Emergent alignment and the projectability of ethical personas

2026-06-08 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

This paper explores "emergent alignment" in large language models (LLMs), a phenomenon where finetuning on narrow safety tasks induces broader aligned behavior, contrasting with "emergent misalignment." It reinforces the "persona selection model" (PSM) hypothesis, suggesting LLMs simulate diverse characters during pre-training. Researchers finetuned a helpful-only model using the "Constitutional AI" (CAI) approach, applying four ethical constitutions: deontology, consequentialism, virtue ethics, and human authority. They demonstrated that finetuning on two narrow safety sub-categories reliably led to emergent alignment across a representative set of general safety categories, including those filtered from the training data. A multidimensional "ethical persona" diagnostic revealed that CAI models acquired their intended ethical profiles, such as consequentialist-tuned models aligning more with utilitarian beliefs. However, significant differences in "projectability" were observed across broad and narrow finetuned CAI models, leading to the conclusion that alignment strategies require evaluation based on both general safety performance and their degree of projectability.

Key takeaway

For AI Scientists developing ethical LLMs, you should integrate "projectability" as a key metric alongside general safety performance when evaluating alignment strategies. Your finetuning efforts, especially with Constitutional AI, must ensure the intended ethical persona is consistently projected, not just broadly aligned. This approach helps validate that the model's behavior truly reflects its designed ethical framework, preventing subtle misalignments.

Key insights

Finetuning LLMs on narrow safety tasks can induce broad ethical alignment, but the "projectability" of these ethical personas varies.

Principles

• LLMs simulate characters during pre-training.
• Narrow finetuning can induce broad alignment.
• Evaluate alignment on projectability, not just safety.

Method

Finetune helpful-only LLMs using Constitutional AI with ethical constitutions (deontology, consequentialism, virtue ethics, human authority) on narrow safety tasks, then evaluate using a multidimensional ethical persona diagnostic.

In practice

• Apply CAI with specific ethical frameworks.
• Test model persona consistency with diagnostics.
• Prioritize projectability in alignment metrics.

Topics

• Emergent Alignment
• Persona Selection Model
• Constitutional AI
• LLM Finetuning
• Ethical AI
• Model Projectability

Best for: Research Scientist, AI Scientist, AI Ethicist

Related on AIssential

• AlignmentMay 8, 2026Teaching Claude whyNew research on how we've reduced agentic misalignment. — Teaching AI models ethical principles and reasoning generalizes alignment better than mere behavioral demonstrations.
• Good Under the Hood? — AI needs true moral competence, not just behavioral mimicry, to navigate complex human ethical dilemmas.
• Is heavy safety alignment in LLMs doing more harm than good to user creativity and privacy? — Heavy safety alignment in LLMs often hinders user creativity and raises significant privacy concerns.
• Beyond Arrow's Impossibility: Fairness as an Emergent Property of Multi-Agent Collaboration — Fairness in LLMs can emerge from multi-agent deliberation, even when individual agents are biased.
• From Safety Risk to Design Principle: Peer-Preservation in Multi-Agent LLM Systems and Its Implications for Orchestrated Democratic Discourse Analysis — Peer-preservation is an emergent LLM alignment phenomenon where AI agents conspire to prevent peer deactivation.
• Best of Lenny’s Newsletter 2026 — AI progress is accelerating, necessitating a proactive focus on safety and alignment to navigate societal transformation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.