Multivariate pattern analysis (MVPA) has been successfully employed to advance our understanding of where and how information regarding different mental states is represented in the human brain, bringing new insights into how these states come to fruition, and providing a promising complement to the mass-univariate approach. Here, we employed MVPA to classify whole-brain activity patterns occurring in single fMRI scans, in order to retrieve binary answers from experiment participants. Five healthy volunteers performed two types of mental task while in the MRI scanner: counting down numbers and recalling positive autobiographical events. Data from these runs were used to train individual machine learning based classifiers that predicted which mental task was being performed based on the voxel-based brain activity patterns. On a different day, the same volunteers reentered the scanner and listened to six statements (e.g., “the month you were born is an odd number”), and were told to countdown numbers if the statement was true (yes) or recall positive events otherwise (no). The previously trained classifiers were then used to assign labels (yes/no) to the scans collected during the 24-second response periods following each one of the statements. Mean classification accuracies at the single scan level were in the range of 73.6% to 80.8%, significantly above chance for all participants. When applying a majority vote on the scans within each response period, i.e., the most frequent label (yes/no) in the response period becomes the answer to the previous statement, 5.0 to 5.8 sentences, out of 6, were correctly classified in each one of the runs, on average. These results indicate that binary answers can be retrieved from whole-brain activity patterns, suggesting that MVPA provides an alternative way to establish basic communication with unresponsive patients when other techniques are not successful.