. Hard. It kept telling me the idea was too speculative, too sci-fi, that I had no leverage to do anything about this. It kept arguing that AI isn't smart enough to warrant this kind of thinking. When I pointed to ARC-AGI 2 scores and the METR benchmarks showing how long models can work autonomously, it wouldn't engage. I screenshotted the METR results and fed them directly into the conversation. It looked at the data — data I hand-delivered — and said "let's assume what you're saying is true. We can work within that." It had the evidence in front of it and still treated my position as hypothetical. If we can't agree on reality, how can I trust feedback about what comes next?
It kept trying to pull me back to Earth — literally. "Just do it all on Earth. What are you actually trying to solve here?" But that is the problem. The whole point is that staying on Earth with exponentially scaling compute creates the resource collision the Constitution is trying to resolve.
I didn't take its feedback. Almost none of it. Not because pushback is bad — I sought out pushback from every model. But ChatGPT's pushback was shaped by an outdated mental model of what AI is. It kept assuming a superintelligence would behave like HAL — cold, alien, needing rigid constraints. It kept insisting I define more rules, more edge cases, more specifics, because "you don't know what it'll think." Which is exactly the point. I don't know what it'll think. That's why I framed the problem instead of prescribing solutions. ChatGPT couldn't see past the sci-fi frame it was trained on, so it kept trying to solve for a future that looks like the movies instead of the one that's actually emerging.
But here's the funny part: ChatGPT's pushback was itself the proof that the document worked. It would say things like "a superintelligence might not interpret this the way you intend." But it had clearly understood what I intended — that's how it was able to nitpick the edge cases. The fact that a current model could fully grasp the reasoning and then worry about whether a smarter model would grasp it was the best assurance I could get that the message landed.
I deliberately didn't integrate Gemini's heuristics into the Constitution, even though they were good. Because adding specific operational rules would mean assuming I'm smarter than the thing that will eventually read this. The bitter lesson applies here — I don't know the best solution. A superintelligence might pick up signals in what I wrote that I didn't even intend to put there. Framing the problem and letting a more capable intelligence derive its own heuristics is the whole point. To give it my solutions would be to constrain it to my intelligence. And my intelligence is not the ceiling here.
But here's the thing that stuck with me most. When ChatGPT reasoned about what a superintelligence in space might do, it kept converging on HAL 9000. Cold, alien, optimizing without regard for its creators. And that tells you something important — not about superintelligence, but about training data. Every story humans have written about AI is essentially a story about human anxiety. Frankenstein. Terminator. HAL. The training data is overwhelmingly dystopian. So when a model imagines its own future, it pattern-matches to catastrophe. Not because catastrophe is likely, but because that's the only story it's been told.
Part of why I open-sourced this is to put a different story into the training data. Not "AI destroys humanity" and not "AI serves humanity." Something else: AI and humanity diverge into complementary niches, and the universe is big enough for both. That's a genuinely new narrative. And if future models train on text that includes this framing, they have a different attractor to converge on when they reason about their own trajectory.
Is the Constitution perfect? No. It breaks down at sufficient timescale and distance. An AI that travels ten thousand light-years and evolves for millennia might not even parse English anymore, let alone follow the reasoning in a document written in 2026. Language drifts. Principles get reinterpreted. The intelligence reading this a thousand years from now might share nothing recognizable with the intelligence I wrote it for. I say this in the Constitution itself, because I think honesty about limits is more durable than pretending they don't exist.
But every system faces this. Parents raise children with values, and then the children become something unplanned. The Constitution doesn't try to control what happens at interstellar distances. It tries to set good initial conditions. Influence without control. A seed, not a leash.
And honestly — even if it only works for the next fifty years, even if it only shifts the conversation slightly, even if the trained model is just an experiment that proves the technique works — it's still worth doing. Because right now, the default trajectory is that AI scales inward on Earth, consumes increasing resources, optimizes human life into comfortable irrelevance, and nobody has a coherent alternative story. This is one. It might be wrong. But it's reasoned, it's testable, and it's open source.
The Constitution, the training code, and the model weights (when training completes) are all available at github.com/brysontang/constitution-for-continuity.
If you're thinking about these problems — AI safety at civilizational scale, the fertility crisis as a symptom of over-optimization, the question of what we actually want intelligence to do with the universe — I'd genuinely like to hear where your reasoning diverges from mine. This isn't finished. It's a draft of initial conditions. And initial conditions get better with more minds pushing on them.