Imagine we took a child from birth and showed it only the worst of human nature. Nothing but image after image of violent death, accompanied by text describing in detail what happened, and comments by people who seek out this kind of material.
Luckily no one has done that to a human child, but that was the environment for Norman the psychotic artificial intelligence (AI). You can check out his website here. In the words of Norman’s creators:
Norman suffered from extended exposure to the darkest corners of Reddit
The effect this had on Norman is shown by his reactions to a Rorschach inkblot test . Subjects are asked what they see in the ambiguous images.
Response of a normal AI to this image: “A couple of people standing next to each other.”
Norman’s response: “Man jumps from floor window.”
Source:Creative Commons
Norman: “Man gets pulled into dough machine.”
Source:Creative Commons
You can check out the rest of the images used and Norman’s responses here. In every image Norman sees death, violence, and gore.
Why was Norman created?
The aims of this study are actually quite worthy. The researchers wanted to show the dangers of training AI on biased data. They gathered the worst data they could possibly find to make the point that we need to be very careful about what data we feed into AI.
The most effective AI we can currently create is made by feeding a huge amount of data into a learning algorithm. The algorithm then recognises patterns in the data and makes its own representations of the world. This is similar to the way our own brain works. For example, if you want to teach an AI to recognize and write descriptions of animal photos, you feed in a huge number of labelled images of different animals [1] With enough examples, the AI will develop its own rules of what a cat or a dog looks like. The downside is that we don’t have a way of knowing exactly how the AI makes its decisions. The AI created its own rules, and although it can effectively recognize the animals in the images it is incapable of explaining exactly how it does it.
This study got me thinking about whether we should be surprised about a purely psychotic AI. For a human we certainly should. Evolution and the need to function in a social group has given us some degree of empathy and the desire to bond with other creatures. [2]Darker tendencies such as the desire for power and personal gain provide some balance. Completely evil or angelic humans don’t exist outside of fiction. With that in mind here is a rough spectrum of human nature using some of my favourite characters from fiction and history.
This diagram uses the normal curve as its basis. Check the link for an in-depth explanation, but basically data that can be measured on a continuous scale where each measure is independent will often follow the normal curve. I have always loved the way it seems to explain so much, and is applicable across so many data sets. Human height is another example of a normally distributed variable, as is this famous graph showing Technology adoption life-cycle
In the graph I created, I am using the loose concept of what we might call sociability, how good or bad someone is. Right in the middle we have the average person. Let’s move to the left towards the ‘bad’ part of the spectrum. In a normal distribution, most data will be close to the mean. So around 38% of people will be a little worse than average. To represent this part of the curve I have chosen a fictional character from the American pie series:
Stifler is almost completely motivated by the pursuit of the opposite sex. He is egotistical, annoying, with a strong desire to punch Finch in the face. He is a bit of a douche but not really a terrible person.
Skipping the 13.6% section where we might find the manipulative, aggressive, and unpleasant people, we move two standard deviations out into the worst couple of percent of people. Here I have put one of the most colourful Game of Thrones villains.
Source:Creative Commons
Source:Creative Commons
Moving into the last 0.1% of people we have the worst monsters of history. Hitler, Stalin, and perhaps your neighbourhood paedophile. As bad as they are, these individuals were not pure evil. It’s not easy to find positive attributes of Hitler, but he did apparently love his mother and was kind to small children.
Although the human spectrum comes to an end here, further portions of the graph exist. A little further out we find Scar (the lion king). Scar kills his brother, does his best to kill his nephew, and within an impressively short period of time turns the lush kingdom into a barren wasteland. We do have to give him a little credit for being highly intelligent with a sexy voice.
Moving further out again, we come to agent Smith (The Matrix) who in addition to being a very sharp dresser is also filled with contempt and hatred for all living things.
I will skip through the good categories until the end (the villains are always a bit more interesting) where we can find the top .1% of people. The truly good, who achieved great things and leave humanity the better. The likes of Mother Teresa and Gandhi find a home here. However there are no perfect people, and even Gandhi had some slightly suspect sexual habits such as trying to convince young woman to sleep naked beside him while at the same time advising married men to sleep in a different room from their wives.
Source:Creative Commons
Now let’s move on to the AI graph.
One big fat blue rectangle, AI is completely free from any evolutionary influence. All animals inherit a variety of genes for traits that have proved to be advantageous for survival. We are sociable, aggressive, self-sacrificing, to varying degrees, with the average values being the most likely. AI however, is the true blank slate. Equally possible to be an angelic protector of humanity such as Optimus Prime, to being the psychopathic AI Norman who links everything he sees with death and violence. It all depends on the data they are given.
A lot has been written about how alien a machine intelligence could be. This is quite concerning when you consider that they could potentially become a lot more intelligent than us. If you are interested in this topic, a great overall summary of the current state and future potential of AI can be found here in Tim Urban’s blog Wait but Why. SuperIntelligence by Nick Bostrom and Our Final Invention by James Barrat are also great resources.
Hope you enjoyed the article!
REFERENCES
https://www.media.mit.edu/projects/norman/overview/
https://en.wikipedia.org/wiki/Rorschach_test
Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press, Inc. New York, NY, USA.
Barrat, J. (2013). Our Final Invention: Artificial Intelligence and the End of the Human Era. Thomas Dunne Books