EBOOK
About
Aligned to what?
The major AI laboratories have, across the past several years, built and deployed systems they describe as aligned. Asked to name what the alignment specifies, the laboratories have produced a vocabulary of formulations that does not, on close examination, name anything at all. Human values. Shared values. Helpful, honest, harmless. Broadly distributed benefits. The collective will. Each formulation has been treated, in public discussion of AI alignment, as if it answered the question. Each, examined directly, does not.
Aligned to What? takes the question seriously. The first part of the book walks, statement by statement, through what twelve major laboratories - Microsoft, Mistral, Cohere, Meta, OpenAI, DeepMind, the Chinese regulator, xAI, Anthropic, and the European national projects - have publicly said about alignment, and demonstrates that none of it names a corpus. The second part assembles, from seventy-five years of intergovernmental treaty work, the cross-cultural corpus the question has been waiting on: nine UN core human rights treaties, six major regional instruments, the environmental and animal-welfare conventions, the first binding international treaty on AI itself. The third part closes with a procedural minimum, the disclosure floor: what any laboratory must publicly disclose for its alignment claim to be assessable by anyone other than itself.
The book does not propose a new vocabulary. The vocabulary exists. It does not propose new institutions. They exist. What it proposes is that the laboratories say what they are aligning to, by name, in the citable form the rest of the world has spent seventy-five years producing.
The empty word is empty because the values have not been named. Aligned to What? makes the naming a public question.
The major AI laboratories have, across the past several years, built and deployed systems they describe as aligned. Asked to name what the alignment specifies, the laboratories have produced a vocabulary of formulations that does not, on close examination, name anything at all. Human values. Shared values. Helpful, honest, harmless. Broadly distributed benefits. The collective will. Each formulation has been treated, in public discussion of AI alignment, as if it answered the question. Each, examined directly, does not.
Aligned to What? takes the question seriously. The first part of the book walks, statement by statement, through what twelve major laboratories - Microsoft, Mistral, Cohere, Meta, OpenAI, DeepMind, the Chinese regulator, xAI, Anthropic, and the European national projects - have publicly said about alignment, and demonstrates that none of it names a corpus. The second part assembles, from seventy-five years of intergovernmental treaty work, the cross-cultural corpus the question has been waiting on: nine UN core human rights treaties, six major regional instruments, the environmental and animal-welfare conventions, the first binding international treaty on AI itself. The third part closes with a procedural minimum, the disclosure floor: what any laboratory must publicly disclose for its alignment claim to be assessable by anyone other than itself.
The book does not propose a new vocabulary. The vocabulary exists. It does not propose new institutions. They exist. What it proposes is that the laboratories say what they are aligning to, by name, in the citable form the rest of the world has spent seventy-five years producing.
The empty word is empty because the values have not been named. Aligned to What? makes the naming a public question.