An article from Bloomberg.com in April 2019 caused a big stir. It described a behind-the-scenes look at how Amazon.com manages its Alexa speaker AI system.30 While much of it is based on algorithms, there are also thousands of people who analyze voice clips in order to help make the results better. Often the focus is on dealing with the nuances of slang and regional dialects, which have been difficult for deep learning algorithms.
But of course, it’s natural for people to wonder: Is my smart speaker really listening to me? Are my conversations private?
Amazon.com was quick to point out that it has strict rules and requirements. But even this ginned up even more concern! According to the Bloomberg.com post, the AI reviewers would sometimes hear clips that involved potentially criminal activity, such as sexual assault. But Amazon apparently has a policy to not interfere.
As AI becomes more pervasive, we’ll have more of these kinds of stories; and for the most part, there will not be clear-cut answers. Some people may ultimately decide not to buy AI products. Yet this will probably be a small group. Hey, even with the myriad of privacy issues with Facebook, there has not been a decline in the user growth.
More likely, governments will start to wade in with AI issues. A group of congresspersons have sponsored a bill, called the Algorithmic Accountability Act, which aims to mandate that companies audit their AI systems (it would be for larger companies, with revenues over $50 million and more than 1 million users).31 The law, if enacted, would be enforced by the Federal Trade Commission.
There are also legislative moves from states and cities. In 2019, New York City passed its own law to require more transparency with AI.32 There are also efforts in Washington state, Illinois, and Massachusetts.
With all this activity, some companies are getting proactive, such as by adopting their own ethics boards. Just look at Microsoft. The company’s ethics board, called Aether (AI and Ethics in Engineering and Research), decided to not allow the use of its facial recognition system for traffic stops in California.33
In the meantime, we may see AI activism as well, in which people organize to protest the use of certain applications. Again, Amazon.com has been the target of this, with its Rekognition software that uses facial recognition to help law enforcement identify suspects. The ACLU has raised concerns of accuracy of the system, especially regarding women and minorities. In one of its experiments, it found that Rekognition identified 28 members of the Congress as having prior criminal records!34 As for Amazon.com, it has disputed the claims.
Rekognition is only one among various AI applications in law enforcement that are leading to controversy. Perhaps the most notable example is COMPAS (Correctional Offender Management Profiling for Alternative Sanctions), which uses analytics to gauge the probability of someone who may commit a crime. The system is often used for sentencing. But the big issue is: Might this violate a person’s constitutional right to due process since there is the real risk that the AI will be incorrect or discriminatory? Actually, for now, there are few good answers. But given the importance AI algorithms will play in our justice system, it seems like a good bet that the Supreme Court will be making new law.
AGI (Artificial General Intelligence)
In Chapter 1, we learned about the difference between strong and weak AI. And for the most part, we are in the weak AI phase, in which the technology is used for narrow categories.
As for strong AI, it’s about the ultimate: the ability for a machine to rival a human. This is also known as Artificial General Intelligence or AGI. Achieving this is likely many years away, perhaps something we may not see until the next century or ever.
But of course, there are some brilliant researchers who believe that AGI will come soon. One is Ray Kurzweil, who is an inventor, futurist, bestselling author, and director of Engineering at Google. When it comes to AI, he has left his imprint on the industry, such as with innovations in areas like text-to-speech systems.
Kurzweil believes that AGI will happen—in which the Turing Test will be cracked—in 2019, and then by 2045, there will be the Singularity. This is where we’ll have a world of hybrid people: part human, part machine.
Kind of crazy? Perhaps so. But Kurzweil does have many high-profile followers.
But there is much heavy lifting to be done to get to AGI. Even with the great strides with deep learning, it still generally requires large amounts of data and significant computing power.
AGI will instead need new approaches, such as the ability to use unsupervised learning. Transfer learning will likely be critical as well. For example, as we’ve covered earlier in the book, AI has been able to realize superhuman capabilities in playing games like Go. But transfer learning would mean that this system would be able to leverage this knowledge to play other games or to learn other fields.
In addition, AGI will need to have the capacity for common sense, abstraction, curiosity, and finding causal relationships, not just correlations. Such abilities have proven extremely difficult with computers. If anything, there will need to be breakthroughs in hardware and chip technologies. This is the opinion of Yann LeCun, one of the world’s top AI researchers and the chief artificial intelligence scientist at Facebook.35 He also thinks there needs to be much more progress with batteries and other energy sources.
Something else that will be critical: more diversity within the AI field. According to a report from the AI Now Institute, about 80% of AI professors are men; and among the AI research staffs at Facebook and Google, women accounted for 15% and 10%, respectively.36
This lopsidedness means that research could be more susceptible to bias. Furthermore, there will be the loss of the benefit of broader views and insights.
Leave a Reply