Computational Analysis on Machine Consciousness in Science Fiction

Research Design and Corpus

Metaphors used in narratives to depict technologies influence how we speculate, perceive, and engage with both technologies and the world around us. The interplay between cultural imagination and technical progress forms a feedback loop, where stories about machines shape how we build them, and the machines we build reshape the stories we tell.

This project grew out of a concern with that feedback loop: if science fiction about machine consciousness is recycling the same human-centered and apocalyptic representations, the genre may be narrowing rather than expanding the imaginative space available to us as we actually build artificial minds. My specific question was whether we keep telling the same dooming story, whether the genre is stuck in a loop of fear and dystopia, or whether there is more ideological diversity than the most canonical texts suggest. To investigate this, I built a curated corpus of 115 science fiction texts centered on machine consciousness and ran four computational methods to trace how the discourse has shifted, or failed to shift, across two centuries.

The corpus spans 1816 to 2022, drawing on Project Gutenberg's robot fiction collection alongside hand-selected canonical works like Frankenstein, R.U.R., and Do Androids Dream of Electric Sheep?, as well as contemporary literary science fiction by Kazuo Ishiguro, Ted Chiang, and Becky Chambers.

The Finding

Across four independent computational methods, the genre does not turn out to be uniformly dystopian. Instead it organizes around what I call a care/threat axis. When science fiction describes machines through violence and control, the portrayal is consistently threatening. When it describes emotional bonds between humans and machines, the portrayal shifts toward hope. The dooming story is real, but it is concentrated in specific thematic registers, not spread evenly across the genre.

Topic Modeling

NMF topic modeling across the corpus surfaces the most theoretically significant finding of the project: a strong correlation between destroyed agent status and a topic I call Interiority, Emotion and Narrative Voice. Frankenstein, Galatea 2.2, Machines Like Me, and He, She and It, all texts in which the artificial being is destroyed, concentrate on this topic. The narratives most richly invested in machine consciousness as subjective experience are precisely the ones that destroy their machines. The genre imagines consciousness most fully when it is about to be extinguished.

Zero-Shot Classification

Zero-shot classification let me define categories directly, rather than finding whatever topics emerge from vocabulary alone, and test the care/threat axis explicitly. Machine violence passages score highest on dystopian/threatening of any thematic category, nearly double the next highest. Human emotional attachment passages are the only category where hopeful scores approach dystopian ones. Fear organizes around violence and control, hope organizes around attachment and relationality.

Fine-Grained Emotion Detection

Emotion detection confirms the pattern through a third, independent method. Fear and disgust together vastly outweigh joy across the corpus. Dystopian passages score highest on fear and anger, while hopeful passages score highest on joy and calm, a near-perfect alignment with the zero-shot results despite the two methods working in entirely different ways. Female-authored texts also register the same threat differently: more disgust, anger, and sadness, where male-authored texts register more fear. Male authors tend to frame machine consciousness as an external threat. Female authors frame it as violation, injustice, and loss.

SBERT Embeddings and UMAP Visualization

Where topic modeling finds thematic patterns in vocabulary, SBERT encodes documents as dense semantic vectors that capture meaning beyond keyword overlap. Frankenstein's five nearest neighbors are Erewhon, The Time Machine, The Sandman, The Future Eve, and Winterson's Frankissstein, a contemporary novel explicitly in dialogue with Shelley. SBERT detects what is a deliberate intertextual relationship spanning 200 years.

The UMAP visualizations showed that consciousness model, my primary metadata category, does not strongly predict semantic clustering. What drives semantic clustering instead appears to be period and generic register, the Victorian and Edwardian prose style that groups Frankenstein with The Time Machine, or the 1950s pulp vernacular that spreads across the center of the plot.