The Danger of AI
This post will not discuss the danger of speculations of what AI can do in the future, but concentrating on what type of problems the current state of AI can bring to our daily life and the modern society. The issues can be categorized into two kinds: the issues within AI itself and the issues with how AI will be used.
Bias is one prominent example of the problems within the AI system. In her thesis, Joy Buololamwini, a researcher from MIT Media Lab, has found multiple concrete pieces of evidence on the gender and racial biases of the service provided by large tech companies like Amazon and Microsoft.
Biases in AI will systemize the discrimination, which will spawn numerous social-economical-cultural-political problem. This can be seen from many instances of AI application. Faception (an Isareal based facial recognition app) was marketed on their website as capable of determining one person’s personalities through their face, including terrorist “personality.” Similar problems can be found with the research “Automated Inference on Criminality using Face Images” by Xiaolin Wu and Xi Zhang; or the criminal predicting model from Harrisburg University.
There have been promising efforts in mitigating the biases in AI by improving diversity in the technological workforce, or diversifying the training data or developing Awareness and Debiasing Tools. An example is IBM’s AI Fairness open-source, extensible library, which can detect biases based on the labels (race, sexual orientation) by running a range of metrics, like disparate impact and equal opportunity difference, to quantify the biases. AI Fairness Library also contains more than ten different debiasing approaches. However, AI Fairness was limited to binary classification only. For other types of machine learning, tools like Aequitas and LIME (The Local Interpretable Model-agnostic Explanations toolkits) can be used to measure feature importance and explain the local behavior of popular models like multiclass classification, regression, and deep learning. LIME allowed human users to understand the decision-making process of the machine, thus allowed more chances for them to spot biases. However, this feature is limited to spotting the biases only, not fixing it, similar to the limitation in Aequitas.
Let’s assume a perfect world where AI behaves exactly like how their creators expected them to behave. The danger of AI still lies in how they can be used. With a robust surveillance system powered by AI and controlled by the government, I would argue that our society will become a dystopia instead of a utopia. We can look at the case in China, Uighurs, where a largely Muslim minority was specifically targeted, being kept track of their comings and goings by the Chinese government using facial recognition tools (New York Times, 2019). Such a political-social disaster could become common in said world if we do not engage with AI development with utmost moral caution.
Let’s bravely assume that we will keep such morally questionable implementation of AI in check. Even then, can we say we are safe and can enjoy the great benefit of efficiency that AI will bring us? I would again argue that problems can come from the success of AI’s efficiency and creativity.
An example of such problems could be how AI can reinforce and strengthen fast fashion dominance in the fashion industry. Fast fashion dominated the fashion industry, with weekly renewed, affordable and fashionable clothing, made possible by a fast supply chain with the cost of the environment, intensive raw material usage, and developing countries labor extortion. Fast fashion brand clothing items are not built to last, encouraging the dangerous culture of disposable clothing, resulting in serious environmental problems like an increase in non-recycled clothing waste and greenhouse gas emissions. One of the key components to the fast fashion business model is quick and new design every week, unlike the traditional fashion seasons released twice a year, to generate new trends or catch up with the current trend to keep customers interested. With the help of AI’s creativity, the design process for fast fashion brands can be further speed up, as shown in the example of my group project called “this dress may exist” in 2020.
In this example, we collected about 1000 images from large fast fashion brands like H&M and Topshop to use as training data and trained the model with Runway ML. The result is more than 500 generated new designs showed on the website. With a tool similar to this, the design process inside these fast-fashion brands can be much more efficient, with less human designer’s involvement. Such reinforcement in the fast fashion industry will increase its negative social and environmental impacts. Another concerning issue is the source of the training data. The model may take designs from small local brands and designers, using their intellectual properties to create new designs that will, in turn, compete with their originals, but on much a wider scale thanks to the big brands’ network of distributions and marketing powers.
The negative impacts of the efficiency in artificial creativity and copyright of intellectual properties are certainly not unique to the fashion industry. AI may bring huge changes to our society with its potential, and we need to be ready for it, be it a legal framework by the government or the understanding mentality of the customers. At the end of the day, even with a perfect working AI system, the danger will always lie in the human using it.
Other AI-related projects:
Rock Paper Scissor Game: Image recognition project using Teachable Machine, make sure the machine always beat you:
https://rock-paper-scissor-machine.glitch.me
Emotion Lightning Round: Facial Expressions and Gestures Recognition project using Teachable Machine, having fun with posing:
