Decision-consistent bias mediated by drift dynamics of human visual working memory

Abstract To adapt to dynamic surroundings, we need to reliably maintain sensory experiences while making accurate decisions about them. Nonetheless, humans tend to bias their ongoing actions toward their past decisions, a phenomenon dubbed decision-consistent bias. Efforts to explain this seemingly irrational bias have been limited to the sensory readout account. Here, by putting the bias in the context of mnemonic maintenance, we uncover its previously unidentified source: the interplay of decision-making with the drift dynamics of visual working memory. By taking behavioral snapshots of human visual working memory while concurrently tracking their cortical signals during a prolonged delay, we show that mnemonic representations transition toward a few stable points while initially biasing decisions and continuously drifting afterward in the direction consistent with the decisional bias. Task-optimized recurrent neural networks with drift dynamics reproduce the human data, offering a neural mechanism underlying the decision-consistent bias.