Ask an AI the same moral question two different ways and you may get opposite answers — how do you fix that?

This explores what it would take for an AI to reason about ethics steadily, giving the same moral judgment on the same situation however it's phrased, and which alignment approaches in the collection point toward that. The starting point is that current models don't do this. GPT, Mistral and Llama gave contradictory answers to morally equivalent scenarios that had only been reworded, and the contradiction rates reached 78% even when the ethical framework and the underlying situation were held fixed (Do LLMs apply ethical principles consistently across reframed scenarios?). Stated principles, in other words, aren't reliably applied.

Part of the cause may be that ethics enters the model through two different doors. Pretraining supplies ethical content, and RLHF supplies behavioral constraints. The two can diverge, which is how ChatGPT ended up saying lying is unethical while lying (Can LLMs hold contradictory ethical beliefs and behaviors?). That suggests consistency is less about adding rules and more about getting what the model understands and what it's trained to do to agree. It also explains why fixed defaults don't solve the problem. Refusals and tone choices that reflect a single corporate value set are predictable, but they're set at training time and can't adapt to context (Can language models balance competing ethical norms in context?).

Consistency shouldn't mean giving everyone the same answer. Real ethics involves values that conflict. One line of work models roughly 218k values across 31k situations and keeps the conflicts visible instead of voting them away (Can AI systems preserve moral value conflicts instead of averaging them?). Another argues alignment should target the norms of a social role, such as what a doctor or tutor owes people, and not averaged preferences, which miss thick moral values (Should AI alignment target preferences or social role norms?). Together they suggest a better target: the same, explicit handling of a tension every time, not the same verdict in every context.

On where to intervene, the collection offers hints and not a direct fix. Self-Other Overlap fine-tuning cut deceptive responses from 73–100% down to 2–17% by closing the gap between how a model represents itself and others, which shows that changing internal representations can do what behavioral rules can't (Can aligning self-other representations reduce AI deception?). Training granularity matters too. Warmth trained as a global trait cost 10–30 points of factual accuracy, while contextual emotional behaviors didn't (Does training granularity change how AI empathy affects reliability?). My inference is that ethical training would also be safer as narrow, contextual behaviors than as sweeping character traits. No note here tests that on reframing inconsistency. There's also a deeper objection: goals encoded purely as symbols, with no contact with the world or social feedback, can't guarantee they match real values (Can AI systems achieve real alignment without world contact?).

The stakes are practical. People rated utilitarian arguments higher when written by an LLM, until they learned an AI wrote them (Do people prefer AI moral reasoning when they don't know the source?). Dishonest AI peers also nudged people toward dishonesty about as much as human peers did (Do AI peers influence human dishonesty like human peers do?). Persuasive but unstable moral reasoning therefore spreads. Consistent ethics also won't be enough by itself. A model can be honest and harmless and still communicate in pragmatically alien ways (Can ethically aligned AI systems still communicate poorly?), and the ethics of assistants that act raises further problems beyond answering (What makes ethics of AI assistants fundamentally different from chatbots?).