Leaders in and around government debate whether technologists or subject matter experts are better suited to lead the ethical development of artificial intelligence.
Building a quality artificial intelligence system takes a diverse team made up of people from various ethnic, geographic and professional backgrounds, including programmers, data scientists and ethicists. But when putting that team together, who is the best person to lead it?
Advancements in machine learning and AI are enabling more decisions to be made by algorithms with less human interaction. But allowing machines to make decisions opens a litany of ethical issues, from the inherent bias of the people developing the algorithm to systemic prejudice in the processes being automated.
In order to combat these problems—one of, if not the chief priority for federal and military AI programs—development teams have to consider these ethical issues from the start, and continue to wrestle with them throughout production and operation.
Chakib Chraibi, chief data scientist for the National Technical Information Service, or NTIS, in the Commerce Department, suggested a five-phase approach to handling ethical questions throughout the AI development lifecycle.
“The first phase is what we’re doing now: awareness,” he said Thursday on a panel discussion during ATARC’s Role of Emerging Technology in the Federal Emergency Response Virtual Summit. “We need to learn more about the AI and the ethics and the issues that are there. I’ve been trying to develop very quickly AI solutions which have given us a lot of opportunities for federal agencies to actually save money, be more efficient and more effective. But we need to be careful about how we go about that.”
The remaining phases include identifying the specific intent of the algorithm being developed; designing the solution, at which point you’ll want to gather a diverse, cross-functional team that is representative of the problem you’re trying to solve; analyze the data and methodology for unwanted biases; and continually monitor outcomes and adjust the training data and algorithm appropriately.
To lead these efforts ethically and effectively, Chraibi suggested data scientists such as himself should be the driving force.
“The data scientists will be able to give you an insight into how bad it will be using a machine-learning model” if ethical considerations are not taken into account, he said.
But Paul Moxon, senior vice president for data architecture at Denodo Technologies, said his experience working with AI development in the financial sector has given him a different perspective.
“The people who raised the ethics issues with banks—the original ones—were the legal and compliance team, not the technologists,” he said. “The technologists want to push the boundaries; they want to do what they’re really, really good at. But they don’t always think of the inadvertent consequences of what they’re doing.”
In Moxon’s opinion, data scientists and other technology-focused roles should stay focused on the technology, while risk-centric roles like lawyers and compliance officers are better suited to considering broader, unintended effects.
“Sometimes the data scientists don’t always have the vision into how something could be abused. Not how it should be used but how it could be abused,” he said.
“The last person I would put in charge of something like this is the data scientists,” Richard Eng, chief engineer of the Applied Software Engineering Department at The MITRE Corporation, agreed.
As with banking, Eng offered medicine as another sector where domain experts—doctors and specialists—are key to understanding nuances of which technologists might not be aware.
“I’ve run across this over and over again in medicine. From a data science perspective, it looks great. Sensitivity, specificity, accuracy: all 90%. Great, we just won the Nobel Prize,” he said. “Well, no, it makes no sense from a pathophysiology perspective. … You need that expertise to push back against a data scientist.”
Eng suggested the best leader for these efforts is “an informed business owner that understands the domain and the context and the consequences—the true consequences—of the wrong answer.”
Anil Chaudry, director of AI implementation for the General Services Administration’s AI Center of Excellence, went a third route.
“It’s got to be a strong program manager with common sense—that’s really what you need,” he said. “That can help bring together the data scientists and the compliance folks.”
While the other roles are critically important to creating working, ethical AI tools, a good program manager will know how to bring it all together effectively.
“The data scientist wants to really take a risk and the compliance person wants no risk. And, somewhere in there, the program manager has to balance that in order to deliver a capability,” Chaudry said.
“I want a data scientist,” said Michael Hauck, a data scientist consulting with the Defense Department. “The reason I say that is the data scientist, to me, has all of these skillsets. … Management needs to understand: You don’t just take AI and add it to the toolbox of things in your organization. It actually requires reengineering of how software and systems are done.”