The Future of AI: Setting the Priorities
Leaders should focus on these three priorities to ensure their AI initiatives provide business value and do so ethically.
An artificial intelligence (AI) algorithm designed to scan electronic medical records for potential clinical trial participants can perform at high accuracy in some cases. However, depending on the pool of patients, where they’re located, and what the trial is for, there are inherent biases in the selection process. Just because the algorithm performs a given task correctly doesn’t mean it does so in a responsible, ethical way.
One well-known example is Amazon’s sexist AI recruiting algorithm that prioritized hiring men over women. The algorithm learned from the company’s existing team -- not inaccurate information -- and was as flawed as the history used to train it. AI has great potential for good, but it is only as effective as the humans and data powering it. These biases may not mean much when it comes to verticals such as retail or to the ads you’re being served, but they can be a life-or-death matter in the healthcare industry.
Fortunately, as AI technology and tools are maturing, so, too, are best practices and regulatory frameworks around ethics. As GDPR is for data protection, the EU has proposed a legal framework for how to ensure AI tools are safer and more trustworthy for users, but we can’t wait until government mandated laws and best practices for AI are passed. For now, it’s on us -- the people who build these products and services -- to ensure AI-powered products and services are doing more good than harm.
Here are three priorities leaders should focus on to ensure their AI initiatives provide business value and do so ethically.
It's one thing for AI to understand the English language, but it's another to understand the nuances of language in domains such as law or clinical practice. Accurately trained models -- and ones that learn quickly -- are key to staying ahead. Achieving this is no easy undertaking. It requires ongoing monitoring, retraining, and tuning. In essence, the job is never quite done. Accuracy is also not just one feature you optimize for. There are different metrics for stability, coverage, bias, and online performance, among other factors.
Think of an AI model going into production as if it were a new car driving off the lot. As soon as it’s out in the wild, the model begins to degrade. Different environments and inputs take a toll on what once worked perfectly in a controlled research setting. To ensure models remain accurate over time, you need to dedicate the appropriate resources -- tech, talent, and software -- to keep them that way. In other words, constant monitoring and tuning is part of the job, in the same way that application and data monitoring, DevOps, and SecOps are ongoing efforts.
Responsible AI Practices
With maturity, growth, and democratizing of AI comes a responsibility to prioritize responsible practices. Think of industries such as the media -- the spreading of fake news, toxic content, and even deep fakes have become serious problems in recent years. This data -- accurate or not -- is the source feeding your AI algorithms. For example, a UC Berkeley study published in Science showed that risk prediction tools used in healthcare exhibited significant racial bias. There are countless other studies that reflect similar problems with AI and patient care.
Healthcare is leading the way for responsible AI and data practices, although it’s still an early work in progress. Bound by stringent regulations and oaths to do no harm that may not be perfect, many other industries can take a page from healthcare’s book. All companies using AI should have a system of checks and balances or ethics committees to ensure appropriate measures are in place. These steps should be implemented before AI is in use to ensure it’s built with good intent. Encourage dialogue around ethical practices and do it from the top-down. A culture of “see something, say something” will help you remain accountable.
No-Code and Low-Code Experiences
Remember when building a website was a major software engineering project? Building an e-commerce website was an eight-figure, multiyear investment in the mid-1990s. Fast forward to today and anyone can start selling in a few hours for $29/month (with a much broader feature set). AI will gradually make this change, with no-code tools getting into the hands of doctors, teachers, lawyers, marketers, and other domain experts. Leveling the playing field for AI is a crucial step for humanity but also for the sake of accuracy and ethics. To truly be representative, systems must be built by people from all demographics, geographies, and backgrounds.
Although the terms are often used interchangeably, no-code offerings are key for getting AI into the hands of the masses, and low-code offerings help with simpler coding tasks, freeing up data scientists to focus on more complex projects. This democratization of AI at all levels has become a growing area of interest, and will help move the needle for accurate and ethical AI. Beyond increasing the diversity of practitioners, no-code tools embody best practices and processes, making it easier to adopt and scale them.
The rapidly increasing evolution of AI brings with it a trove of new ethical questions and concerns. Whether an algorithm is delivering on its intended promise is one thing, and whether its downstream effects are positive is another. An algorithm may be inaccurate or unstable, but at worst it could be harmful and cost lives. Getting AI into the hands of more people through low- and no-code functionality is a step in mitigating some of these risks. By prioritizing accuracy, responsible practices, and usability, you can make your AI initiative part of the solution, not the problem.
Author: David Talby