TimeCapsule AI: Victorian-only model uses hallucinations for historical sense-making

Researchers Hayk Grigorian and Hamed Yaghoobian introduced TimeCapsule, a 1.2B-parameter causal language model trained from scratch on a corpus of Victorian-era texts spanning 1800 to 1875. Unlike conventional LLMs that encode present-day concepts through contemporary pretraining, TimeCapsule is deliberately isolated from modern knowledge. This temporal isolation forces the model to generate 'hallucinations' when encountering unfamiliar modern concepts — but these hallucinations are historically grounded, producing interpretations that mimic how a 19th-century thinker might reason about the future. Quantitatively, the model achieves a 45.4% perplexity reduction over a GPT-2 baseline on held-out Victorian prose, though larger modern models still achieve lower raw perplexity due to broader pretraining.

In a qualitative hermeneutic probe, two humanities scholars were asked to distinguish between genuine Victorian excerpts and machine-generated text. Both misclassified approximately 40% of authentic Victorian passages as AI-produced, revealing a deep crisis of authenticity in historical text analysis. The paper, accepted to Creativity and Cognition (C&C '26), argues that this structural ignorance of the future transforms model hallucinations into interpretive probes of 19th-century ontologies. For example, TimeCapsule described a computer as a 'hypertrophied lung' — a phrase that captures Victorian conceptual frameworks while being factually inaccurate. The work challenges the assumption that better factual accuracy is always the goal, proposing instead that purposeful ignorance can unlock novel historical sense-making.