Just as Artificial Intelligence can detect patterns that escape our eyes and reasoning, exacerbating biases that we do not perceive in advance, sometimes AI can provide results that make no sense to us, but are technically correct.
Let’s look at the following example presented by Janelle Shane, an AI researcher[78]. In one of her experiments, she decided to give an AI a disassembled robot in a virtual space and instructed it to assemble it in such a way that it could move from a starting point A to an endpoint B. As Janelle points out, the problem with AI is not that it will turn against us but that it will do exactly what we ask it to do without being clear about the limits or variables it must follow. Her AI was able to assemble the robot and fulfill the given task, but not in the way we would have imagined.
What went wrong? This is the question we ask ourselves whenever an AI provides unexpected results. Did our AI rebel, or was there a problem with our initial instructions? In movies, when something goes wrong with AI, it’s because the machines gain consciousness and decide to rebel. That might happen on TV channels like TNT, but in the real world, it doesn’t. This means that if we give a series of instructions to an AI, it will do whatever it can to achieve the desired goal, but not necessarily in the way we expect. The following example, borrowed from Janelle’s TED talk titled “The danger of AI is weirder than you think,” is key to illustrating what we mean. The disassembled robot looked like this:
Based on our human logic, the first thing we would do is assemble these parts, give it a familiar shape with legs, and then teach it to move its legs to effectively travel from A to B, as shown in the following image.
Based on AI logic, things change significantly. We don’t tell our AI how to solve the problem; we just assign the challenge or the final result it needs to achieve, which it will do through trial and error. To the surprise of many, the AI concluded that the problem could be solved by assembling itself into a tower and letting itself fall from A to B. Technically, this solves the problem we gave it.
Once again, we see that the problem is not the AI itself, nor that it turns bad against us, but that it will do exactly what we ask it to do. This means that the difficult part of working with AI is how to correctly set the challenges to get results closer to what we expect. Another curious result occurred with an AI similar to the one used in the previous experiment, but in this case by David Ha[79]. The AI was given a straight path to cross, once again starting from point A to reach a destination B, but it was told that to do so, it needed to create a system of feet and legs. The AI developed a system of giant feet and legs that could cross the entire space between A and B in one step. Did it fulfill its task? Absolutely! Did we expect it to design a set of giant feet and legs? No. We expected something smaller, to scale, comparable to the proportions of our feet and legs, not those of animals or giant limbs that were not even attached to a complete body.
If algorithms remain private, locked in a black box, we will never fully understand why our AI solved something one way and not another. We have already discussed how negative this can be socially. Now it’s time to see what happens when we open its source code to see what’s inside that black box. Let’s not repeat the same mistake as Epimetheus, who didn’t know the content of Pandora’s box, according to Greek mythology. The moral of the myth resonates with the current debate, where we must also consider the dangers of rushing to create something we don’t fully understand (Ioanna Lykiardopoulou, 2023).
A few years ago, a group of scientists put an AI model to work differentiating Siberian huskies from wolves. While the program showed significant effectiveness, there was a case where the AI identified a Siberian husky as a wolf when it was not. Since the research team didn’t know why this error occurred, they rewrote their code to make the program explain why it had made that decision. The answer? The photo of the Siberian husky showed snow in the background and around the dog. The researchers hadn’t noticed, but most of the photos the program had absorbed of wolves also showed snow. This means that in this case, the AI paid more attention to the environment than to the object it should have been trying to identify in the first place.
This reveals another problem: what happens when AI makes a mistake? What if this decision had cost a person’s life? Note that Artificial Intelligence itself will not make mistakes; at most, it will replicate errors that we inserted when programming it, feeding it our data set, or giving unclear instructions. While we’ve seen advances in recent years that have left us speechless, look at the following photo analyzed by Artificial Intelligence in 2015. It was a long time ago, I know, but look at the effectiveness it achieved!
The eye of an AI in 2015[80]
Be careful when asking to open these black boxes; there are limits. Or rather, sometimes the move can backfire on us. While we don’t want a social network to privately define the outcome of an election through its algorithm, we can’t ask them to publish their content moderation algorithms either. If these were public, there would not only be more people trying to exploit them to position their content better, but terrorist groups, spammers, and bot networks could also openly explore how to bypass these controls, and that’s not something we want to see.
[78] Shane, J. (2019). The danger of AI is weirder than you think. TED Talks. Viewed on October 12, 2021, at https://www.ted.com/talks/janelle_shane_the_danger_of_ai_is_weirder_than_you_think.
[79] Ha, D. (2018). RL for Improving Agent Design. GitHub. Viewed on June 19, 2021, at https://designrl.github.io.
[80] McNamara, T. (2015). Twitter. Viewed on July 23, 2021, at https://twitter.com/timClicks/status/619734363362557953.