Decoding morality through EEG with ML: rigour, risks, and the road ahead

The quest to decode morality through electroencephalography (EEG) with machine learning (ML) is as much about confronting the risks of opaque algorithms as it is about harnessing their scientific promise. This systematic review synthesises 17 empirical studies published between 2018 and 2025 that employed ML or deep learning (DL) approaches to EEG data within moral cognition paradigms. Following PRISMA guidelines, studies were identified across four electronic databases and categorised into eight experimental paradigms spanning three thematic branches: Evaluation, Emotion, and Prosociality. The Ultimatum Game was the most frequently applied paradigm (n = 6), while classical ML approaches dominated methodological choices, accounting for 65% of the included studies. Despite the growing interest in DL, interpretability was inconsistently addressed, with Explainable Artificial Intelligence methods applied in only one of the six DL studies and in just two of the 17 included studies overall. This review highlights three pressing methodological gaps: the need for multimodal fusion to capture richer neural signals, the integration of transfer learning to enhance generalisability, and the adoption of standardised interpretability reporting to ensure transparency. We argue that ML offers a scientifically rigorous pathway for advancing EEG-based moral cognition research, but only when models are robustly validated and their outputs rendered explainable. Without these safeguards, the promise of decoding morality risks being undermined by opaque algorithms and methodological blind spots.

The online version contains supplementary material available at https://doi.org/10.1007/s11571-026-10559-4.