Artificial intelligence excels at performing specific tasks, but it has trouble keeping up with the latest dance trends, for instance, without some human intervention.
That was one of the lessons that Jay Swartz, now chief scientist at the consultancy Blackbox AI, learned building a series of bots designed to answer questions about health insurance coverage.
"I didn't teach it about twerking, OK, which is a term in the English language. It's a dance move," Swartz told an audience at the Global Artificial Intelligence Conference in Boston on Tuesday. "Somebody literally asked it a question about: 'I hurt my hip twerking.'"
The power of AI to sort massive amounts of data can speed up decision-making in business and medicine, but the blind spots in the technology mean the guiding hand of a person should be close by, according to Swartz.
"No AI learns on its own. None," Swartz said after his talk. He said, "It is not that smart and it needs to be taught again and again, and tuned and watched. So you can't really implement a self-running AI because it will go off the rails too fast -- today."
Swartz's caution, which was echoed by others, could serve as a slight check on fears promulgated by AI's ascension, while also exposing the limitations of AI in applications where the technology might be expected to be fully in control soon, as in autonomous vehicles.
The power of AI is its capacity to take on specific chores, not take over whole jobs, according to Swartz and Robbie Allen, chief executive at Infinia ML, which helps businesses develop machine learning solutions.
Jay Swartzchief scientist, Blackbox AI
Allen's prior company, Automated Insights, developed Wordsmith, a program that tasked computers with writing up earnings reports for The Associated Press and summarizing game day stats for Yahoo fantasy football players. Machine learning supplanted relatively straightforward reporting and writing for financial journalists, but it didn't threaten their jobs, according to Allen, who also gave a talk at the Global AI conference.
"When I did interviews specifically with financial journalists, they were very thankful," Allen said, because by handling the rote work, the AI allowed those journalists to focus on projects that better used their talents. He said, "We're not aware of anybody being let go because now we have Wordsmith."
Cesar Argueta, the vice president of client success and product innovation at Wylei Inc. who attended Allen's talk, agreed that executives should be aware of the limitations of AI.
"Where AI is today, it has to be very structured. It has to be very applied. And it's like anything. Unless it's programmed, unless it's created or designed to do something, it won't do it," Argueta said after the talk. "It just can't do everything."
Power of AI to eliminate jobs
While the AI-assisted Wordsmith program developed by Robbie Allen may not be smart enough to replace journalists, the power of AI to eliminate humans on the job was evident just a stone's throw from the Global AI conference.
AI replaced jobs along the Massachusetts Turnpike. The state's transportation department eliminated all the toll taker jobs in 2016 when it implemented automatic tolling with gantries along the highway that replaced the old toll booths.
Many of the 470 former toll takers were offered new positions in state government, while others opted to retire from careers as the gatekeepers of the interstate.
The Turnpike runs below-grade just across the street from the Boston Convention and Exhibition Center.
How to parse the ROI of AI
When determining whether to implement an AI application, executives should consider whether it is worth the substantial cost, Swartz said.
"Machine learning can do better than humans if you bound it nicely enough or tightly enough," Swartz said. He said, "It doesn't generalize as well as many people think it should."
According to Swartz, factors that indicate AI might be worthwhile for a process are:
- High volume of work
- Limited judgment required
- Low cost of mistakes
- Repetitive work
- Work that follows a pattern
The power of AI and the limitations of AI are displayed in its use in radiology: AI excels at reading scans to detect the presence of breast cancer, but it is ill-suited to replace all of the insights, knowledge and information-gathering required of a radiologist, Swartz said.
When Swartz built the health insurance chatbot, which is branded as Welltok and uses IBM's Watson, he actually built several different bots to handle the various tasks with an interface on top. The user experienced it as a singular program rather than what it was: a network of different bots that each had limited knowledge in a specific area of health.
"What they didn't realize is they were actually sweeping in a whole collection of Watsons when they were asking questions," Swartz said.
New job title: Machine learning assistant
Although Welltok required a network of bots, AI in this case was well-suited to handle questions about health insurance because of the volume of factors that go into determining whether something is covered.
While the human brain excels over computers at thinking about the big picture, it generally makes decisions based on only about five data points, according to Swartz.
"We make it at high speed. We make it two or three times a second. We readjust. We move according to what we see in people's eye. It's a very complex phenomenon. Healthcare has a penchant for lists of 27 different things. And you can't hold that in your head, so you start getting lost," Swartz said. Of health insurance, he said, "That's why none of us can navigate it. It's contrary to human capability. So we built a bot that solves that."
The task performed by Transportation Security Administration employees of reading X-ray scans to detect weapons in travelers' luggage is ripe for replacement by AI, according to Allen, who said that humans should stay involved in that security process even if the task of identifying guns and knives is automated.
The development of AI to fit new niches will also create new work for people in the form of "machine-learning assistants" and those who prepare databases for AI, according to Allen.
"Everyone thinks their data's ready, and it's almost never ready," Allen said.
One hurdle to implementing AI is that many managers feel threatened by computers taking over their responsibilities, Swartz said. He theorized that is why AI is usually implemented by startups, where it is integral to their business model, or by giant corporations that are equipped to handle that type of change.
The best way for businesses to keep up with the latest technology is by using a module-based system throughout the enterprise so they can "roll out the old stuff, and slide in the new stuff" without disturbing the other parts, according to Swartz.