A Test for Evaluating Performance in Human-AI Systems

Abstract

Many important uses of AI involve augmenting humans, not replacing them. But there is not yet a widely used and broadly comparable test for evaluating the performance of these human-AI systems relative to humans alone, AI alone, or other baselines. Here we describe such a test and demonstrate its use in three ways. First, in an analysis of 79 recently published results, we find that, surprisingly, the median performance improvement ratio corresponds to no improvement at all, and the maximum improvement is only 36%. Second, we experimentally find a 27% performance improvement when 100 human programmers develop software using GPT-3, a modern, generative AI system. Finally, we find that 50 human non-programmers using GPT-3 perform the task about as well as –- and less expensively than –- the human programmers. Since neither the non-programmers nor the computer could perform the task alone, this illustrates a strong form of human-AI synergy.

Interacting with Large Language Models: A Case Study on AI-Aided Brainstorming for Guesstimation Problems

“…our last study, the tool did not show significant effects on performance. This finding is consistent with other work on LLMs [36,37] that also showed that improving performance synergistically can be difficult in various human-AI interaction scenarios [37]. Although GPT-3 produces impressively reasonable results in our studies, subjects remarked that the tool often made suggestions they already thought of.…”

“…However, when human creativity and judgment are required, AI systems will support humans and increase their productivity but are unlikely to replace them completely. Good examples for this are programming or writing assistants [24,1,36,37,22,3]. When there is no clear division of labor and full automation is not the aim human-AI interaction remains challenging [18,38], and guidelines are needed [39].…”

“…Even though many tasks require human creativity and judgment, AI systems have the potential to support humans to increase their productivity and improve the overall results of the human-AI team, e.g. with programming or writing assistants [24,1,36,37,22,3]. However, designing human-AI interaction for such scenarios is not easy [18,38], and updated guidelines for interaction and interface design are required [39].…”

“…However, designing human-AI interaction for such scenarios is not easy [18,38], and updated guidelines for interaction and interface design are required [39]. This might also be why such hybrid teams sometimes do not improve performance [37]. Furthermore, the best possible hybrid team performance is not necessarily achieved by combining the most accurate systems with its user but rather those that are complementary [40,41] and use them as decision support [42].…”

“…Furthermore, the best possible hybrid team performance is not necessarily achieved by combining the most accurate systems with its user but rather those that are complementary [40,41] and use them as decision support [42]. Therefore, how the interaction of a hybrid team needs to be designed, how the team's performance is evaluated, and how successful they are is case-specific and achieving beneficial team synergy can be difficult [37]. Overall, how to best achieve human-AI synergy remains an important open research question.…”