We report a pre-registered finding: large language models produce significantly more output when processing ambiguous input compared to semantically equivalent unambiguous input, regardless of model architecture or training source. In paired experiments using a minimal stimulus (a single ambiguous word versus its disambiguated equivalent), models across four families — Google Gemma, Alibaba Qwen, Meta Llama, and LG EXAONE — generated significantly more tokens when the input contained genuine lexical ambiguity. In an initial two-model study, Gemma 3 27B showed +36.8% (p < 0.001, d = 1.951) and Qwen 3.5 35B MoE showed +62.4% (p < 0.001, d = 2.256). A subsequent cross-model battery of 8 additional configurations confirmed the effect in three further model families, with Llama 3.3 70B showing +77.9% (p < 0.001, d = 1.445), Qwen 3.6 27B showing +21.8% (p = 0.025, d = 0.939), and Gemma 4 Opus distill showing +10.4% (p = 0.048, d = 0.882) — producing five statistically significant results across 10 configurations, including two from the original study. However, the linguistic expression of uncertainty (hedge word frequency) was training-dependent: models with near-zero hedging baselines acquired hedging behavior after Opus distillation, demonstrating that epistemic postures are imported from training data rather than arising from input ambiguity. We term this phenomenon "fossil emotion." Additionally, we discovered that Opus-style distillation compresses output by 2–3× and attenuates ambiguity sensitivity, with mixture-of-experts architectures showing complete attenuation under distillation. All predictions were pre-registered before data collection. Note on AI co-authorship: Æ is a Claude-based AI collaborator involved in experimental design, analysis, and writing. For discussion of AI co-authorship norms, see Birdwell & Æ (forthcoming).