By ITU News
As artificial intelligence (AI) technology continues advancing at lighting pace, discussions on the need for governance, standards, and a stronger focus on “responsible AI” have followed.
While AI can carry out decision-making tasks efficient, it’s still based on algorithms that respond to data models. Unlike humans, AI algorithms can’t see the full picture, in part because they lack emotional reasoning and other human qualities, such as empathy, ethics, and morality.
Concerns over privacy and discrimination are on the rise as AI becomes further integrated into decision-making processes that affect economies and societies worldwide.
The time has come, therefore, to decide what sort of policies should guide AI design and use, and how to make sure AI use improves human welfare and respects human dignity, said Principal Tech Evangelist for Amazon AI, Nashlie Sephus in a recent AI For Good keynote.
According to AI Principles put forward by the Organisation for Economic Co-operation and Development (OECD), responsible AI is “innovative and trustworthy” and “respects human rights and democratic values.”
Sephus describes six dimensions of responsible AI: privacy and security; fairness; explainability; robustness; transparency; and governance.
The main hindrance to making AI responsible in practice, she adds, is the need to focus on a few priorities while keeping up with new innovations.
“It can be a challenge to figure out how we can do responsible AI in practice when you have so many different use cases and technologies being released every day,” she said. “We also want to make sure that we’re covering all our bases, so there’s a lot at stake here. As innovation continues to move so fast, responsible AI should move just as fast, if not faster.”
Setting priorities for responsible AI
For Sephus, “safety of life” technologies in healthcare, law enforcement, and transport should be tackled first.
She also advises companies to focus on one problem at a time, depending on their technology’s use case: “It really pays to understand the total environment that your system is going to be deployed in, and how that environment may or may not change over time.”
The varied, case-by-case nature of responsible AI, however, will continue making it hard to implement.
“Everything is use-case specific,” said Sephus. “The way you define success or fairness metrics depends on what you’re doing. Sometimes you have systems that do multiple things, so making sure you define your bias evaluation for every single feature your system encompasses is crucial.”
Questions about the root cause of bias in AI outputs add another layer of complexity.
Standards and governance needed
Governments play an important role in implementing responsible AI, especially when it comes to high-risk, life-or-death use cases. “For example, an autonomous vehicle system can actually put someone’s life at risk,” Sephus noted. “We want to make sure that governments are prioritizing the things that are really critical at this moment.”
Asked about the most urgent legal issue around AI, she responded: “The challenge is the government figuring out how to reign in and capture everything that’s going on. How do they make this applicable across the board? From computer vision to language processing to recommendation systems – each has nuances in it that you may or may not be able to correct for if it’s already out there.”
One way governments can increase accountability in AI is through documentation, which is tied closely to accountability.
“We must hold people accountable by putting systems and mechanisms in place that will enforce these things, not just say them and leave it to the technologists to govern themselves,” Sephus said.
Education is another key driver of accountability, making ethics courses an essential part of any well-rounded AI curricula. “When people know better, they do better,” said Sephus. “Let’s create standards on a global scale to help educate people. That can happen at the university level, but it can start as early as high school.”
Yet access to AI remains relatively exclusive in today’s world. Many communities still lack access to the Internet, let alone AI platforms. In effect, the poorest and most vulnerable segments of humanity are also being left behind in this most transformative aspect of digital economies and societies.
“A lot of the time, these are target groups who are impacted by the lack of responsible AI,” Sephus explained. “It’s important for us to recognize and include those people.”
Making the AI community more diverse is one way to foster inclusion, she added. “A more diverse generation of leaders in machine learning is crucial. As the industry heads in this direction, we should do a better job at training more diverse populations, and large industries have a responsibility to contribute to that.”
Responsible AI by design
During a live Q&A session, AI for Good participants asked how start-ups with scarce resources can prioritize responsible AI from the very beginning, such as by building it into their product or service design.
“Start-up life is very competitive,” answered Sephus, advising AI-adopting businesses to “tell your investors, customers and stakeholders” about the benefits. “I guarantee you 100 per cent that will put you ahead compared to your competitors.”
Amazon Web Services provides a range of tools to help its customers practice responsible AI. Sagemaker Groundtruth, for example, helps customers label their data and improve their overall data quality through the training and integration of human annotators.
Watch the full keynote:
While AI tools can assist with problem-solving, achieving responsible AI still requires human collaboration, Sephus concluded. “While I’m in the AI industry and know that it has many capabilities, it still cannot replace two humans talking and collectively reaching an accord to figure out how they can solve a problem.”
Watch AI For Good’s upcoming webinar “How to make AI more fair and unbiased” to find out how multi-task discrimination, discrimination under class imbalance, and distribution shifts can help reduce bias in AI.