Machine learning is the driving force behind data-driven, adaptive, and increasingly intelligent products and platforms. Algorithms of artificial intelligence (AI) systems, such as Content Recommender Systems (CRS), intertwine with users and content creators, in turn shaping viewer preferences and the available content on these platforms.
However, the current design and evaluation methodologies of these AI systems do not adequately address how users and AI systems mutually influence each other. For instance, training large static datasets using supervised learning doesn’t illustrate how the AI system modifies the operational environment. Additionally, the deployment of AI systems without considering their systemic response can detrimentally affect overall performance and societal condition.
To overcome such shortcomings, researchers from four universities, Cornell University, the University of California, Princeton University, and the University of Texas at Austin, proposed the concept of Formal Interaction Models (FIM). This mathematical model explicates how AI systems and users mutually effect each other.
FIM functions as a dynamic system entangling the AI system with the users, thus improving the design and evaluation of AI designs. In particular, it includes four significant applications; it stipulates interactions for implementations, monitors with empirical analysis, anticipates societal impacts via counterfactual analysis, and regulates societal influence through interventions.
The model also facilitates the creation of additional metrics to quantify societal impacts, which in turn can ensure superior design objectives. These metrics can be optimized using supervised learning or RL-based algorithms for controlling societal effects. While certain societal impacts can be evaluated with a single parameter of FIM, others might emerge as a complex amalgamation of multiple parameters.
The researchers have further discussed methods of measuring value, optimizing downstream user welfare, and enhancing ecosystem health through mechanism design tools and recommender systems. They conducted multiple analyses, focusing on anticipating and controlling societal impacts, and introducing new strategies.
The incorporated model designs are primarily homogeneous with a distinguishable gap between viewer and creator interactions. Notably, dynamic models have been avoided due to their tendency to create feedback loops due to viewers’ feedback on the recommended content and its usage.
In conclusion, the Formal Interaction Models (FIM) represent a mathematical model, proposed by researchers from four universities, that formalizes the mutual shaping of AI and users. The model’s dynamical systems language is utilized to highlight the limitations in the use cases for future works. They encourage further exploration and research into this area. Whilst FIM may appear complex, its application and resultant benefits for societal impacts, downstream user welfare, and ecosystem health are evident. It is a promising step towards creating a more aligned, balanced, and mutually beneficial interaction between AI systems and their end users.