Abstract
<p>Decoded neurofeedback (DecNef) has been widely interpreted as evidence that humans can learn to regulate specific neural representations without awareness of the reinforced target. However, this conclusion depends critically on how awareness is measured. The present study revisited this issue using two complementary approaches. First, we conducted a meta-analysis of post-training forced-choice identification tasks from DecNef studies to assess whether participants could explicitly identify the trained target. Second, we applied large language model (LLM)-based semantic analyses to participants’ verbal strategy reports to test whether these reports contained information related to the DecNef target or trained condition. The forced-choice judgments meta-analysis showed little evidence for above-chance identification of the DecNef target, consistent with the standard interpretation that participants lacked explicit awareness of the reinforced target. However, Bayesian evidence for chance-level performance was only moderate, indicating that existing awareness measures remain limited in sensitivity. In contrast, semantic analyses revealed that participants’ verbal reports contained information predictive of the DecNef experimental context and trained condition, indicating that their strategies may have been indirectly aligned with the reinforced neural representation. Together, these findings support a correlated-hypothesis view of awareness in DecNef studies, whereby participants may fail to explicitly identify the target yet still use strategies or internal states that are indirectly related to the reinforced neural pattern. Future DecNef studies should therefore move beyond single forced-choice tests and adopt more sensitive, multidimensional awareness measures.</p>