Artificial intelligence holds tremendous promise, has been a major disappointment or is a threat that will derail civilization as we know it.

All of those views, in fact, were apparent during discussions this week at EmTech Digital 2018, the MIT Technology Review conference held this week in San Francisco. Which characterization is right depends a great deal on the use case involved and the strongly held opinions of those in the technology industry.

At its core, however, AI is a lot like human beings: imperfect and still learning. “Artificial intelligence and human intelligence have always been inextricably entwined,” Gideon Lichfield, Technology Review’s editor in chief, said during his opening remarks to the gathering Monday.

AI seems to be everywhere these days, from smart toothbrushes to attempts to thwart illegal deforestation. Yet major progress is still slow, thanks in part to the laborious work needed to build models that rely on human-generated categories, labeled data and algorithms to allow a machine to make sense of it all.

Understanding common sense

The challenge continues to be finding ways to give systems key traits that humans have innately learned, such as context and common sense. A doctor, when presented with data on a five-year-old whose weight is listed at 500 pounds, will instantly recognize the error and correct it to 50 pounds. A machine will likely go ahead and prescribe a dosage with potentially disastrous results.

“If you want to understand the limitations of AI, just talk to Alexa or talk to Siri,” said Oren Etzioni, chief executive of the Allen Institute for Artificial Intelligence. “My seven-year-old is more autonomous than an AI system.”

AI’s limitations are not confined simply to intelligence power. Some researchers are expressing concern about bias through the use of training data for AI models that may inherently be weighted to favor people of a specific gender or race.

Courts and parole boards are using criminal risk scores, generated using AI models, to determine actions such as granting bail or early release. Researchers from ProPublica found that formulas used in risk scoring appeared to ensure that black defendants would be erroneously tagged as future criminals at higher rates than white offenders.

“These algorithms are being used in very high-stakes scenarios,” said Timnit Gebru, a research scientist at Microsoft Corp., who cited the ProPublica study as a warning sign. “AI has a lot of opportunities, but we should have safety measures in place.”

Struggling to grasp objects

Aside from concerns over intelligence and bias, another broad field for AI is still struggling to make a leap forward: training robots in basic tasks for warehouse fulfillment. The difficulty stems from challenges in perception and control. A simple task such as pushing a bottle across a particular surface can confound many robots because of unpredictable friction generated by the action and the machines’ limited ability to understand the concept.

“It’s not surprising that today’s robots are remarkably clumsy,” said Ken Goldberg, a professor at the University of California at Berkeley who has been working in the robotics field for over 35 years. “I have to confess, we’ve made remarkably little progress.”

The central issue continues to be grasping, the ability of robots to lift objects ranging from the simple to the complex. Examples of the latter which have confounded robots in Goldberg’s work include a paper clip and a squeezable bottle of chocolate syrup.

The industry measurement for robotic progress in this area is based on mean picks per hour or MPPH. Humans can achieve 400 to 600 MPPH, while the best robotic rates so far in Goldberg’s research have topped out at 200. Suction cups help, yet mimicking the human hand remains a difficult proposition. But the professor remains optimistic. “I do think we can get to the point in the next five years where we can achieve the same picks per hour as humans,” Goldberg said.

China blazing new trails

Despite the slow advancement and bias concerns surrounding AI, there are also promising signs that the technology is improving. One region of the world where such progress can be found is China. Several Chinese researchers at EmTech Monday talked about the integration of AI into many facets of daily life, particularly in mobile media content.

SenseTime is one of the world’s fastest-growing AI companies, showing a revenue increase of 420 percent in fiscal year 2017, according to Dahua Lin, assistant professor at the Chinese University of Hong Kong and director of a joint research lab with SenseTime. “[China] sees this as a great opportunity for the country to catch up with the West,” Lin said.

One of Lin’s demonstrations included scenes from the 1997 film “Titanic.” The firm’s AI engine capably recognized and extracted footage on demand that showed either romantic interludes or disaster segments from the movie.

Another Chinese company, Bytedance, is building its AI capability based on content creation for the mobile platform. The company has 700 million monthly active users and processes 20 million new videos per day, according to Wei-Ying Ma, vice president and head of AI Labs at Bytedance.

“We want to equip the user with AI technology directly on their phone,” Ma explained. “We are like a YouTube for everything on a mobile device.”

Ma demonstrated a phone app that seamlessly changed a user’s hair color in live action video and another that rewarded participants for dance moves, based on a body motion tracking algorithm built into the mobile platform. Bytedance also runs a digital service, called Toutiao, powered by machine and deep learning algorithms to write and distribute news content.

If progress in moving AI forward is uneven, the field still continues to capture enterprise interest. Adobe Systems Inc. released data this month showing that just 15 percent of enterprises are currently using AI, but 31 percent have it on the agenda for the next year.

“The key question we face today is where we are headed in the next several years to come,” Lin said. It’s a question we always ask ourselves as humans. Now we just have to see if machines can give us the answer.

Photo credit: A Health Blog via photopin