Moral boundaries of AI automation

Moral boundaries of AI automation track where people treat AI replacement as categorically inappropriate even if the system is technically capable, cheaper, or more accurate. This is not the same construct as low AI literacy, general algorithmic aversion, or fear that the tool will perform badly. It is the social-feasibility boundary around work that is understood as carrying care, dignity, spiritual authority, human presence, or other forms of role meaning.

2025-friis-performance-principle gives the topic its first direct measurement spine. In a quota-matched U.S. survey of 2,357 adults rating 940 O*NET occupations, support for full AI automation rose from 30.0% of occupations under current AI to 58.3% when AI was described as outperforming humans at lower cost, suggesting that much resistance is performance-contingent. But 11.9% of occupations remained morally repugnant for AI use, especially roles centered on caregiving, therapy, spiritual leadership, and moral or relational authority.

The practical implication is that substitution forecasts need a moral-acceptability layer, not just a capability or exposure layer. 2025-friis-performance-principle reports only weak alignment between moral repugnance and technical suitability (R² = 0.134), which means technically exposed work can still face social resistance, while technically immature AI can be socially acceptable where the role lacks protected meaning.

Connections

• Connects to automation-and-substitution because social acceptance can slow or redirect replacement even when AI capability and cost favor automation.
• Connects to ai-adoption because adoption can fail for principle-based reasons that do not disappear with better model performance, usability, or training.
• Connects to ai-workforce-impact-measurement because exposure, applicability, worker desire, public support, and moral repugnance are different inputs to workforce-impact inference.
• Connects to human-ai-task-allocation because morally protected work may require human-held authority or presence even when a machine can complete narrow task components.
• Connects to ai-receptivity because willingness to use AI is not a single attitude; people may accept AI assistance while rejecting AI replacement in the same occupation.

Contradictions & open questions

• Public attitudes vs. organizational decisions. 2025-friis-performance-principle measures public support and moral repugnance, not employer adoption, legal rules, professional norms, customer behavior, or actual labor demand.
• Occupation-level morality vs. task-level design. The paper argues that moral boundaries attach to socially recognized roles, but AI systems enter through tasks, workflows, and interfaces. The wiki needs evidence on when a morally protected occupation still accepts AI support for narrow backstage tasks.
• Protection vs. inequality. Moral repugnance may protect some human roles from automation, but 2025-friis-performance-principle finds that higher-repugnance occupations tend to have higher wages and higher proportions of White and female workers, so social protection may not be evenly distributed.
• Stable principle vs. shifting norm. The performance-principle decomposition treats some objections as durable, but moral boundaries around work can shift as technology, institutions, and professional identities change.