Blog 4: How GenAI Works

Case Study:
How Generative AI Works and How It Fails

Summary

The case study on how generative AI works aims to provide an introductory explanation into the behind-the-scenes of commercial AI chatbots. It discusses the harms of AI and some of the tasks it has gotten good at. Ultimately, it ends up discussing ethical issues related to the use of AI and some changes that are needed in order for AI to be safe and non-exploitative.

Discussion Question Topic: Deepfakes

Question: Clone your own voice and video using a commercial product. How good is it? Show a friend a real video of you talking, and a cloned video, and see if they can tell which is which.

Answer: The video is incredibly good for the first 10 seconds. After that, however, it becomes apparent how head movement and facial expressions are uncoordinated with speech. Personally I can see when my head starts moving in a different way from how I normally do it. To my friend, the only way they could differentiate the two videos was through voice. The original video audio is not super crisp and clear and has somehow long pauses. The cloned video, on the other side, contains no blemish at all which made it sound fake to my friends. Additionally, there is not switch in tonations like how I normally speak which also became suspicious as my friends were getting more and more confused. Overall, the AI was beginner good but still lacks subtle details in the way humans speak in day-to-day.

To understand why my video had these specific flaws, I looked further into how deepfakes are made. I found out that deepfakes primarily rely on a technology called Generative Adversarial Networks which use two AI models, the “generator” and the “discriminator”. These two models compete with the “generator” constantly fooling the “discriminator” until the later cannot find the difference between real and fake content. The likely reason for my video being terrible is because the discriminator does not have a lot of data with people that talk similarly to me in its training dataset so it did not have high standards to judge the generator model. This highlights how crucial the diversity of an AI training dataset is to the success of the model.

My Own Question Related To The Case Study

What jobs do you think are most likely to be impacted by genAI? Which ones do you think are practically untouchable by AI? Do you have any fear about what AI might be capable of doing in the next 5 years? What is your biggest fear?

Reflection On The Question

This question highlights an important aspect of our capitalist economy. The ability to feed ourselves and dependent. A lot of people probably wonder the same day and night so that is why I came up with it. Also, people constantly fear being left out so the second part about fear is there to try to expose those fears and hopefully prompt the audience to address them.

Personal Reflection

This assignment was so beneficial in my academic and professional pursuits. A does not pass without hearing the word AI so getting such a gentle and easy-to-read introduction into the basics of how it works was so satisfying. One thing I appreciated a lot from the case study was the discussion of labor exploitation by AI companies. This has definitely left food for thought in me so as to not think of the current AI revolution as all rainbows and sunshines.