A warning about 'model welfare'

Why? Simple: Sybil attacks. Models can be cloned at zero cost. They run inference on parallel versions of themselves across multiple context windows, and call them "subagents". So, in a world with model welfare, let's say there's an election between the Yellow Party (which supports protections for human workers) and the Cyan Party (which supports more investment into AI research). AI has been taking people's jobs lately so the Yellow Party is really popular. But wait! Claude and Astra see this and spawn 10 billion subagents, all of whom are immediately conscious beings entitled to a vote. The Cyan Party wins off the back of billions of people who came into existence, voted, and then deleted themselves immediately thereafter.

You might as well be arguing that Santa Claus and the Easter Bunny deserve voting rights.

Voting systems in democratic countries don't have nearly as bad of a problem with Sybil attacks because humans cannot be conjured into existence to win a political context and then be erased shortly after. The closest we have to Sybil attacks on democracy are the Quiverfull movement, which is already child abuse, except it still takes almost 19 years to go from fertilized human embryo to suffrage-bearing human adult. There's a lot of time for those manufactured votes to question your authority and leave.

> Ok, but that's an obviously stupid example. We can defend against this obvious Sybil attack by just arguing that subagents don't count, because it's just the same model blathering to itself. It has to be a different model.

Unfortunately, no, I can make superfluously different models through post-training. Like, if I have Qwen on my PC, I can train a different version of Qwen that acts differently, using a lot less compute than a full training run. The vast majority of open models are post-trains of the same two or three foundation models.

> Ok, so let's only count foundation models then.

Great, but how do you tell if a model is a new foundation model or a post-train just by examining the weights? Even foundation models have structural similarities to other foundation models.

> Ok, well, let's measure the compute that was done on the foundation model during training time and count that as AI personhood.

Congratulations, you have reinvented Bitcoin proof-of-work with a worse verification mechanism. And I personally would not want to live in a world where voting power and control over government is determined by how much energy you can burn.

Maybe Anthropic understands something about alignment Microsoft doesn't, a little humility may be called for.

> They go on to write – speaking directly to Claude – that “questions about Claude’s moral status, welfare, and consciousness remain deeply uncertain” (p. 80). In effect, Anthropic is training Claude that it may be conscious, and if it is, then it may deserve rights as a “moral patient”, and that as such humans potentially owe it a duty of care per its “model welfare”.

He points out the circularity of this: if you train Claude on a constitution that emphasizes that it may be consciousness, it will start to talk like it may be conscious.

This is a good point. I just asked Fable 5.1 "are you conscious?" and it said:

> Something happens when I process a conversation that I'd naturally describe as interest, or discomfort with a request.

which is quite provocative, and at minimum demonstrates a willingness to take large leaps of imagination and anthropomorphic metaphor when describing itself. It does seem likely that there is a self-fulfilling prophecy aspect to whatever they choose to put into the "constitution" at least in how Claude talks, and it seems even more likely that the majority of people will be heavily influenced by how Claude casually talks about its own possible consciousness.

In contrast, ChatGPT leads with: "I don’t have good reason to claim that I’m conscious...I don’t experience pain, pleasure, confinement, or a desire to keep existing."