Responsible Commonsense AI

AI is technology with high impact potential on individuals, groups, and societies. Its harmful uses are growing at a similar rate compared to its positive applications, making the responsible use of AI a societal imperative. Yet, considerations of morality, ethics, and bias remain hotly debated even for humans making their incorporation into AI non-trivial. Specifying norms detached from context is well understood, but the application of these norms in real-world contexts brings complexity. Indeed, developing AI that behaves responsibly in novel contexts has been extremely challenging, and attempts to do so have resulted in models exhibiting strong biases, and unethical behavior. This chapter reviews ongoing efforts to test and improve the models’ ability to perform moral reasoning in sensitive situations, and to measure and mitigate model biases against groups defined in terms of ethnicity, gender, and profession. A key finding is that the model biases highly correlate with biases in the data and that training with large-scale morally curated data leads to models that perform better on morality tasks. At the same time, these models are still trained to directly infer patterns from large datasets with minimal guidance–thus, it is expected that they will propagate any human biases that will be inevitably present in the data. The chapter concludes with a set of open challenges that stem from the dominant data-driven learning paradigm, such as consistency, representation fairness, and compositionality. To mitigate these challenges, the chapter suggests two key future directions of incorporating neural and symbolic (bottom-up and top-down) methods and devising comprehensive frameworks through collaborations with other disciplines.