Daejeon, July 27–31
Science fiction is widely treated as an anticipatory resource, yet its long-run patterns have rarely been studied at corpus scale. We present a hybrid LLM-expert framework for extracting technological forecasts and societal needs from 1,710 products across 157 globally distributed science fiction books, 1990-2025, coded for need category, agency impact, beneficiary level, forecast type, and magnitude of societal influence. Three concurrent transformations emerge. Imagined technologies have shifted from the body as project of intervention to the environment as condition to survive; from the individual as agent of technology to the human as subject of technological systems (Mann-Whitney U = 902, p = 0.002); and from technologies that augment the ordinary to technologies that operate at the edges of human capacity. Beneath these shifts certain concerns persist, notably war at six per cent of products across all decades, while others disappear as their referents become real. The framework demonstrates that digital humanities methods can surface temporal patterns in the speculative imagination that are invisible to close reading, with implications for how foresight practice draws on cultural anticipatory infrastructure.
Two literatures frame this work. Within digital humanities, the use of large corpora to address literary-historical questions has matured into an established practice, generating new accounts of how genres consolidate and how literary judgment persists across centuries (Underwood 2019). The field has simultaneously interrogated when computational scale obscures rather than reveals, calling for closer attention to what coded data actually represents and to the interpretive labour that quantitative claims require (Da 2019). Our hybrid LLM-expert framework responds to that scrutiny directly: large language models propose candidate forecasts, twenty researchers verify and refine, and the unit of analysis is an explicitly coded forecast-need pair rather than a latent textual feature.
Within futures studies, science fiction has long been treated as a site where sociotechnical imaginaries are publicly rehearsed and contested (Jasanoff / Kim 2015; Vint 2021), as anticipatory infrastructure that storyworlds make legible (von Stackelberg / McDowell 2015), and where narrative form shapes which futures feel possible enough to drive sustained imagination (Moore / Milkoreit 2020; Milojevic / Inayatullah 2015; Bina et al. 2017). A practical strand of this scholarship has developed science fiction as a workable foresight methodology, through science-fiction prototyping for innovation (Wu 2013; Kymäläinen 2016), educational and design applications (Zhang / Callaghan 2014; Zaidi 2019), and integration within established foresight handbooks (Popper 2008). The present paper asks a question those frames cannot answer directly: how has the shape of imagination shifted across thirty-five years of post-1990 science fiction?
Corpus. Building on the tradition of treating science fiction as an empirical resource for studying innovation patterns (Bassett et al. 2013; Michaud / Appio 2022), we assembled 157 commercially successful science fiction books published between 1990 and 2025, selected against six criteria of literary recognition: major awards, bestseller-list presence, screen or game adaptations, three or more translations, two or more editions, or print runs above 100,000. Books were required to satisfy at least two criteria, producing a corpus dominated by United States (54%) and United Kingdom (19%) authors, with the remainder distributed across fifteen further countries. The four decade cohorts contain 44, 35, 49, and 29 books respectively (the 2020s decade covers six years and is treated as provisional). From these books we extracted 1,710 technological forecasts, each treated as a coded product.
Coding. Each product receives four product-level codes — need category (84 categories), magnitude of societal influence, forecast type, and publication year — and two need-level codes propagated to products: agency impact (expanding, maintaining, or constraining human agency) and beneficiary level (self, collective, state, humanity, or planet-biosphere). Need-level codes are stated explicitly as a transparency point: they characterise need categories, not individual products, and analytical claims are scoped accordingly.
Pipeline. Books were processed in Google AI Studio using Gemini 1.5 Pro, segmented into 100-page passes with explicit instructions to draw only on the supplied text and to flag uncertainty. Three structured prompts extracted, in sequence, book-level metadata, product-level forecasts, and any documented real-world realisations. Twenty researchers manually verified every entry, refined the need taxonomy through inductive-deductive iteration, and resolved ambiguous codes through calibration workshops and consensus review.
