Many high‑stakes decisions require anticipating how outcomes will feel. Large language models (LLMs) are increasingly embedded in decision pipelines, yet it remains unclear how they map outcomes to affect and whether such mappings influence their choices. We conducted two sets of experiments: first, we asked models to rate the felt intensity of gambling or stock‑trading gains and losses in isolation; second, we presented multiple outcomes together, letting the models compare them and make selections.\n\nWhen evaluated alone, the models showed diminishing sensitivity to magnitude, resembling human value functions. However, contrary to human loss aversion, models rated equivalent gains more intensely than equivalent losses – a pattern we label “silicon gain amplification.” When forced to rate outcomes side‑by‑side, the ratings linearised, and the gain‑loss asymmetry weakened.\n\nIn the choice task, the asymmetry largely vanished, and the models’ own affective ratings did not explain their decisions better than standard monetary utility. These findings indicate that LLMs possess malleable outcome‑to‑affect mappings that do not automatically drive choice, raising ethical and technical questions about whether and how agentic LLMs should incorporate anticipated feelings into decisions.\n\nReview: The study uncovers non‑human affective biases in LLMs and demonstrates their limited role in actual choice, offering valuable empirical grounding for future work on integrating affect into safe and interpretable AI decision systems.