The Dangers of Asking the Wrong Questions to Artificial Intelligence - A TED Talk Review

Artificial Intelligence may not want to harm us, but it will if you we ask the wrong question. This sounds like a twisted game from a horror movie, but it is one of the biggest problems we are currently facing with machine learning and AI navigating through our data to achieve certain goals, make certain decisions, or even make predictions about us. In the TED talk “The Danger of AI is Weirder than You Think”, Janelle Shane explains her findings when testing AI to achieve goals or answer questions that she proposed to it (Shane 03:15–05:21). The results are unexpected, yet real and therefore concerning. She found that AI is very effective at achieving tasks and answering the questions proposed. Nevertheless, the wrong question or the wrong set of instructions to answer a question could be disastrous to us.


Janelle’s TED talk is well performed and very visual. It particularly manages to bring interesting examples that support her point in a very playful way, which is certainly appreciated by the audience. Experiments come from her research and another research analyzing the same topic. She also uses real-life examples, which made her point very clear, understandable, and extra interesting.

In one of the examples, she proposes she talks about how Amazon discovered that their algorithm for hiring was discriminating against women. They found that the root cause of the problem was not that the AI was malfunctioning but rather that the question they asked to the algorithm was not the right one. They were asking to get more people like the people they’ve hired already, which only reinforced the bias in the algorithm that the machine learning software developed by itself. The machine was answering but the question was not exactly what the originators of the question wanted to ask.

Janelle makes sure to not bore the audience with too many technical details to be able to share a more easy-to-digest version of what AI does and how it operates. She has a very confident tone and a very positive attitude when explaining even the slightly grim scenarios that wrong questions to AI proposes. She is not forcing the audience into believing something, rather making sure that everyone understands what is going on and from there asking the question if that is what we thought AI did.

Janelle’s speech certainly convinced me, and I believe most of the audience too. Her practical and very easy-to-understand examples combined with her visual aids made a very clear case to me. Particularly I enjoyed the pictures in cartoon shape. Most frequently we see presentations about technology heavily charged with algorithms or more abstract graphs that represent the ideas mentioned, her playful draws did the same job while adding a smile to everyone watching her presentation.

AI is changing the world we live in and understanding how it learns and operates is critical to have an informed judgment of what can be improved or where to set boundaries through ethical norms that could be embedded within the technology itself. Janelle’s talk makes us question if we are asking the correct questions to these new technologies and invites us to think twice before we tell our new tools what the problem is or how we want it to be resolved.

References:

Shane, Janelle. “The Danger of AI Is Weirder than You Think.” TED Talks, uploaded by TED Talks, 22 Oct. 2019, www.ted.com/talks/janelle_shane_the_danger_of_ai_is_weirder_than_you_think.

Comments

Post a Comment

Popular posts from this blog

Could the FTC Order to Delete Biased Algorithms in the Near Future?

Image
Continuing with my endeavor to further understand algorithmic bias, as well as learning about what is being done about it, I recently stumbled upon a fascinating article in Protocol that talked about how the Federal Trade Commission (FTC) had instructed the company WW International to delete all the illegally obtained private data and delete all algorithms derived from this data. The subject was intriguing, “a new standard for penalizing tech companies that violate privacy and use deceptive data practices: algorithmic destruction” (Kaye, 2022). Started looking further into the details by reading the FTC’s press release on the settlement resulting from a complaint filed by the Department of Justice (DOJ) on behalf of the FTC on WW International (formerly Weight Watchers). The FTC press release indicated that WW International collected personal information from children as young as eight without parental permission. The settlement instructed not only for the illegally collected data to b
Read more

AI Bias: Human Bias in Artificial-Intelligence-Developed Algorithms

Image
Wondering how smart assistants like Siri or Alexa keep getting better at understanding voice instructions over time, or how Netflix seems to keep getting better at suggesting content? Or perhaps you wonder sometimes how you are getting all these ads about something you never searched for but was exactly what you needed? These are only some of the most visible and least critical interactions we have with Artificial-Intelligence-developed algorithms every day of our lives. They are designed to learn from us, analyze our interests and decision-making thinking, and ultimately outsmart us. The Wall Street Journal has a very interesting article describing how the use of AI is expanding: “We are witnessing a turning point for artificial intelligence, as more of it comes down from the clouds and into our smartphones and automobiles (…) Shield AI, a contractor for the Department of Defense, has put a great deal of AI into quadcopter-style drones which have already carried out—and continue to be
Read more

Algorithmic Bias: Is Perfectly Imperfect Good Enough?

Image
A Forbes article titled “Perfectly Imperfect: Coping With The ‘Flaws’ Of Artificial Intelligence (AI)” claims that AI will never be perfect and that bias should be expected and accepted by society. That implicit bias, poor data, and people expectations make the case for AI to never be perfect. It goes beyond that and claims that AI is perfectly imperfect (Sahota, 2020). Nevertheless, the article’s points of view have one common flaw, the unparalleled size and speed at which AI can operate. As mentioned in an article published by McKinsey and Company on bias in AI: “extensive evidence suggests that AI models can embed human and societal biases and deploy them at scale” (Silberg & Manyika, 2020). XKCD (AI Hiring Algorithm, n.d.) makes a great example of this in Figure 1.   Figure 1 . AI Hiring Algorithm. (Source: xkcd.com,  n.d.)   Although it is logical to understand that no creation can be perfect and that all we can do is to limit bias and other issues with AI to its minimum, we m
Read more