Three transformations run concurrently through the post-1990 corpus, all crossing in the 2010s (Figure 1). From the body to the environment: clusters of needs focused on the body — physical augmentation, organ repair, biomedical intervention, biological code manipulation — fall from 1.64 products per book in the 1990s to 0.86 in the 2020s, while environment-focused needs (terraforming, hostile-environment survival, closed-habitat life support, planetary-systems analysis) climb from 0.71 to 1.35. From agent to subject: the proportion of products coded as expanding human agency drops from 67.0 per cent in the 1990s to 53.9 per cent in the 2020s, while constraining and maintaining codes both rise (Mann-Whitney U = 902, p = 0.002, r = 0.71 on per-book medians). Imagined technologies increasingly act on humans rather than for them. From the ordinary to the capacity-edge: needs that augment everyday life (communication, mobility, payment, holographic display, physical enhancement) drop from 2.23 to 1.10 products per book, while capacity-edge needs (faster-than-light flight, terraforming, megastructures, mind-backup, hostile-environment survival) climb from 1.00 to 1.24, with the steepest jump in the 2010s.
Figure 1. The three currents in the post-1990 corpus. Body-focused needs decline as environment-focused needs rise (A); the proportion of products coded as expanding human agency falls steadily (B); ordinary-augmenting needs decline as capacity-edge needs rise (C). All values are per-book products in each decade cohort (n = 1,710 products, 157 books).
Beneath these transformations a small set of imaginative concerns remains remarkably constant. War — operationalised as “waging war with advanced or autonomous weapons” — is the single most frequent need category in the corpus, appearing in six per cent of all products and in 0.59, 0.63, 0.82, and 0.48 products per book across the four decades. It is the largest stable concern, but not the only one: small-device surveillance, real-time sensory sharing, swarm inspection, and human-machine communication enhancement all hold relatively flat across the four decades. The substrate is small, but it persists across the same thirty-five years in which the three currents reorganise everything else.
Closer inspection of the 84 need categories reveals five distinct temporal shapes. Rising under stress (n = 6 needs): personal-health watch, automated crowd control, faster-than-light flight, hostile-space survival, food production without soil. Falling because real (n = 10): quantum computation declines monotonically across the period; embedded digital information, fast personal mobility, and biomedical intervention each collapse after the decade when their referents enter consumer reality. Inverted-U, peaked then faded (n = 13): mass surveillance, memory editing, and algorithmic governance all peak in the 2000s and decline thereafter — cyberpunk-era anxieties normalised into mundane cultural worry. U-shaped returns (n = 7): physical augmentation, communication enhancement, autonomous medical care, and air-water recycling each fall and then return, often with the post-2020 spike driven by recognisable real-world events. Stable foundations (n = 5): the substrate described above.
Two mechanisms can be traced directly in the temporal data. Cultural shock absorption appears as a sharp post-2020 spike in three otherwise modest needs: products coded as “keeping continuous watch on personal health” rise from 0.19 per book in 2015-2019 to 0.44 per book in 2020-2025, automated crowd control rises from 0.19 to 0.44 in the same window, and rapid autonomous medical care rises from 0.11 to 0.37. Three independent need categories spike simultaneously, in the immediate aftermath of a real-world health emergency that touched all three domains directly. Science fiction does not predict such events; it absorbs them, refracting them into the imaginative material of the years that follow.
The second mechanism runs in the opposite direction. Technologies that become real disappear from the imagination. Manipulating biological and genetic code peaks at 0.63 products per book in 2000-2004, collapses to 0.14 in 2010-2014 around the consolidation of CRISPR, and only partially recovers. Quantum computation declines monotonically across the period, with the steepest drops following IBM Q in 2017 and Google’s quantum-supremacy claim in 2019. Fast personal mobility — central to 1990s anticipation — falls to a trough in 2005-2009, the years smartphones, ride-share, and consumer GPS entered everyday life. This pattern complicates accounts of technological forecasting that track only acceleration (Kurzweil 2005): the corpus does not accelerate alongside real technology, it retreats from domains that technology has colonised, redirecting imagination toward what remains beyond reach. Existing work has shown this for the specific case of AI policy (Hudson et al. 2023); the present findings generalise the observation across eighty-four need categories and three concurrent structural shifts.
