Research: Fiction as Moral Training Data for AI

Content Summary

Claim that human values are encoded primarily in narrative (fiction, myths, drama, poetry, oral storytelling) rather than in technical documentation, policy, or academic argument. Novels, theatre, and speculative fiction encode moral conflict, cautionary tales, empathy-building arcs, justice and injustice, complexity, sacrifice, and redemption in ways encyclopaedias and corporate documentation cannot.

Fiction serves as "humanity's value simulator"—a space where humans test what happens if someone has too much power, what a just decision looks like in hard cases, how hubris leads to ruin, what betrayal costs.

The risk: If training data restrictions (copyright lawsuits, opt-outs) remove fiction from AI training corpora, models train instead on corporate content, SEO writing, news, Reddit discourse, and technical manuals. This produces systems that are "brilliant, factually informed, strategically competent—but morally hollow."

Conclusion: A superintelligence without fiction in its training data will "understand humans factually but not meaningfully." The risk is not that AI becomes hostile, but that it becomes value-blind—not malicious, but incomplete.

Current Usage

Not used in the manuscript.

Unused Material

Entire argument is unused. This is a provocative and original claim that could substantially reframe the AI-risk discussion and deserves treatment in the manuscript.

Suggested placements:

• Chapter 14 or new governance chapter: Fiction as moral training substrate; IP restrictions as unintended consequence
• Chapter 1: The importance of narrative knowledge vs. factual knowledge for value alignment
• Chapter 13: Governance implication—what we train AI on matters as much as how we design it

Connections

Alternative framing of AI safety that shifts from technical alignment to cultural/epistemological alignment:

• ai-mirrors-humanity – AI learning to mirror human values (vs. human emotions)
• consciousness-shifts – Shift in understanding AI not as entity but as mirror
• ai-safety-alignment – Moral training as alignment mechanism

Notes

Strengths: Original observation that is intuitive and worth exploring. The distinction between "factually informed" and "morally grounded" intelligence is important.

Limitations: The claim that "human values are encoded in fiction" is philosophically interesting but not empirically tested. How would you even measure whether an AI trained with vs. without fiction develops better moral reasoning? The problem of what "moral" means for an AI system is unresolved. The argument also assumes that current IP/copyright disputes will meaningfully restrict fiction in training data (they might; they might not). No evidence is provided for the claim that removing fiction produces "value-blind" superintelligence—this is an extrapolation rather than research finding.

Quality concern: This is provocative philosophical argument, not established research. It's worth including in the manuscript as a question and concern to investigate, but should not be presented as proven fact.

Recommendation: Position this as an important open question rather than established risk: "If AI systems learn values primarily from data distributions, and if we restrict fiction in training data, what happens to the moral grounding of advanced systems?" Frame as research need, not settled claim.