Week 11 Reflection (Savannah)

Introduction

This week’s topic covered AI and Large Language Models, specifically ChatGPT, and its interactions with various privacy concerns. 

Monday

To begin, we were informed on a recent development with NYU college admissions, hackers, privacy, and affirmative action. This led to the topic of the five data privacy concerns with AI and their details, which included: data collection, data security, data use, data persistence, and bad actors. Next, we transitioned to Carsen’s blog post discussing ChatGPT’s privacy policy, as it was updated on April 2, 2026. We talked about the largely broad concept of “Personal Data” and the long list of third-parties ChatGPT included in their policy. In addition, we covered the potential information government authorities can potentially access. The different versions and personalized advertisements were highlighted. The class then focused conversations on Carsen’s privacy suggestions for AI platforms/LLM models: 1. By default, LLM models should not train its models on user inputs unless a user specifically opts in and 2. There should be “automated mechanisms that reduce the chance of sensitive information sharing and use for model training”. The question of whether protecting companies from consumers/their privacy concerns or/while balancing the protection(s) of consumers from themselves arose. 

Wednesday

We began class with news regarding recent developments, particularly with motor vehicles and tracking information. We then chatted briefly about AI’s (Gemini’s) accuracy via the publication from the New York Times. The class learned that, depending on the version, roughly 9-15% of outputs can be inaccurate. After, we transitioned to a discussion about the difference between the deep and dark web. Topics such as the Silk Road and the pardon of Ross Ulbricht were mentioned. This led to the question of the week. There is no law from any level (local, state, or federal) that has weighed in on this issue, making the question challenging to respond to. The conversation considered the idea of a potential “viewer discretion advised” label before inputs made by users. In addition, an action of consent by the user with/and incentives were brought to light. Again, the question of whether companies or users are more at fault for the shared information and what is done with it arose.

Takeaways

One detail from the week that stood out to me was the overarching question, whether stated directly or indirectly: who is and who should really be at fault–companies or consumers? There were arguments for both cases, which was interesting to listen to and explore. I tend to agree that it is on the company to be as transparent as possible. However, some of the examples that were defined as “sensitive information” pushed me to lean more towards the stance of the user being responsible. With that being said, I do not think AI companies should necessarily retain and use sensitive information. On a similar topic, the idea of ‘what “sensitive information” is’ was intriguing. The way I thought about it was that the concept would have different definitions, meanings, and roles based on the individual. Therefore, how could a company or entity specifically generalize what “sensitive information” is, looks like, and/or should be treated if there is ambiguity? 

As a side note, I found the reading about the black market to be quite insightful, as I did not know very much about it before this class. More specifically, the chart illustrating the levels/depths of the internet was informative and intriguing, while also slightly nerve-racking. 

We, as a society, have invented technologies with the goal to tackle restrictive obstacles limiting some form of collective advancement. Although efficiency has become a strong seller, we bear the weight of the unknown. Yes, there is information about these platforms; however, many of these platforms have not existed long enough for extensive research to be conducted. Therefore, we are only aware of the surface level side-effects, not of the long term damages. I am interested to see how these models grow, what they are used for, where they take us, what restrictions and/or legislation is implemented, and how privacy factors are considered overtime.