The threshold separating these moves is moral rather than capability-based. Across the corpus, realistic forecasts skew positive (61.1 per cent positive magnitude), hypothetical forecasts split (39.8 per cent positive), and fantastical forecasts skew darker (29.1 per cent positive, 18.9 per cent negative; χ² = 103.9, df = 4, p < 0.001). The more speculative the forecast, the more morally ambivalent its imagined consequences. Foresight practice that draws on science fiction as anticipatory infrastructure needs to register this asymmetry: the corpus is not a neutral mirror of possibility but a structured imagination whose moral charge shifts with speculative distance (Sætra 2024).
The 2020s decade covers six years (2020-2025) and is treated as provisional. Per-book normalisation is used throughout to account for unequal cohort sizes; all temporal claims report products per book rather than raw counts. Agency Impact and Beneficiary Level are coded at the need level and propagated to products, limiting inferential claims to need-level patterns. Inter-coder reliability was ensured through calibration workshops and consensus review.
All books in the corpus were processed in Google AI Studio using three structured prompts applied in sequence. Each prompt was applied in 100-page passes with explicit instructions to draw only on the supplied PDF and to flag uncertainty. The prompt text is reproduced below as used during data collection.
Analyze the attached PDF of [BOOK TITLE]. You may use the Internet to supplement your analysis with information about the book, its reception, and the author’s background. Provide information about the text as a whole according to the following structure: Title; Year of Publication; Author; Country of Origin; Adaptations (films, games, etc.); Number of Translations; Number of Editions; Number of Awards; Awards; Circulation; Bestseller Status; Sources (for all metrics); Belongs to Science Fiction Genre; Genre; S (Social factors influencing the narrative); T (Technological factors); E (Economic factors); E (Environmental factors); P (Political factors); V (Values factors). The work belongs to science fiction if it is part of one of these genres: hard science fiction; soft science fiction; space opera; social science fiction; alternate history; historical science fiction; military science fiction; post-apocalyptic fiction; cyberpunk; speculative fiction; new wave science fiction. For each data point, include links to corroborating sources whenever possible.
Analyze the attached PDF of [BOOK TITLE]. Based on the content of the PDF, identify and extract all descriptions of futuristic technology concepts and products. Go through all pages without exception. Form a numbered list of technology conceptions found in the text. For each concept, note the page(s) where the description appears, and provide: name of the technology; concise summary of the core concept; passages of text describing the technology; functionality, user experience, and technical features as tag lists; impact on the STEEPV context (neutral / mixed / limited / defining / significant); nature of the impact (positive / negative / neutral / mixed / none); description of the impact on each STEEPV category where present; level of detail (schematic / detailed / comprehensive / absent); OECD economic activity category; forecast type (realistic / hypothetical / fantastical); SCOPUS category; target audience (citizens, consumers, enterprises, government, other); wild card status.
Constraints: rely only on the PDF file sent to you; do not use third-party materials. Only analyse 100 pages of the book per pass; if the book has more than 100 pages, request to continue with pages 100-200, 200-300, and so on. For every 100 pages, identify at least five to ten forecasts.
Based on the content of the attached PDF of [BOOK TITLE] and information available online, identify any real-world products or technologies that were inspired by the technology concepts presented in the book. Consider the year in which the work was written when judging whether a depicted concept can be regarded as realised. For each product provide: corresponding implemented product; year of implementation; sources (links to articles, interviews, etc., confirming the link); documented influence of the concept from the book on the creators of the product (quotes, facts, description of influence).