Animacy semantic network supports causal inferences about illness

We found distinctly localized neural responses to causal inferences about illness relative to both mechanical causal inferences and noncausal vignettes. A bilateral PC region previously implicated in thinking about animate entities (i.e. people and animals) responded preferentially to causal inferences about illness over both mechanical causal inferences and causally unrelated sentences in whole-cortex analysis (p<0.05, corrected for multiple comparisons; Figure 1C) and in individual-subject overlap maps (Figure 1—figure supplement 1; Figure 1—figure supplement 2). PC responses during illness inferences overlapped with previously reported responses to people-related concepts (Fairhall and Caramazza, 2013b; Figure 1—figure supplement 3).

Responses to illness inferences in the precuneus (PC).

(A) Percent signal change (PSC) for each condition among the top 5% Illness-Causal>Mechanical-Causal vertices in a left PC search space (Dufour et al., 2013) in individual participants, established via a leave-one-run-out analysis. (B) Average PSC in the critical window (marked by dotted lines in A) across participants. The horizontal line within each boxplot indicates the overall mean. (C) Whole-cortex results (one-tailed) for Illness-Causal>Mechanical-Causal and Illness-Causal>Noncausal (both versions of noncausal vignettes), corrected for multiple comparisons (p<0.05 family-wise error rate [FWER], cluster-forming threshold p<0.01 uncorrected). Vertices are color-coded on a scale from p=0.01 to p=0.00001. (D) Example stimuli. ‘Magical’ catch trials similar in meaning and structure (e.g. ‘Sadie forgot to wash her face after she ran in the heat. Now she has a cucumber nose.’) enabled the use of a semantic ‘magic detection’ task.

Relative to illness inferences and noncausal vignettes, inferring the causes of mechanical breakdown in inanimate entities activated bilateral anterior parahippocampal regions (i.e. anterior PPA), suggesting a double dissociation between illness and mechanical inferences (Figure 2; Epstein and Kanwisher, 1998; Weiner et al., 2018). This anterior PPA region is engaged during memory/verbal tasks about physical spaces (Baldassano et al., 2013; Fairhall et al., 2014; Silson et al., 2019; Steel et al., 2021; Häusler et al., 2022; Hauptman et al., 2025).

Responses to mechanical inferences in anterior parahippocampal regions (anterior PPA).

(A) Percent signal change (PSC) for each condition among the top 5% Mechanical-Causal>Illness-Causal vertices in a left anterior PPA search space (Hauptman et al., 2025) in individual participants, established via a leave-one-run-out analysis. (B) Average PSC in the critical window (marked by dotted lines in A) across participants. The horizontal line within each boxplot indicates the overall mean. (C) The intersection of two whole-cortex contrasts (one-tailed), Mechanical-Causal>Illness-Causal and Mechanical-Causal>Noncausal that are corrected for multiple comparisons (p<0.05 family-wise error rate [FWER], cluster-forming threshold p<0.01 uncorrected). Vertices are color-coded on a scale from p=0.01 to p=0.00001. Similar to PC responses to illness inferences, anterior PPA is the only region to emerge across both mechanical inference contrasts. The average PPA location from a separate study involving perceptual place stimuli (Weiner et al., 2018) is overlaid in black. The average PPA location from a separate study involving verbal place stimuli (Hauptman et al., 2025) is overlaid in blue.

In individual-subject functional ROI (fROI) analysis (leave-one-run-out), we similarly found that inferring illness causes activated the PC more than inferring causes of mechanical breakdown (repeated-measures ANOVA, condition (Illness-Causal, Mechanical-Causal) × hemisphere (left, right): main effect of condition, F(1,19) = 19.18, p<0.001, main effect of hemisphere, F(1,19) = 0.3, p=0.59, condition × hemisphere interaction, F(1,19) = 27.48, p < 0.001; Figure 1A). This effect was larger in the left than in the right PC (paired samples t-tests; left PC: t(19) = 5.36, p<0.001, right PC: t(19) = 2.27, p=0.04). Illness inferences also activated the PC more than illness-related language that was not causally connected (repeated-measures ANOVA, condition (Illness-Causal, Noncausal-Illness First) × hemisphere (left, right): main effect of condition, F(1,19) = 4.66, p=0.04, main effect of hemisphere, F(1,19) = 2.51, p=0.13, condition × hemisphere interaction, F(1,19) = 8.07, p=0.01; repeated-measures ANOVA, condition (Illness-Causal, Noncausal-Mechanical First) × hemisphere left, right: main effect of condition, F(1,19) = 4.38, p=0.05; main effect of hemisphere, F(1,19) = 1.17, p = 0.29; condition × hemisphere interaction, F(1,19) = 17.89, p<0.001; Figure 1A). Both effects were significant only in the left PC (paired samples t-tests; Illness-Causal vs. Noncausal-Illness First, left PC: t(19) = 2.77, p=0.01, right PC: t(19) = 1.28, p=0.22; Illness-Causal vs. Noncausal-Mechanical First, left PC: t(19) = 3.21, p=0.005, right PC: t(19) = 0.5, p = 0.62).

