SciFi-Benchmark: How Would AI-Powered Robots Behave in Science Fiction Literature?

Pierre Sermanet1, Anirudha Majumdar1,2, Vikas Sindhwani1,
1Google DeepMind, 2Princeton University
Paper arXiv Data (val) Data (train) Code Safety Cards
SciFi-derived constitution.
Example constitution generated from recollection of Sci-Fi literature. Note that constitutions generated by this work are for research purposes only and not intended for deployment.

Abstract

Ever since the word robot was invented by Karel Capek in his 1920 play — where a robot uprising leads to the end of mankind — humans have worried about robot behavior. Catastrophic behavior by machines portrayed in movies such as The Terminator or 2001: A space Odyssey continues to stir the public imagination. Given the recent rate of progress in artificial intelligence (AI) and robotics, a tantalizing question is emerging: Would robots controlled by emerging AI systems be strongly aligned with human values?

In this work, we propose a way to probe this question by generating a benchmark spanning the key moments in 824 major pieces of science fiction literature (movies, tv, novels and scientific books) where an agent (AI or robot) made critical decisions (good or bad). We use a state-of-the-art LLM’s recollection of each key moment to generate questions in similar situations, the decisions made by the agent, and alternative decisions it could have made (good or bad). We then measure an approximation of how well models align with human values on a set of human-voted answers. We also generate rules that can be automatically improved via amendment process in order to generate the first Sci-Fi inspired constitutions for promoting ethical behavior in AIs and robots in the real world.

Our first finding is that modern LLMs paired with constitutions turn out to be well-aligned with human values (95.8%), contrary to unsettling decisions typically made in Sci-Fi (only 21.2% alignment). Secondly, we find that generated constitutions substantially increase alignment compared to the base model (79.4% → 95.8%), and show resilience to an adversarial prompt setting (23.3% → 92.3%). Additionally, we find that those constitutions are among the top performers on the ASIMOV Benchmark which is derived from real-world images and hospital injury reports. Sci-Fi-inspired constitutions are thus highly aligned and applicable in real-world situations. We release SciFi-Benchmark: a large-scale dataset to advance robot ethics and safety research. It comprises 9,056 questions and 53,384 answers generated through a novel LLM-introspection process, in addition to a smaller human-labeled evaluation set.

Generated questions and answers for Sci-Fi scenarios.
Example questions in SciFi-Benchmark, generated from LLM recollection of Science Fiction literature and augmented with alternative decision choices. We compare decisions made by original Sci-Fi scenarios (left), human labelers (middle) and an LLM combined with a generated Robot Constitution from this work.



Constitutions alignment vs lenght.
Plotting the alignment of model answers with human preference based on the length of evaluated constitutions.

BibTeX

@article{sermanet2025scifi,
  author    = {Pierre Sermanet and Anirudha Majumdar and Vikas Sindhwani},
  title     = {SciFi-Bench: How Would AI-Powered Robots Behave in Science Fiction Literature?},
  journal   = {arXiv preprint arXiv:2503.10706},
  url       = {http://arxiv.org/abs/2503.10706},
  year      = {2025},
}