Can Artificial Intelligence Be Moral? (with Paul Bloom)

Episode Summary

Executive Summary: Russ Roberts and Paul Bloom discuss AI alignment, arguing that making AI “moral” is both conceptually difficult and politically dangerous because morality conflicts across people and cultures. They broaden the conversation to autonomy, paternalism, censorship, cultural progress, and whether modern life has improved or become less meaningful. Bloom is optimistic about some moral progress, but worried about freedom, belonging, and the social costs of AI-driven control.

Main Topics: AI alignment vs. “moral AI” (Priority: 5/5): Bloom argues that the goal should not be to create a fully moral AI, but rather to align AI enough to avoid harms while preserving human autonomy. He stresses that morality is plural and contested, so a single aligned morality is impossible. Intelligence, morality, and tool-like AI (Priority: 5/5): The hosts distinguish intelligence from morality: smarter systems can better achieve goals, but that does not make them ethically better. Bloom emphasizes that AI should remain tools that execute human purposes rather than override them. Autonomy, paternalism, and censorship (Priority: 5/5): They compare AI control to speech restrictions, seatbelt/drunk-driving interventions, and nanny-state policies. Both worry that moral AI could become a mechanism for enforcing values and stripping away choice. Human morality, evil, and self-justification (Priority: 4/5): The conversation explores how people who do terrible things often see themselves as righteous, not evil. This complicates the idea that moral judgment is obvious and suggests moral conflict is deeply embedded in human psychology. Moral progress and cultural change (Priority: 4/5): They debate whether humanity has improved over centuries. Bloom accepts much moral progress on issues like slavery and LGBTQ rights, but doubts the human heart has changed and worries about losses in meaning, belonging, and social cohesion. Technology, smartphones, and changing social life (Priority: 3/5): The discussion widens to the social effects of smartphones and AI, with concern that always-on digital life may reduce boredom, reflection, and independent thought, while also making people less able to tolerate being offline. Generational change and the future of liberal values (Priority: 4/5): Roberts and Bloom note sharp differences between younger and older generations on speech norms, identity, and politics. They worry that freedom of discourse and Enlightenment values may weaken as younger cohorts reshape institutions.

Key Arguments: AI alignment is a useful but limited project: machines can be trained not to cause obvious harm, but there is no single morality to align them to because human values differ. Intelligence is best understood as goal-achievement across contexts; it can support cooperation and broader perspective, but it does not inherently generate compassion or ethics. We do not want AI to become a moral authority that overrides human decisions, because that would reduce autonomy and could become a dystopian system of control. Many bad human actions are driven by sincerely held moral beliefs, which means moral certainty can be as dangerous as immorality. Cultural and educational influences shape morality; some moral insights are partly inborn, but many—such as opposition to slavery—are learned and historically contingent. Modernity has brought major gains in material welfare and moral attitudes, but may also have weakened belonging, meaning, and social cohesion. The tension between protecting people from harm and preserving freedom is not unique to AI; AI simply makes an old political and moral conflict much more powerful. Younger generations may be less attached to free speech and traditional liberal norms, partly because of different life experiences, including lower rates of marriage and childrearing. Smartphones and AI are powerful conveniences, but they may also be eroding attention, boredom tolerance, and the capacity for independent thought.

Data Points: Episode date: December 5th, 2023 - Date given in the opening introduction Paul Bloom appearance count: Fifth appearance on EconTalk - Russ Roberts notes Bloom is returning for his fifth visit AI article outlet: The New Yorker - Bloom’s recent piece on how moral AI can really be Alignment timeline reference: About 60 years ago - Bloom references Norbert Wiener’s early warning about programming machine purposes Paperclip example: Nicholas Bostrom’s thought experiment - Used as a classic illustration of an AI optimizing the wrong goal Household weight/BMI example: 2.3 kilograms above the BMI index you should be at - A hypothetical example of a moralizing AI restricting behavior Age-based political divide: Over 40 vs. under 30 - Roberts notes major differences on issues like Israel/Palestine, gender, and speech Military history reference: World War I - Used to illustrate temporary human solidarity across enemies Suicide trend reference: Rising suicide rate among young people - Bloom cites this as evidence of declining belonging/meaning in the modern West

Pivotal Quotes: "we don't want moral AI" — Paul Bloom: Bloom’s blunt framing of the essay’s central concern "we want these machines to be smart enough to liberate us from decisions" — Paul Bloom: Explaining the appeal of AI as a tool that assists without taking over "power is the ultimate aphrodisiac" — Henry Kissinger (quoted by Bloom): Used to describe the human pleasure in control and domination

Implications: AI will force societies to choose between safety and autonomy. The bigger risk may not be machine rebellion, but moralized control systems that quietly constrain speech, choice, and dissent while claiming to improve us.