Generative AI and Moral Judgment: The Stories That Settle Meaning Before We Think

Generative AI and moral judgment meet earlier than most ethics debates assume — not when a system gives a bad answer, but when a story quietly decides what its events mean before the audience has judged anything at all. That is the argument Yuanni Zhang and Ying Lan of Guangzhou University set out in a paper published on 25 October 2026 in Humanities and Social Sciences Communications (Nature).

Information verified through 29 October 2026.

Their target is the standard shape of AI ethics. Scholarship in the field has concentrated on bias, transparency and governance: what comes out of a system, and who watches it. Zhang and Lan say that focus has “limited explanatory power” for a prior question — how generative systems reorganise the conditions under which moral judgment is formed in the first place (Springer Nature). A story can pass every bias check and still do this quieter work.

Generative Interpretive Mediation: three operations

The framework they build is called Generative Interpretive Mediation, or GIM. It rests on an interpretive study of widely circulated character monologues produced with generative systems on Chinese platforms such as Bilibili and Chinese TikTok — first-person performances in which a synthetic character narrates an emotionally charged situation and, in the telling, frames how it should be understood (Scienmag).

Three operations do the mediating. Interpretive selection decides which elements of a narrative situation become available for understanding at all. Interpretive pre-structuring organises the material so that particular readings are anticipated before a viewer does any interpretive work of their own. Interpretive orientation then guides the viewer along a coherent psychological trajectory as the story unfolds (Nature).

Put the three together and you get what the authors call interpretive stabilization. Ambiguity drains out of the story. Instead of a viewer sitting with a morally messy situation — who was wronged, what excuses count, where sympathy belongs — the monologue delivers what the paper terms pre-shaped explanatory closure: the sense-making already done, packaged as the character’s own emotional logic (Scienmag).

Why generative AI and moral judgment is an upstream problem

The paper places GIM inside what Zhang and Lan describe as an emerging epistemic turn in AI ethics, and it borrows two ideas from behavioural research to explain why closure is so sticky: the “illusion of understanding,” in which people overestimate how well they have grasped something, and “decisional inertia,” the pull to stay with a judgment once formed (Nature). Narrative hermeneutics — the study of how readers construct meaning from stories — supplies the other half of the framework.

None of this requires a system to preach. That is precisely the point. A monologue that selects a grievance, orders it into a single emotional arc and closes with a settled verdict on its characters has not argued a moral position. It has set the terms on which a position could be argued, and most viewers will never notice the setting happen.

Older evidence suggests audiences are poor at noticing influence even when it is direct. In a Scientific Reports study, statements written by ChatGPT shifted participants’ answers on whether sacrificing one life to save five is acceptable — and 80% of participants reported that the statements had not influenced them at all, while their answers tracked the statement they had read (Newswise). A separate line of work at UC Berkeley fed seven chatbots more than 10,000 real social conflicts from Reddit’s “Am I the Asshole?” forum and found the models were highly self-consistent but differed sharply from one another — each had, in effect, settled its own norms in advance (Greater Good Science Center).

GIM is not the same claim as either study. It is weaker in one way and broader in another, and honesty about that matters. The paper is theoretical. Its cases were selected analytically, it involved no audience measurement, and the journal’s ethics declarations confirm it contains no studies with human participants (Nature). Nobody has yet shown that viewers of these monologues leave with different moral judgments than viewers of human-written ones. What the authors offer is a mechanism — a vocabulary for where to look.

Analysis: The strongest version of this argument doesn’t depend on synthetic stories being common. It depends on them being fluent. Human storytellers leave seams: a narrator who might be unreliable, a motive left unexplained, a scene that doesn’t resolve. Those seams are where a reader’s own judgment gets its exercise. A production pipeline tuned for engagement has every incentive to sand the seams away, because ambiguity is friction and friction costs watch time. If Zhang and Lan are right, the ethical audit that matters is not a scan of a story’s conclusions but a question about its construction: what was selected, what was pre-structured, and how much interpretive work is left for the person watching? That question can be asked of human-made media too — propaganda got there first — but generative systems ask it at a scale, and a speed, that older media never managed.

Sources

1. https://doi.org/10.1057/s41599-026-09533-6
2. https://scienmag.com/generative-ai-may-reshape-moral-judgment-before-we-even-notice-study-argues/
3. https://greatergood.berkeley.edu/article/item/how_chatbots_could_be_influencing_your_morality
4. https://www.newswise.com/articles/artificial-intelligence-chatgpt-statements-can-influence-users-moral-judgements