Artificial Intelligence: The Downside
Enriching computers with human-like intelligence has been a fantasy of computer experts since the beginning of electronic computing. Since there have been made huge advancements in Artificial Intelligence, the technology tends to outperform humans in very defined tasks, such as playing strategic games, language translation, self-driving vehicles and image recognition.
ImageNet Roulette
Consequently, everyone is talking positively about Artificial Intellegence these days. In contrast, the application ImageNet Roulette shows the drawbacks of this phenomenon. This viral app examines photos and uses AI to allocate personality traits to people. The application’s designers have trained the algorithm on almost three thousand subcategories that can be found on ImageNet. These personality traits are mostly described with relatively harmless terminology. But many assigned descriptions, for instance, “hypocrite”, “loser”, “drug addict” and “racist”, are deeply disturbing. As ImageNet is free to use for everyone, it is one of the most important and complete datasets within Artificial Intelligence. Therefore, it is very important to cope with the technology appropriately.
Computers can’t act like human
While using ImageNet Roulette it is a fun activity for most people, the application also shows the limitations of using Artificial Intelligence. When a company is using machine learning to build a neural network, the network with different layers is modeled on how we might think the brain works. As we give input to the system, we enable the machine to learn, for example to identify images. But since, the machine is relying on the input we gave it, the neural network is not as smart as it looks. The machine is just learning how to better perform a certain task, depending on our input.
Machine learning, and thus Artificial Intelligence, can be used for lots of different task, from virtual personal assistants to the filtering of spam in your mail. But, neural networks don’t truly understand anything. Therefore, we should be very careful in the application of this fast-moving technology.
Sources:
Hamilton, I.A. (2019) Deze selfie-tool die viral gaat, zegt wat voor persoon je bent – en geeft opmerkelijke resultaten. Available at: https://www.businessinsider.nl/imagenet-roulette-selfie/
Schwab, K. (2019) Who does AI think you are? This groundbreaking new exhibit will show you. Available at: https://www.fastcompany.com/90400613/who-does-ai-think-you-are-this-groundbreaking-new-exhibit-will-show-you
Hoffman, C. (2019) The Problem With AI: Machines Are Learning Things, But Can’t Understand Them. Available at: https://www.howtogeek.com/394546/the-problem-with-ai-machines-are-learning-things-but-cant-understand-them/
Oberoi A. (2017) 9 Machine Learning Examples from Day-to-Day Life. Available at: https://insights.daffodilsw.com/blog/9-machine-learning-examples-from-day-to-day-life
Wong, J.C. (2019) The viral selfie app ImageNet Roulette seemed fun – until it called me a racist slur. Available at: https://www.theguardian.com/technology/2019/sep/17/imagenet-roulette-asian-racist-slur-selfie
Hi Thije,
Thanks for this interesting read! The application of AI on pictures has been debated for quite some time. Googles AI was fooled several years ago by recognizing a turtle for an automatic machine gun, and mistaking a picture of a cat for guacamole. This was done by altering small pixels in the picture, which are unidentifiable by the human eye. I completely agree with you on the fact that we should be careful upon trusting AI, as its black box way of making decisions is impossible to understand. On the other hand, could it be that ImageNet Roulette has found a way to artificially express inaccurate prejudices which we desperately oppress?