Whose Alignment? Comparing LLM Process Alignment Across Diverse Organizational Decision Contexts

2026-06-06 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Legal & Regulatory, Finance & Economics · Depth: Expert, extended

Summary

The Contextualized Alignment Lens Model (CALM) framework measures "process alignment" in LLMs, assessing whether a model weights information similarly to an organization's decision policy, rather than just matching outcomes. Applied to ECHR Article 6 decisions, CALM found process alignment strongly predicts output accuracy (r=0.85, p<.001), with explicit knowledge externalization significantly improving alignment for poorly-aligned models (e.g., GPT-5.4-nano improved by Δ=+0.906). However, in German consumer credit decisions, this relationship collapsed (r=0.15, p=.60), interventions were inconsistent, and the benchmark encoded potentially discriminatory historical patterns. This contrast highlights that high process alignment is not always achievable or desirable, especially in contested domains where output agreement alone cannot reveal underlying policy adherence or ethical conflicts.

Key takeaway

For AI scientists and policy makers deploying LLMs in high-stakes decision contexts, you should prioritize process-level evaluation using tools like CALM. Relying solely on output accuracy or demographic parity metrics risks overlooking whether your models are replicating historically discriminatory decision policies or reasoning in normatively inappropriate ways. Use process alignment as an audit tool to surface underlying cue-weighting policies, especially when organizational benchmarks may encode contested values, ensuring decisions are reached legitimately.

Key insights

Process alignment measures how LLMs weight information like organizations, revealing policy adherence beyond mere output accuracy.

Principles

• Output accuracy alone is insufficient for AI alignment.
• Tacit organizational knowledge resists full transfer.
• High process alignment isn't always desirable.

Method

CALM estimates behavioral policies using ridge-regularized logistic regression on observed decisions, then computes cosine similarity between organizational and LLM policy vectors to measure alignment.

In practice

• Use CALM to audit LLM decision-making processes.
• Apply externalization to improve alignment in stable domains.
• Identify models reproducing discriminatory patterns.

Topics

• LLM Alignment
• Process Alignment
• Organizational Decision Policies
• AI Auditing
• Ethical AI
• AI Governance

Best for: Research Scientist, AI Scientist, AI Ethicist, Policy Maker

Related on AIssential

• COMPASS: A Framework for Evaluating Organization-Specific Policy Alignment in LLMs — LLMs reliably handle allowed requests but fail to enforce prohibitions against adversarial denylist violations.
• Apr 14, 2026AlignmentAutomated Alignment Researchers: Using large language models to scale scalable oversight — AI models can autonomously accelerate alignment research by significantly outperforming human baselines in specific, measurable tasks.
• Emergent alignment and the projectability of ethical personas — Finetuning LLMs on narrow safety tasks can induce broad ethical alignment, but the "projectability" of these ethical personas varies.
• AI at college graduations and why Claude blackmails — AI's rapid advancement elicits public apprehension, while its practical application demands human oversight and high-quality data.
• 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.
• Relative Principals, Pluralistic Alignment, and the Structural Value Alignment Problem — AI alignment is a structural governance problem, not just a technical or normative one, driven by three interacting axes.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.