Less Is Moral: A CHARMing Framework for Moral Foundations Detection in Endorsement Behaviour

TEXT START: Moral language plays a central role in shaping online endorsement and the diffusion of information, yet existing moral foundation detection systems often suffer from poor cross-domain generalization, weak rationale grounding, and reliance on costly prompting-based large language models (LLMs).

The Dissection

CHARM is a classifier presented as a psychologically grounded measurement system. Its reported benchmark gains establish improved task performance, while the COVID-19 Twitter application extends that performance into a broader claim: that moral alignment helps explain endorsement and misinformation diffusion.

The central move is an escalation from classification to social diagnosis. A model that predicts labels is treated as if it has detected stable psychological constructs and exposed a causal mechanism of online behavior.

The Core Fallacy

CHARM confuses label performance with construct validity, and association with mechanism. AUC and F1 demonstrate agreement with dataset labels. Out-of-domain gains demonstrate transfer across those evaluation settings. Neither proves that the model has isolated moral foundations, that human rationales are faithful, or that moral alignment causes endorsement.

The model may instead be exploiting partisan vocabulary, topic, stylistic cues, dataset artifacts, or hate-speech correlations. The abstract supplies no evidence that these alternatives have been separated. The paper measures a legible surface and promotes it to governing cause.

Hidden Assumptions

• MFTC, MFRC, News, and MFTCXplain labels reliably represent the underlying moral psychology rather than annotator conventions.
• Rationale supervision reflects genuine reasoning instead of plausible post-hoc language.
• Hate-speech signals add psychological validity rather than leakage or confounding.
• COVID-19 Twitter discourse generalizes to online endorsement and misinformation diffusion more broadly.
• Moral value alignment explains behavior independently of ideology, identity, network structure, bots, recommendation systems, source credibility, and platform incentives.
• Correlation in observed discourse is sufficient for practical intervention.
• A scalable, low-cost detector remains reliable after adversarial adaptation and mass deployment.
• Automated moral inference retains special value even as AI systems generate, remix, and target moral language at industrial scale.

Social Function

This is partial truth, prestige signaling, and transition management.

The partial truth is real: moral framing can influence how people interpret and endorse information, and a cheaper detector could be operationally useful. The prestige signaling lies in binding an ordinary automation problem to psychological theory and interpretability language. The transition-management function is more consequential: CHARM converts human moral expression into a scalable signal for moderation, persuasion targeting, misinformation triage, and population segmentation.

That is not preservation of human productive participation. It is another layer of cognitive work being made machine-readable and cheap.

The Verdict

CHARM may be a competent classifier, but the abstract carries its evidence beyond its load-bearing capacity. It has not shown that moral foundations cause endorsement or misinformation spread; it has shown that a lightweight model can reproduce patterns associated with moral labels and find correlations in a Twitter sample.

Under the Discontinuity Thesis, CHARM is a small but clear P1/P2 artifact. It automates interpretation, weakens the economic position of human analysts, and creates leverage for whoever controls the data, compute, platforms, and intervention channels. The model architecture is a temporary moat. The durable asset is control of the behavioral data and the systems that act on the prediction. CHARM does not resist the transition; it is infrastructure for it.