Coproduction for algorithm adherence in AutoML applications
How to cite
Protte, Marius. 2022. "Coproduction for algorithm adherence in AutoML applications." AEA RCT Registry. AEARCTR-0010449. https://doi.org/10.1257/rct.10449-1.0
Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.
* Primary investigator
Abstract
AutoML is a promising field of Machine Learning in which even non-experts can participate in the advantages of data science and, for instance, benefit from predictions of algorithms. However, non-experts cannot necessarily evaluate the benefits of AutoML predictions and, hence, decide against those predictions and rather trust in their own expertise of their field even if it is not rational to do so (i.e., algorithm aversion; Burton et al. 2020, Jussupow et al. 2020). For instance, an engineer who is responsible for the maintenance of a production unit and worked at the machine for several years, determines the timing of maintenance based on his experience no matter what a recently installed maintenance artificial intelligence (AI) predicts. As AI is more efficient for many tasks, we investigate how to foster trust of domain experts in AI technologies. Coproduction (i.e., the engagement of customers in the production process) has shown the strived effects in the provision of services as well as classic production set-ups. There is also evidence, that algorithm aversion decreases when users can modify these algorithms (Dietvorst et al. 2018)
In our experimental study we examine whether this effect holds in an AutoML set-up. We create an environment, in which subjects act as engineers and are confronted with an AI that helps to predict malfunctions of their machine. We implement two groups (i.e., baseline group and coproduction group) and compare between those groups whether coproduction indeed fosters trust of the subjects resulting in decisions that are more conform with the AI predictions.
Intervention
2 group between-subject design: coproduction treatment (active training of AI) vs. control (passive witnessing of AI training results) --> see experimental design
Hypothesis 1: We expect individuals who actively contributed to the AI’s training to be more likely to adhere to the AI’s advice compared to individuals who were not actively involved in the algorithm’s training.
Hypothesis 2: We expect that for those individuals whose assessment of the situation differs from that of the AI, those individuals are more likely to follow the AI's advice if their initial assessment differs fundamentally from the AI's advice.
Experimental design
In our experiment, participants play in the role of an engineer that is responsible for assessing the likelihood of a production machine malfunctioning and deciding whether machine maintenance should take place in a given situation.
The experiment consists of four stages. In the first, participants read the experimental instructions and subsequently answer extensive comprehension checks to ensure a sufficient understanding of the experimental design and task rules. In the second stage, participants set their personal acceptance ranges for three machine parameters (temperature, speed and voltage) for which they deem operating the machine innocuous. In the third stage (“training stage”), the training of the AI takes place. In the control group, this training happens automatically with no decision-making authority to the participants, who are simply informed about the training’s success, i.e., the AI’s precision, in the end. In the treatment group, participants actively evaluate the training data themselves and thereby contribute to training the AI and consequently its precision. In stage four, the participants supervise the machine for 25 rounds. In each round, they decide whether machine maintenance should or should not take place. In these maintenance decisions, participants receive recommendations from the AI on the likelihood of a machine malfunction. Participants give an initial assessment of the machine status and their maintenance decision first and are then able to revise their decision after learning about the AI’s advice.
Participants receive performance-dependent monetary incentives for their decisions plus a fixed show-up fee.
Design details
Cited by
No citation data available for this registration yet.