Mary Shacklett
"The issue of mistrust in AI systems was a major theme at IBM's annual customer and developer conference this year," said Ron Poznansky, who works in IBM design productivity. "To put it bluntly, most people don't trust AI—at least, not enough to put it into production. A 2018 study conducted by The Economist found that 94% of business executives believe that adopting AI is important to solving strategic challenges; however, the MIT Sloan Management Review found in 2018 that only 18% of organizations are true AI 'pioneers,' having extensively adopted AI into their offerings and processes. This gap illustrates a very real usability problem that we have in the AI community: People want our technology, but it isn't working for them in its current state."
Poznansky feels that lack of trust is a major issue.
"There are some very good reasons why people don't trust AI tools just yet," he said. "For starters, there's the hot-button issue of bias. Recent high-profile incidents have justifiably garnered significant media attention, helping to give the concept of machine learning bias a household name. Organizations are justifiably hesitant to implement systems that might end up producing racist, sexist or otherwise biased outputs down the line."
On the other hand, Poznansky and others remind companies that AI is biased by design—and that as long as companies understand the nature of the bias, they can comfortably use AI.
As an example, when a major AI molecular experiment in identifying solutions for COVID was conducted in Europe, research that deliberately did not discuss the molecule in question was excluded in order to speed time to results.
That said, analytics drift that can occur when your AI moves away from the original business use case it was intended to address or when underlying AI technologies such as machine learning "learn" from data patterns and form inaccurate conclusions.
Find a midpoint
To avoid skewed results from AI, the gold standard methodology today is to check and recheck the results of AI to confirm that it is within 95% accuracy of what a team of human subject matter experts would conclude. In other cases, companies might conclude that 70% accuracy is enough for an AI model to at least start producing recommendations that humans can take under advisement.
Arriving at a suitable compromise on the degree of accuracy that AI delivers, while understanding where its intentional and blind bias spots are likely to be, are midpoint solutions that organizations can apply when working with AI.
Finding a midpoint that balances accuracy against bias allows companies to do three things:
They can immediately start using their AI in the business, with the caveat that humans will review and then either accept or reject AI conclusions.
They can continue to enhance the accuracy of the AI in the same way that they enhance other business software with new functions and features.
They can encourage a healthy collaboration between data science, IT and end-business users.
"Solving this urgent problem of lack of trust in AI … starts by addressing the sources of mistrust," Poznansky said. "To tackle the issue of bias, datasets [should be] designed to expand training data to eliminate blind spots."
No comments:
Post a Comment