Last week, we discovered that there is a new side to AI. And I don’t mean to gloat, but I saw this potential pitfall as fairly obvious. It is interesting that the real world event that triggered all the talk occurred within days of episode 159 of The Lightstream Chronicles. In my story, Keiji-T, a synthetic police investigator virtually indistinguishable from a human, questions the conclusions of an Artificial Intelligence engine called HAPP-E. The High Accuracy Perpetrator Profiling Engine is designed to assimilate all of the minutiae surrounding a criminal act and spit out a description of the perpetrator. In today’s society, profiling is a human endeavor and is especially useful in identifying difficult-to-catch offenders. Though the procedure is relatively new in the 21st century and goes by many different names, the American Psychological Association says,
“…these tactics share a common goal: to help investigators examine evidence from crime scenes and victim and witness reports to develop an offender description. The description can include psychological variables such as personality traits, psychopathologies and behavior patterns, as well as demographic variables such as age, race or geographic location. Investigators might use profiling to narrow down a field of suspects or figure out how to interrogate a suspect already in custody.”
This type of data is perfect for feeding into an AI, which uses neural networks and predictive algorithms to draw conclusions and recommend decisions. Of course, AI can do it in seconds whereas an FBI unit may take days, months, or even years. The way AI works, as I have reported many times before, is based on tremendous amounts of data. “With the advent of big data, the information going in only amplifies the veracity of the recommendations coming out.” In this way, machines can learn which is the whole idea behind autonomous vehicles making split-second decisions about what to do next based on billions of possibilities and only one right answer.
In my sci-fi episode mentioned above, Detective Guren describes a perpetrator produced by the AI known as HAPP-E . Keiji-T, forever the devil’s advocate, counters with this comment, “Data is just data. Someone who knows how a probability engine works could have adopted the characteristics necessary to produce this deduction.” In other words, if you know what the engine is trying to do, theoretically you could ‘teach’ the AI using false data to produce a false deduction.
I published Episode 159 on March 18, 2016. Then an interesting thing happened in the tech world. A few days later Microsoft launched an AI chatbot called Tay (a millennial nickname for Taylor) designed to have conversations with — millennials. The idea was that Tay would become as successful as their Chinese version named XiaoIce, which has been around for four years and engages millions of young Chinese in discussions of millennial angst with a chatbot. Tay used three platforms: Twitter, Kik and GroupMe.
Then something went wrong. In less than 24 hours, Tay went from tweeting that “humans are super cool” to full-blown Nazi. Soon after Tay launched, the super-sketchy enclaves of 4chan and 8chan decided to get malicious and manipulate the Tay engine feeding it racist and sexist invective. If you feed an AI enough garbage, it will ‘learn’ that garbage is the norm and begin to repeat it. Before Tay’s first day was over, Microsoft took it down, removed the offensive tweets and apologized.
Apparently, Microsoft, though it had considered that such a thing was possible, but decided not to use filters (conversations to avoid or canned answers to volatile subjects). Experts in the chatbot field were quick to criticize: “‘You absolutely do NOT let an algorithm mindlessly devour a whole bunch of data that you haven’t vetted even a little bit.’ In other words, Microsoft should have known better than to let Tay loose on the raw uncensored torrent of what Twitter could direct her way.”
The tech site, Arstechnica also probed the question of “…why Tay turned nasty when XiaoIce didn’t?” The assessment thus far is that China’s highly restrictive measures keep social media “ideologically appropriate”, and under control. The censors will close your account for bad behavior.
So, what did we learn from this? AI, at least as it exists today, has no understanding. It has no morals and or ethical behavior unless you give it some. Then next questions are: Who decides what is moral and ethical? Will it be the people (we saw what happened with that) or some other financial or political power? Maybe the problem is with the premise itself. What do you think?