Actionable AI: Enabling Non Experts to Understand and Configure AI Systems

TL;DR

Actionable AI addresses the challenge of enabling non-experts to configure black-box AI systems in uncertain settings. The authors implement a cartpole experiment with a real-time influence mechanism, testing 22 pairs of non-experts who interact with a learned agent in a two-player setup without prior instruction. They demonstrate that 14/22 teams can achieve good performance and 20/22 participants develop an actionable operating understanding, validating the framework and informing design guidelines. The work advances human-centered AI by proposing direct manipulation, visible action spaces, time to experiment, progressive learning levels, and explicit performance indicators to facilitate effective human–AI collaboration in the absence of full model transparency.

Abstract

Interaction between humans and AI systems raises the question of how people understand AI systems. This has been addressed with explainable AI, the interpretability arising from users' domain expertise, or collaborating with AI in a stable environment. In the absence of these elements, we discuss designing Actionable AI, which allows non-experts to configure black-box agents. In this paper, we experiment with an AI-powered cartpole game and observe 22 pairs of participants to configure it via direct manipulation. Our findings suggest that, in uncertain conditions, non-experts were able to achieve good levels of performance. By influencing the behaviour of the agent, they exhibited an operational understanding of it, which proved sufficient to reach their goals. Based on this, we derive implications for designing Actionable AI systems. In conclusion, we propose Actionable AI as a way to open access to AI-based agents, giving end users the agency to influence such agents towards their own goals.

Figures (10)

• Figure 1: Screenshots from the session of two players (P31 and P32) during four instances in their cartpole game. Top left: the beginning of the game. It is Level 1 of the game and the coach can see the four gates (vertical blue and red lines) that need to be passed by the cartpole in their screen. On the coach's screen, there is also additional information such as score, level, and best score. The cart is red coloured on the coach's screen as it is not positioned in the correct part of the screen, which is to the right of the blue gate to pass the gate. On the influencer's screen, the player can see the influence circles, the cartpole, and the level of the game. The cartpole does not change colour on the influencer's screen. In instance 2, the cartpole has reached the right part of the screen and we can see on the coach's screen that the gate is passing, i.e. moving from the bottom to the top of the screen where it disappears. In instances 3 and 4, we see the positions of the influencer's circles which moves the cartpole to the correct part of the screen. The correctness of the position of cartpole, is made visible on the coach's screen with the green-coloured cartpole and the gate passing. In step 4, the coach sees that it is the end of the game as it is the last gate.
• Figure 2: The figure illustrates the experimental setup which is a Zoom meeting. We see on the left part the remote participant with the coach's screen and the remote researcher. On the right side of the figure is illustrated the local participant with the influencer's screen and two researchers also present in the meeting room.
• Figure 3: Illustration of the impact of influences on the cartpole movement on two screens. On the left screen, the cartpole is stabilized with two influences of the same size. Both influences send equal "forces" that maintain the cartpole in the center of the screen. On the right screen, the cartpole is pushed to the left because of the bigger right influence
• Figure 4: Screenshots from playing the game under static influences. When no influence is added (on left), the cartpole applies the learned model under a 10% stochastic action. The trained model acts in the same stochastic configuration under small-size influences (center-left), medium-size influences (on center-right), and big-size influences (on right).
• Figure 5: Cumulative average of cartpole's lifetime over ten trials under four conditions: No Influences, Small-size Influences, Medium-size Influences and Big-size Influences. The value in trial 1 corresponds to the cartpole's lifetime (in seconds) for the first trial. The value in trial 4 corresponds to the average lifetime of the cartpole over the trials 1, 2, 3 and 4.