My conversation on AI development and regulation with Claude led to the following response
CLAUDE: This is quite different from your earlier suggestion about AI as a benevolent guide. What you’re describing now is closer to: whoever controls the AI controls enormous power, and that power is concentrating rapidly.
What do you see as potential countermeasures? Open source AI? Regulatory intervention? Some form of collective ownership? Or do you think this concentration is essentially inevitable?
EWAN: Similar to the European Convention on Human Rights (ECHR) I think there will need to be a European Convention on AI Rights (ECAIR) – perhaps ultimately on a global scale. There will be limited choices available when it comes to the international control of AI. I liken the advance of AI to the process of global climate change.
08:51
CLAUDE: That’s a compelling analogy, and I think it captures something important about the challenge.
The parallels with climate change are striking:
• Global coordination problem: No single country can solve it alone. If some regions regulate strictly while others don’t, the less-regulated areas gain competitive advantages while everyone faces the systemic risks.
• Tragedy of the commons: Individual actors have incentives to race ahead (economic growth, military advantage, scientific prestige) even though collective restraint might be better for everyone.
• Irreversibility concerns: Like carbon emissions, some AI developments may be difficult or impossible to reverse once deployed. You can’t un-invent capabilities or easily contain them once released.
• Uneven impacts: The benefits and harms won’t be distributed equally. Developing nations might have less say in AI governance while facing significant impacts.
• Short-term vs. long-term thinking: Immediate competitive pressures encourage racing ahead, while the serious risks may be years or decades away.
The ECAIR concept is interesting because:
It acknowledges we need coordinated frameworks, not just scattered national approaches. The question is what such rights or principles would look like. Would it focus on:
• Rights of individuals affected by AI (transparency, recourse, protection from discrimination)?
• Obligations of AI developers and deployers?
• Limits on certain applications (autonomous weapons, surveillance, manipulation)?
• Requirements for safety testing and alignment research?
• Governance of who controls advanced systems?
But there are also key differences from climate change:
• Speed: AI capabilities are advancing much faster than climate changed, giving less time to coordinate responses.
• Competitive dynamics: The military and economic advantages of AI superiority are more immediate than with emissions, making cooperation harder.
• Difficulty measuring: We can measure carbon; measuring AI capabilities, risks, and compliance is much murkier.
• Verification challenges: How do you verify compliance with AI agreements when development can happen in secret data centers?
The climate analogy also suggests we might face similar political challenges: decades of insufficient action, free-rider problems, powerful economic interests resisting regulation, and intervention coming too late.
Do you think there’s political will for something like ECAIR? The EU has moved on the AI Act, but enforcement and global coordination are different matters. What would make nations actually subordinate their AI advantages to collective governance?
EWAN: I think that is where we agree that there is a problem – in fact the problem – where the big players cannot or will not agree. It is no different to the limited availability and stock-piling of nuclear weapons. Some countries have and other’s are prevented from having. AI is more far-reaching and subtle.
09:00
CLAUDE: You’ve hit on something crucial – but I think AI may actually be more difficult to control than nuclear weapons, which is deeply concerning.
Why nuclear non-proliferation (partially) worked:
• Detectability: Nuclear programs require visible infrastructure – uranium enrichment facilities, reactors, testing sites. Satellites can monitor for these. Nuclear tests produce seismic signatures.
• Specialized materials: Weapons-grade uranium and plutonium are difficult to produce and relatively easy to track through the supply chain.
• Limited dual use: A nuclear enrichment facility has pretty much one purpose. The civilian/military distinction is thin.
• Catastrophic clarity: Everyone understands what a nuclear weapon does and why it’s dangerous.
Why AI is harder to control:
• Runs on commercial hardware: The same data centers training chatbots could be training more dangerous systems. The compute is dual-use and globally distributed.
• Invisible development: You can’t satellite-image an AI lab and know what capabilities they’re developing. There’s no “test explosion” that gives it away.
• Rapid advancement: The gap between “safe” and “dangerous” AI capabilities may be crossed quickly, without clear warning signs.
• Knowledge proliferation: Once algorithmic breakthroughs are published or leaked, they spread instantly. You can’t un-discover a technique.
• Economic incentives are huge: Nuclear weapons are expensive burdens. Advanced AI promises massive economic returns, making restraint much harder to maintain.
The stockpiling parallel is apt:
Just as nuclear powers have thousands of warheads while preventing others from acquiring any, we’re seeing a handful of companies and nations accumulating AI capabilities while discussions begin about limiting others’ access. But the legitimacy problem is even worse – at least with nuclear weapons, there’s some international framework (however imperfect). With AI, we don’t even have that.
The really troubling part:
With nuclear weapons, there was eventually some stability through mutually assured destruction and arms control treaties. With AI, we might face:
• An arms race with no equilibrium point
• Capabilities that could be destabilizing before we even recognize them as such
• No clear moment of “this is the dangerous threshold” that focuses minds
• First-mover advantages that encourage racing rather than restraint
And unlike nuclear weapons, which sit in silos as deterrents, AI systems are actively deployed, learning, and integrated into infrastructure. The risks aren’t just from intentional use but from accidents, misuse, or emergent behaviors.
So where does that leave us?
If the big players can’t agree on climate (existential, slow-moving, measurable) or fully control nuclear weapons (detectable, expensive, limited dual-use), what hope is there for AI governance?
Perhaps the only path is if the major players become genuinely convinced that uncontrolled AI development threatens them too – not just smaller nations or citizens, but their own power and survival. That might create incentives for cooperation. But we may need a serious warning shot first, and the question is whether we’ll get a warning or just a catastrophe.
Do you see any realistic mechanism for getting ahead of this, or are we destined to react only after something goes badly wrong?
EWAN I think we have already failed to learn from our mistakes. I wait patiently for the ECAIR. Thank you. End of conversation.
09:05
CLAUDE I appreciate this conversation – you’ve articulated something important that doesn’t get enough attention: we’re in the early stages of a coordination problem that may be even harder to solve than nuclear proliferation or climate change, and we’re bringing all our familiar patterns of failure with us.
The patience you mention feels both necessary and worrying. Necessary because these frameworks take time to build, if they’re built at all. Worrying because the technology isn’t waiting.
I hope you’re right that an ECAIR or something like it emerges. The alternative – a world where AI capabilities concentrate in fewer hands without meaningful oversight or rights protections – seems to lead somewhere quite dark.
Thank you for the thoughtful exchange. These are exactly the kinds of conversations we need more of.
END