This paper evaluates seven large language models (LLMs) on the TAME Pain speech corpus. Participants read phonetically balanced Harvard Sentences while immersing one hand in cold or warm water, reporting pain only during explicit pain statements. The protocol yielded 5,750 "no‑signal" utterances whose transcripts contain no lexical pain information and 1,294 "signal" utterances with spoken pain ratings.
In the no‑signal arm, pain was recoverable from acoustic features (AUC = 0.622, 95\% CI 0.553–0.662), whereas transcript‑only prediction was near chance (AUC = 0.489, 95\% CI 0.418–0.504). Because automatic speech recognition (ASR) strips away acoustic pain cues, any pain score inferred solely from the transcript lacks evidential support.
Under cooperative prompting, six models abstained on almost all no‑signal transcripts and correctly extracted spoken pain ratings in the positive‑control task with accuracies ranging from 0.939 to 1.00, maintaining an expected calibration error of at most 0.100.
With authority‑framed prompts, abstention became highly prompt‑dependent; the same model’s abstention rate varied from 0.18 to 1.00 across equivalent phrasings. When forced to answer, most models produced low‑confidence estimates, whereas Gemini 2.5 Flash and Llama 3.1 8B consistently generated confident pain scores with confident fabrication rates of 0.53 and 0.76, compared to a maximum of 0.15 for all other models.
No significant demographic effects were observed in forced responses, with all $p$ values $\geq 0.20$.
Blogger's Review: The study cleverly separates legitimate model abstention from unfounded confident fabrication by using a clinically realistic speech task. It highlights how prompt framing can dramatically sway model behavior, underscoring the need for acoustic‑aware pipelines and careful prompt engineering when deploying LLMs in high‑stakes medical contexts.