San Francisco-based a non-profit Artificial Intelligence research firm OpenAI has released a paper, where it showing that long-term AI safety research requires social scientists to ensure that AI alignment algorithms succeed as actual humans involved. AI alignment works the task of making sure that AI systems reliably perform what human beings want them to do. As the OpenAI team noted, since they are attempting to behave accordingly people’s value, the most significant data will be humans’ data regarding their values.
Though, to appropriately align the advanced AI system with human values, several uncertainties related to the psychology of human rationality, emotion, and biases would have to be resolved. The analysts deemed that these can be resolved through testing where they guide the AI system to do what humans want them to perform reliably by understating human beings. This would involve on the basis of data regarding what people want from AI, and then it will train the Machine Learning models. After that, they can be optimized to do well as per these models once it trained. However, the company said it’s not simple, because humans can’t be totally relied upon when it comes to answering queries related to their values. The researcher believes that a question can be presented in several ways that can interact differently with human biases, as a result, can generate either low or high-quality answers.
So, researchers have come out with experimental debate to further solve this issue, comprising only humans instead of the Machine Learning agents. Albeit, these experiments will be provoked by Machine Learning algorithms, but they won’t include any Machine Learning systems or any sort of this background. According to the published paper, the team’s goal is ML+ML+human debates, but currently, Machine Learning is too primitive to perform the range of interesting tasks. So, the OpenAI proposed replacing Machine Learning debaters with human debaters to learn how to best organize debates in this machine-less setting.