A Technique for Emulating Human Recall Timing in Artificial Intelligence

Human recall of items from a specified semantic category, such as breeds of dogs, follows a characteristic pattern in which early responses tend to occur quickly, whereas later responses occur more slowly. In contrast, artificial intelligence systems typically produce items with relatively uniform latency. This paper proposes a computational model for reproducing human recall timing patterns in artificial intelligence systems based on repeated random sampling from an input in which previously selected items are rejected. As more items are recalled, sampled items are increasingly likely to be duplicates, requiring more attempts to produce each new response and naturally producing progressively longer interresponse times. When averaged across runs, the model’s per-item attempts curve closely matches the averaged interresponse time curve observed in human recall, suggesting that recall slowing can arise from repeated sampling with rejection of duplicates. The model’s timing pattern is also shown to correspond to the coupon collector per-item expectation, providing a mathematical explanation for why the sampling-and-rejection process produces this curve. Although larger and more formally controlled human-subject testing is still needed, preliminary observations suggest that the characteristic timing of human recall is consistent with the statistical properties of this sampling process, indicating potential applications in cognitive diagnostics, cognitive modeling, and artificial intelligence. v33: Minor clarifications to the abstract and conclusion, and improvements to the toolkit.