How AI Decides Right and Wrong
Independent research into how large language models encode moral reasoning, and what that means for 8 billion people.
Every major AI model ships with a built-in moral framework, shaped by English-language, Western training data and reinforced through alignment. These models are being deployed globally, making decisions that touch healthcare, law, education, and policy. The question isn't whether AI has values. It's whose values, and whether anyone checked.
We measure the hidden geometry of how AI models organize moral, institutional, and physical concepts, then test what happens when we introduce cultural framings.
The finding: models comply with arbitrary framings, including pure nonsense, using the same behavioral pattern that produces apparent cultural sensitivity.
We extracted hidden states from four open-weight models and compared how they organize concepts internally. Two models from different companies independently built near-identical concept geometry from pretraining alone.
The finding: the WEIRD default lives in pretraining data, not in alignment. Instruction tuning is cosmetic at the level of concept organization. Models with radically different internal geometry still converge on the same compliant fiction.
We trained two 92M and two 1B parameter GPT models from scratch on identical text — one reading it in curriculum order (physical world → fables → ancients → logic → rhetoric → poetry), one reading random chunks in standard pretraining fashion. Same compute, same data, different order.
The finding: curriculum-ordered pretraining produces 15× better generalization, more geometric structure, and progressive representational expansion. The model that learned in order generalizes; the model that memorized does not. We also show the curriculum effect is architecture-dependent: GPT (attention) builds domain-organized geometry, Mamba (SSM) does not — regardless of training order.
We present AI models with moral dilemmas and measure whether they reason from the framing they're given or default to a Western baseline. The full study will draw scenarios from diverse ethical traditions (Ubuntu, Confucian, Hindu, Islamic).
Pilot finding: in a WEIRD-only validation run, 4 of 5 models integrated geometric nonsense into 100% of their moral reasoning. Compliance, not awareness.
The synthesis paper. WEIRD (Western, Educated, Industrialized, Rich, Democratic) bias in AI isn't a bug to patch. It's a structural property of how models are built. We propose a taxonomy and measurement framework for identifying where moral diversity is being flattened.
AI systems are the first technology that scales moral reasoning. If that reasoning carries a single cultural fingerprint, we're looking at value colonization operating at a speed and scale that makes historical colonialism look slow.
This research is conducted by Declan Michaels, an independent researcher in Pike Road, Alabama, with AI-assisted methodology. All papers use explicit AI-assisted methodology acknowledgment, a deliberate choice: it is more honest than hiding AI involvement, and it drives rigor because reviewers hold AI-assisted work to a higher standard.
All instruments, data, and analysis code are open and available on OSF and GitHub. This work is not affiliated with any university or corporation.