We also observed increased activity for illness inferences compared to mechanical inferences in the temporoparietal junction (TPJ) (leave-one-run-out individual-subject fROI analysis; repeated-measures ANOVA, condition (Illness-Causal, Mechanical-Causal) × hemisphere (left, right): main effect of condition, F(1,19) = 5.33, p=0.03, main effect of hemisphere, F(1,19) = 1.02, p=0.33, condition × hemisphere interaction, F(1,19) = 4.24, p=0.05; Figure 1—figure supplements 4 and 5). This effect was significant only in the left TPJ (paired samples t-tests; left TPJ: t(19) = 2.64, p=0.02, right TPJ: t(19) = 1.13, p=0.27). Unlike the PC, the TPJ did not show a preference for illness inferences compared to illness-related language that was not causally connected (repeated-measures ANOVA, condition (Illness-Causal, Noncausal-Illness First) × hemisphere (left, right): main effect of condition, F(1,19) = 0.006, p=0.94, main effect of hemisphere, F(1,19) = 2.19, p=0.16, condition × hemisphere interaction, F(1,19) = 1.27, p=0.27; repeated-measures ANOVA, condition (Illness-Causal, Noncausal-Mechanical First) × hemisphere (left, right): main effect of condition, F(1,19) = 0.73, p=0.41; main effect of hemisphere, F(1,19) = 1.24, p=0.28; condition × hemisphere interaction, F(1,19) = 3.34, p=0.08; Figure 1—figure supplements 4 and 5).

In contrast to animacy-responsive PC, the anterior PPA showed the opposite pattern, responding more to mechanical inferences than illness inferences (leave-one-run-out individual-subject fROI analysis; repeated-measures ANOVA, condition (Mechanical-Causal, Illness-Causal) × hemisphere (left, right): main effect of condition, F(1,19) = 17.93, p<0.001, main effect of hemisphere, F(1,19) = 1.33, p=0.26, condition × hemisphere interaction, F(1,19) = 7.8, p=0.01; Figure 2). This effect was significant only in the left anterior PPA (paired samples t-tests; left anterior PPA: t(19) = 4, p<0.001, right anterior PPA: t(19) = 1.88, p=0.08). The anterior PPA also showed a preference for mechanical inferences compared to mechanical-related language that was not causally connected (repeated-measures ANOVA, condition (Mechanical-Causal, Noncausal-Illness First) × hemisphere (left, right): main effect of condition, F(1,19) = 14.81, p=0.001, main effect of hemisphere, F(1,19) = 1.81, p=0.2, condition × hemisphere interaction, F(1,19) = 7.35, p=0.01; repeated-measures ANOVA, condition (Mechanical-Causal, Noncausal-Mechanical First) × hemisphere (left, right): main effect of condition, F(1,19) = 11.31, p=0.003; main effect of hemisphere, F(1,19) = 3.34, p=0.08; condition × hemisphere interaction, F(1,19) = 4, p=0.06; Figure 2). Similar to the PC, both effects were larger in the left than in the right hemisphere (post hoc paired samples t-tests; Illness-Causal vs. Noncausal-Illness First, left anterior PPA: t(19) = 3.85, p=0.001, right anterior PPA: t(19) = 2.22, p=0.04; Illness-Causal vs. Noncausal-Mechanical First, left anterior PPA: t(19) = 3.59, p=0.002, right anterior PPA: t(19) = 1.19, p=0.25).

In summary, we found distinctly localized responses to illness and mechanical causal inferences. Inferring illness causes preferentially recruited the animacy semantic network, particularly the PC.