Computer Science > Computation and Language
arXiv:2607.24750 (cs)
Submitted on 22 May 2026
Title: TimeCapsule: Generative Hallucination as a Method for Historical Sensemaking
Authors: Hayk Grigorian, Hamed Yaghoobian
Abstract
Large Language Models (LLMs) are inherently temporally overexposed: trained on vast contemporary corpora, they encode present-day concepts that render them unreliable narrators of the past. We introduce TimeCapsule, a 1.2 billion-parameter LLaMA-style causal model, trained exclusively on Victorian texts from 1800 to 1875, functioning as an epistemologically isolated generative archive. Quantitative evaluation demonstrates a 45.4% perplexity reduction over a GPT-2 baseline on held-out Victorian prose, while larger contemporary causal models achieve lower raw perplexity through broader pretraining—at the cost of losing temporal isolation. TimeCapsule exhibits a form of computational sensemaking, generating historically plausible analogical explanations for unfamiliar modern concepts. For example, it describes a computer as a "hypertrophied lung." A qualitative hermeneutic probe involving two humanities scholars revealed a crisis of authenticity: both misclassified approximately 40% of genuine Victorian excerpts as machine-generated. We argue that structural ignorance of the future transforms model hallucinations into powerful interpretive probes of nineteenth-century ontologies.
Comments
- 10 pages, 4 figures
- Accepted to Creativity and Cognition (C&C '26)
Subjects
- Computation and Language (cs.CL)
- Human-Computer Interaction (cs.HC)
How to Cite
- arXiv: arXiv:2607.24750 [cs.CL]
- Version: arXiv:2607.24750v1
- DOI: 10.48550/arXiv.2607.24750
- Related DOI: 10.1145/3803784.3807554
via ArXiv CL