Short answers, pulled from the story.
The ethics of artificial intelligence covers topics such as algorithmic bias, fairness, accountability, transparency, privacy, regulation, machine ethics, AI safety, autonomous weapons, and existential risk. It also addresses questions about AI welfare and the possibility of machine consciousness.
COMPAS is a program used in the United States justice system to predict which defendants are likely to reoffend. Although it is calibrated to produce the same overall error rate across racial groups, Black defendants were found to be almost twice as likely as white defendants to be falsely flagged as high-risk, and half as likely to be falsely flagged as low-risk.
The EU Artificial Intelligence Act entered into force on the 1st of August 2024. It follows a risk-based approach and becomes fully applicable 24 months after entry into force. The Act either prohibits high-risk AI applications or requires them to meet specific requirements before being placed on the market.
Training large AI models produces greenhouse gas emissions estimated at around 626,000 pounds of carbon dioxide per training run. Data centers also consume approximately two liters of water per kilowatt-hour of energy used, and around two-thirds of data centers are located in water-scarce regions. Additional concerns include electronic waste containing hazardous materials such as lead and mercury.
Writing in 1976, Weizenbaum argued that AI should not replace people in roles requiring authentic empathy, including judges, therapists, soldiers, and nursemaids. He warned that placing machines in these positions would leave people alienated and devalued, and that even entertaining the possibility reflected what he called an "atrophy of the human spirit that comes from thinking of ourselves as computers."
Machine ethics is the field concerned with designing AI systems that behave morally or as though moral. Key contributors include Wendell Wallach and Colin Allen, who wrote Moral Machines: Teaching Robots Right from Wrong; Stuart Russell, who proposed that beneficial AI should pursue human preferences rather than fixed goals; and Nick Bostrom and Eliezer Yudkowsky, who have debated the relative merits of transparent decision trees versus adaptive machine learning.