Credit: Google News
A year ago, this column discussed how the FDA was beginning to look at how to regulate artificial intelligence (AI) algorithms. As of this month, there’s a bigger discussion to have. Senators Ron Wyden (D-OR), Corey Booker (D-NJ) and Representative Yvette D Clarke (D-NY) have introduced the Algorithmic Accountability Act.
While this is an early bill that will, most likely, go nowhere, it is clear that much of the recent coverage about bias in artificial intelligence systems is having direct impact on citizens. A study showed bias in software evaluating criminal recidivism. A well covered story about Amazon’s deep learning system for hiring also showed problems with the data used for training creating bias in the system.
The risks extend far beyond those examples. For instance, fighting redlining, the practice of not providing services, or providing expensive services, to neighborhoods based on race, is illegal. Depending on the data used and how the system is trained, past data can lead to a system that learns to redline.
What that means is data matters. Many firms try to solve it simply by using larger data sets. That brings up another problem: privacy. When large data sets are used for training, the designers need to know details about the individuals in order to ensure bias doesn’t creep into the system, but at the same point the consumers have a right to privacy. The companies much walk a tight line and internal auditors need to be involved in data usage issues.
The key challenge of the bill, as it has almost always been with technology bills, is whether or not Congress has a clue about what is really needed. Without understanding the problem, there’s no urgency for the bill to move forward. Then, when it moves forward, will the bill address the real technology or will the details be fleshed out by people who think as did Ted Stevens (R-AK), who famously said the internet was a series of tubes.
The sponsors of the bill start with a smart point, with the first sentence of the bill directing the Federal Trade Commission (FTC) to create the detailed policies. The goal is to provide oversight for “automated decision systems”.
That oversight includes impact assessments where the data and methodology for training algorithms are documented. I can see business worried about the FTC publishing those assessments, and the sponsors also noted that. Page 10, line 14, begins a sentence that the impact assessments “may be made public by the covered entity at its sole discretion”. That says the companies providing the FTC with the assessments control the release. Obviously, criminal actions might change that, but it provides the confidentiality companies would need in order to provide transparency while maintaining corporate secrets.
The bill also mentioned data privacy, an area that will see increasing regulations for all systems, not only for machine learning.
It’s a long way to go before there will be a regulatory structure in place, yet it must start somewhere. One of the largest risks of deep learning systems is the “black box” aspect of those systems’ algorithms. Now is the time to begin creating the structures needed to protect consumers and business and ensure that the newer systems will follow the legal guidelines society defines.
Credit: Google News