AIs Being Designed to be Safer to Humans

Within the world of AI lurks the human prejudices that we don’t even notice in the world around us. Undesirable behavior in AIs is an issue that has come up in several situations, including facial recognition technology, and AI Twitter bots. However, scientists from the University of Massachusetts Amherst have designed a framework that should make it easier to keep AI focused on the task and less susceptible to their programmers’ inherent biases.

New Algorithms with Fairness Built In

The framework is designed to help AI algorithm developers to specify fairness and safety constraints for their work. The structure was initially tested in Brazil, with a sample size of 43,000 students to predict the distribution of grade point averages. The framework was able to avoid different kinds of gender bias in the scores and have a more standardized output.

The Key is in the Definition

Machine learning researcher and lead author of the paper Philip Thomas noted that the framework was the first of its kind to introduce this type of constraint into results. The idea of what is fair is at the core of the system. Researchers don’t need to figure out fairness immediately. Thanks to the framework’s Seldonian algorithm (named after a character in an Isaac Asimov story called Seldon), all the developer needs to define is what constitutes undesirable behavior. The system will then endeavor the AI to avoid this undesirable behavior in its processing and final results.

Ethics in AI in Focus

The framework is the beginning of a study of ethics in AI, a field that many researchers have expressed should be a priority as the design of AI systems increases in complexity. Humans have the built-in ability to understand and enforce ethical solutions to problems. However, AI doesn’t yet have that moral framework, and ideas such as fairness, validity, reliability, and transparency are alien concepts to it. One hopes that this framework will eventually be able to offer AI an ethical framework to prevent the wonder of artificial intelligence from turning into a nightmare.