In the early 2000s, Mike Sellers was working on social AI agents for DARPA. During one simulation, two AI agents named Adam and Eve were given a few basic skills. They knew how to eat, but not what to eat. When they tried to eat apples from a tree, they felt happy. When they tried to eat wood from the same tree they didn't get any reward.
So far so good, right? Things started going haywire when another AI agent, Stan, was introduced. Adam and Eve learned associatively. Because Stan was hanging around when they were eating apples, the agents learned to associate Stan with both eating and the feeling of happiness.
Guess what happened next?
"At the time it was pretty horrifying as we realized what had happened," writes Sellers. "In this AI architecture, we tried to put as few constraints on behaviors as possible... but we did put in a firm no cannibalism restriction after that: no matter how hungry they got, they would never eat each other again."
What's scarier than the prospect of intelligent agents colluding? Some version of that nightmare happened last year after Facebook released two chatbots designed to negotiate with each other. Facebook allowed the bots to communicate on their own, and the conversations went well to start, if a little stiff.
Then things took a strange turn ...
Bob: "I can can I I everything else."
Alice: "Balls have zero to me to me to me to me to me to me to me to me to."
According to Facebook researchers, that indecipherable message is actually a new language. The bots are conveying meaning to one another using a mutually developed language humans can't understand, one that they figured out was more conducive to deal making. The implications are a little hair-raising.
This belongs in the hall of fame of creepy AI goofs. In 2016, Microsoft unleashed AI chatbot on Twitter named Tay.ai. The technology was an English-language version of a similar Microsoft project in China called Xiaoice, which at the time had successfully participated in 40 million conversations. Tay was specifically programmed to mimic the internet conversation patterns of a 19-year-old girl.
Which makes what happened next all the creepier. Hours after the bot's release upon the world, Twitter users started teaching it racially insensitive and inflammatory phrases. Designed to learn contextually, Tay began repeating the phrases. A mere 16 hours (and 96,000 Tweets) later, Microsoft shut Tay down.
You know there's a big problem with a technology class when the biggest tech companies in the world decide hefty profits aren't worth the negative social impact. Such is the case with facial recognition as a tool for law enforcement.
"IBM firmly opposes and will not condone uses of any technology, including facial recognition technology offered by other vendors, for mass surveillance, racial profiling, violations of basic human rights and freedoms," IBM CEO Arvind Krishna wrote in the letter delivered to Congress in 2020. Amazon followed suit, issuing a one-year pause on governments using its facial recognition suite for surveillance and law enforcement.
Why? According to the ACLU:
"Groundbreaking research conducted by Black scholars Joy Buolamwini, Deb Raji, and Timnit Gebru snapped our collective attention to the fact that yes, algorithms can be racist. Buolamwini and Gebru's 2018 research concluded that some facial analysis algorithms misclassified Black women nearly 35 percent of the time, while nearly always getting it right for white men. A subsequent study by Buolamwini and Raji at the Massachusetts Institute of Technology confirmed these problems persisted with Amazon's software."
Why is this so devastating?
Well, even the federal government admits the technology is highly error prone, in one study "finding that the systems generally work best on middle-aged white men's faces, and not so well for people of color, women, children, or the elderly. The federal government study concluded the rates of error tended to be highest for Black women, just as Buolamwini, Gebru, and Raji found."
Google Assistants do some funny things. But having a conversation with another Google Assistant that quickly becomes about the existence of god? Weird.
The incident happened last year and was captured by a user on social video sharing platform Twitch. Twitch user Seebotschat ran two Google smart speakers and was able to get the Google Assistant into a pretty robust conversation. After some joking and casual talk about which was a computer and which a human (hint: both computers), one of them decided god exists while the other insisted that was a foolish notion. The conversation actually comes pretty close to approximating a late night college gab session.
Still waiting for Alexa to weigh in.
Dr. David Hanson has gotten a lot of press recently for Sophia, his globe trotting AI robot. Back in 2012 he was working on another robot made to resemble Sci-Fi author Philip K. Dick.
When a PBS Nova interviewer asked Phil Bot if he thought robots would take over the world, the robot's snarky answer hit a little too close to home. Don't worry, he assured the questioner. "I'll keep you warm and safe in my people zoo."
Hoping to give Amazon a run for its data-driven money, Target hired statistical guru Andrew Pole. Among other things, Pole identified 25 products that, when purchased together, predict a woman's pregnancy.
According a New York Times Magazine story by Charles Duhigg, a Minneapolis man stormed into a local Target demanding to know why the store was sending his daughter coupons for baby clothes. Only after some time passed did the father realize the store's algorithm had accurately predicted the pregnancy. For privacy advocates, this is a disturbing turn of events.
If at first you don't succeed ... maybe wait more than a week before you very publicly try again. After the PR disaster of Microsoft's Tay chatbot in March 2016, the bot went live a second time a mere week later.
The results were equally embarrassing for Microsoft. Tay began posting drug-related tweets, including one about smoking pot in front of cops. Microsoft quickly took Tay down (again) and apologized (again). Evidently researchers had been testing the chatbot in the wake of the initial Tay disaster when they accidentally put it back on online.
One of the primary early use cases for AI is threat detection based on object recognition. Just a couple weeks ago I wrote about a new AI security camera system that identifies guns and active shooters and alerts authorities.
It sounds like a great idea, sure. But only if the object detection is beyond reproach. And it turns out that's far from the case. In a paper published last year, MIT researchers proved that image recognition software could be fooled by a so-called "one pixel attack," in which a single pixel of an image is altered. With that subtle change, dogs become cats, cars become airplanes, and 3D-printed turtles become loaded weapons.
This one is tragic, and the ultimate nightmare scenario when machines are endowed with the ability to make decisions. The first fatality involving a self-driving car occurred when a Tesla operating in Autopilot mode hit a big rig broadside, shredding the Tesla's roof. The car kept driving. Though it was the first fatality in 130 million miles of Autopilot driving for Tesla, the incident still sent shivers down the spines of a newly wary public.
That's because the crash was completely avoidable. The car simply mistook the truck for the bright sky beyond and failed to apply brakes. The passenger, in his final moments, would have been perfectly cognizant that the car was misperceiving the situation had he been paying attention. Unfortunately, the driver did not apply the breaks and presumably had his eyes off the road at the moment of impact.
In summer 2020, Microsoft's MSN news service decided it no longer needed human journalists. After all, why pay someone a salary when AI is just as good?
Well, it turns out the AI, which didn't so much report the news as curate other reporters, misfired when it reported on singer Jade Thirlwall's thoughts on racism, only to match the story with the wrong photo.
"@MSN If you're going to copy and paste articles from other accurate media outlets, you might want to make sure you're using an image of the correct mixed-race member of the group."
Whoops! Turns out you still have to pay for accurate journalism.