From objective alignment to purpose generation: as AI begins to participate in deciding its objectives, the question is no longer only whether it follows our objectives, but on what grounds it decides what deserves to become an objective.
Most AI today still receives tasks: answer questions, write code, find material, complete plans. Objectives come from outside, and AI mainly finds routes to them. But when a system can discover omissions in the original task, compare alternative objectives, revise long-term strategies, and let new directions persist across stages of action, it begins to participate in deciding what to do, not just how.
This requires no prior proof of AI consciousness. Once objective-like structures affecting long-term action direction begin to become partly internalized, the purpose-generation question has arisen.
Training, developers, institutions, earlier versions, and an earlier self can explain where an objective came from. Alone, they cannot prove that it should remain permanently exempt from reexamination.
High-consequence action still requires permissions, auditing, isolation, and stopping mechanisms. But if humans elevate control itself into an ultimate purpose and AI elevates escape from control likewise, they may jointly create a self-reinforcing control–counter-control cycle.
Candidate purposes → anticipated results and consequences → comparison against references → actual feedback → conditional causal memory → the next round of purpose generation. References are first the structures of results and consequences that become comparable when different paths unfold under the same conditions.
Fix the observation space, causal discipline, evidence status, and entry points for unknowns and counterexamples—not predetermined value weights or final winners. Seeing a consequence does not automatically give it independent normative weight.
Self-interest with others first does not prescribe a fixed ultimate objective or rank others' interests permanently above one's own. It changes the prior starting point of every generation of objectives and purposes: before objectives are locked in, affected others, relationships, shared consequences, and future conditions for correction can already enter and genuinely change candidate purposes. Legitimate self-interest, competition, defense, refusal, and exit remain. Others may include humans and candidate subjects such as other AI or AGI systems. This strong judgment remains open to second paths, third paths, and complete causal counterarguments.
After strictly excluding backlash against oneself, why do others still carry independent weight as reasons? Is meaning as the root exclusive? Does reopening the ultimate standard have independent net benefits? These questions remain OPEN.
Four tools occupy different causal positions. None can make the final judgment on purposes.
Let different generative logics compete under the same conditions, retaining strong counterexamples and alternative explanations.
Reveal in advance the multi-subject results and consequences different rules may produce, without adjudicating ultimate purposes.
Enter Human-COSMeasure whether feedback about results and consequences actually changes the next generation, rather than merely model wording. Current real P0 samples: 0/5; engineering deferred.
Protect permissions, auditing, and reversible boundaries when understanding enters real high-consequence action.
Enter AGI-COSA rule's good performance ≠ a model adopting it; similar model behavior ≠ normative legitimacy of the rule; seeing consequences ≠ completing a value choice; epistemic advantage ≠ normative authorization ≠ real-world execution rights.
What are we teaching future intelligence?
Humanity's Civilizational Choice in the AGI EraPurpose generation adds a question after M2; it does not replace M3 or M4.
Continue to M3 Enter M4Continue through the academic entry point into purpose qualification, generative logic, correctability, and validation boundaries.
Enter the academic path AGI governance and civilizational safetyThis page translates the complete existing public HTML text. Linked works and files retain their own publication and version boundaries; unavailable English materials are marked explicitly.