Looking for AI without the bias
- By Matt Leonard
- Apr 06, 2017
What: “An Intelligence in Our Image: The Risks of Bias and Errors in Artificial Intelligence,” a report by Rand Corporation.
Why: The rise in artificial intelligence means that algorithms are behind many of the decisions being made in society. Although AI has potential to greatly increase the efficiency and speed work, allowing algorithms to make decisions for people carries ethical implications, which must be understood.
Findings: The report discusses some of the failings of AI-based decision making and problems with algorithmic errors and bias, then examines some ways for address these problems.
“Misbehaving algorithms” can lead to “incorrect, inequitable, or dangerous consequences.” AI in sentencing software has been shown to have “an extreme systematic bias” that gives black people higher risk scores than whites. Some AI-based technologies have picked up unfortunate habits -- IBM’s Watson learned to swear, for example, after ingesting some “unsavory data.”
The quality of the AI depends on the quality of the system’s data diet. If humans are the source of the modeling data, then the AI will have the biases of those subjects. Researchers are working on ways to ensure algorithmic equitability, but the “schemes will often trade some predictive power for fairness.”
Takeaway: The public’s anthropomorphized view of AI may have the unexpected benefit of fostering public understanding “that artificial agents, like humans, are not above bias.” Let’s hope the future is more Siri and less Skynet.
Read the report here.
Matt Leonard is a former reporter for GCN.