OpenAI is investing in academic research to develop algorithms that can predict human moral judgments. In an IRS filing, OpenAI Inc. , the nonprofit arm of OpenAI, revealed it had awarded a grant to researchers at Duke University for a project titled "Research AI Morality. " According to an OpenAI spokesperson, this grant is part of a broader $1 million, three-year initiative aimed at "making moral AI. " Details about this morality research are scarce, with the grant set to conclude in 2025. Walter Sinnott-Armstrong, the principal investigator, and a Duke professor specializing in practical ethics, mentioned in an email to TechCrunch that he could not discuss the project. Sinnott-Armstrong and Jana Borg, the co-investigator, have previously explored AI's role as a "moral GPS" to aid human decision-making. Their past work includes developing a "morally-aligned" algorithm for kidney donation decisions and examining scenarios where people prefer AI's moral decisions. The press release states that the aim of the OpenAI-funded research is to design algorithms capable of predicting human moral judgments in situations involving conflicts among morally relevant factors in medicine, law, and business. However, capturing the complexity of morality remains a significant challenge for today's technology. In 2021, the Allen Institute for AI created Ask Delphi, a tool intended to provide ethical recommendations. While it performed well with straightforward dilemmas, minor changes in wording could lead Delphi to condone almost anything, even smothering infants. This issue arises from how modern AI systems operate. Machine learning models are essentially statistical tools.

By analyzing a vast number of examples from the web, they learn patterns to make predictions, such as associating "to whom" with "it may concern. " These models lack an understanding of ethical concepts and the reasoning and emotions involved in moral decisions. Consequently, AI often reflects the values of Western, educated, industrialized societies, since its training data comprises predominantly such viewpoints. Not all values are represented in AI outputs, especially if the individuals holding those values do not contribute to online content used for training. AI systems also incorporate a range of biases beyond Western perspectives. For instance, Delphi deemed being straight as more "morally acceptable" than being gay. The task facing OpenAI and its researchers is complicated by morality's inherent subjectivity. Philosophers have debated ethical theories for centuries without achieving a universal framework. Claude advocates Kantianism, emphasizing absolute moral rules, while ChatGPT leans slightly toward utilitarianism, focusing on the greatest good for the greatest number. The superiority of one approach over the other is subjective. Creating an algorithm to predict human moral judgments must consider all these factors. Achieving this goal is exceptionally challenging—assuming it's even feasible.