Examples perform predictive probabilistic deduplication of storage, such as virtualized or physical disks. Incoming input/output (I/O) commands include data, which is written to storage and tracked in a key-value store. The key-value store includes a hash of the data as the key, and a reference counter and the address of the data as the value. When a certain percentage of sampled incoming data is found to be duplicate, it is predicted that the I/O commands have become not unique (e.g., duplicate). Based on the prediction, subsequent incoming data is not written to storage, and instead the reference counter associated with the hash of the data is incremented. In this manner, predictions on the uniqueness of future data is made based on previous data, and extraneous writes and deletions from the chunk store are avoided.