When comparing ChatGPT and Notebook LM’s abilities to act as a research assistant on racial bias in AI facial recognition software, it became evident that Notebook LM offered a lot more detailed and relevant information than Chat did.

My prompt for Chat was: “Act as a research assistant for me. Researching racial bias within AI facial recognition software using language friendly for college students.” This resulted in many different subsections, all of which contained two or three bullet points that contained brief headings of more that could be found. Some of these subsections included current “solutions” to the bias problem, the “Gender Shades” study mentioned in class, and explained how Ai became biased, blaming training professionals as being biased themselves.

After importing the same sources pulled from ChatGPT into Notebook LM, I then asked Notebook: “Search sources for the reason for the bias in facial recognition software.” Notebook LM provided a lot more detail and was able to reference specific sections within the sources that provided statistics of bias in facial recognition. However, Notebook LM also has a section that places part of the Ai bias on the trainers and the data provided, stating that Ai is not a perfect tool.

Based on these findings, I can gather that Notebook LM would be more useful when writing a research paper and when it comes to creating in-text citations. Nonetheless, based on what both Chat and Notebook provided me, I can safely say that it is not a good choice for police to use Ai facial recognition in investigations, as there is still much training to be done to shrink the size of bias within these systems.

Sources:

National Institute of Standards and Technology (NIST). (n.d.). Demographic Effects in Face Recognition. Retrieved from https://pages.nist.gov/frvt/reports/demographic/

• Buolamwini, J., & Gebru, T. (2018). Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification. Proceedings of Machine Learning Research, 81, 1–15. Conference on Fairness, Accountability, and Transparency.

Sarridis, I., Koutlis, C., Papadopoulos, S., & Diou, C. (2025). VB-Mitigator: An Open-source Framework for Evaluating and Advancing Visual Bias Mitigation. AEQUITAS 2025: Workshop on Fairness and Bias in AI

In my research in class today, my group experimented with asking ChatGPT, and Notebook LM and asked about the ethics of AI facial scanning. Both softwares found that AI has a harder time scanning both women’s, and people with darker skin’s faces compared to men’s and people with lighter skin’s faces. For me, a helpful prompt in my research with Chat GPT was, “Generally, is it easier to scan male’s or female’s faces? Generally, is it more difficult to scan darker skinned people’s faces than lighter skinned individuals?” When Chat was asked this, the response was honest, and said that studies show that there is a higher rate of error for face scanning of women and individuals with darker skin, compared to the error rate of face scanning of men and individuals with lighter skin. A prompt that gave me a confusing response was, “Are you racially biased? Are gender biased?” When asked this, chat immediately says no. However, Chat then points out how biases do show themselves at times, and how it is because of the way it has been trained. Notebook LM differed with Chat, only in the format of its responses. Notebook LM likes to use more detailed responses in paragraph format. Chat uses less detailed responses in bullet point format.

https://www.scu.edu/ethics/focus-areas/technology-ethics/resources/examining-the-ethics-of-facial-recognition

https://journalofethics.ama-assn.org/article/what-are-important-ethical-implications-using-facial-recognition-technology-health-care/2019-02

For this class activity, I explored the Notebook LM and its capabilities in summarizing the implications of A.I. in facial recognition. I asked Notebook LM to scan the web for surface-level sources relevant to this topic. There was an option for a “deep search” as well, but the surface-level search returned a vast number of peer-reviewed sources, all relevant. Some sources examined the implications for law enforcement, while others examined the implications for health care.

One of the most troubling implications that Notebook LM laid out for me was the high likelihood that darker-skinned females were misidentified. Light-skinned males had an error rate of 0.8%, compared to 34.7% for darker-skinned females. Scraping social media to find training data is another concern. Just because an image is publicly viewable does not mean its biometric data should be a free resource for corporations to train their LLMs.

I think Notebook LM can be used to synthesize sources. I would ask Notebook LM to find its own sources sparingly, as I am concerned about the credibility of the sources that LM generated for me. I think, after finding sources for yourself, LM could be a good tool for summarizing the information sources have in common, though it’s definitely not a substitute for actually reading these sources if you need to write a literature review, or something like that.

Group members who explored ChatGPT found it was very confident in all its responses, even though the ideas were more general than those of Notebook LM. One thing I found interesting about Notebook LM is that it often backs up its passages with citations directly from the sources, so you can go back and read the author’s words to make sure LM isn’t hallucinating. I think Notebook LM has uses in writing papers. They could help you see connections between sources that are hidden in plain sight.

Source: https://notebooklm.google.com/notebook/e8482b60-de6d-4b3c-bd07-23e7e34e3b71

This made me think about how important it is for developers to test AI systems on diverse groups of people. AI is going to keep growing and being used in things like security, jobs, and schools, so it needs to be fair. In the future, I think AI will become even more common, but people will also push harder for rules and accountability to make sure it’s used responsibly. One thing that really stood out to me was learning how facial recognition systems don’t always work the same for everyone. Some studies showed that these systems can make more mistakes when identifying people with darker skin tones. That made me realize that technology isn’t always neutral—it reflects the data it’s trained on.

During todays class, I used ChatGPT as a generative AI source, while other members of my group used NotebookLM. We made an attempt to research and discover AI hiring tools as a topic and see which sources came up to compare reliability. Group members that did not use ChatGPT found sources like “15 best AI hiring tools” and “10 Best AI hiring sources.” While using ChatGPT, I found sites that were about AI sourcing and outreach such as hireEZ and Zoho Recruit. I believe that ChatGPT is more reliable because I have used it to find me sources on topics, and I’ve used it for citations to format properly. Often, it’s easier to do because you can just copy and paste the source into the search bar and then ask it to format the source properly. I’m sure others prefer to use NotebookLM, but I’ve seen both sides and I have used what I prefer. I do not have any plans to use Ai any differently going forward, but it was brought to my attention that ChatGPT is not great for sources to find, and I have to be more careful with the usage of it while avoiding bias. This is something I’m taking into consideration moving forward heavily. If I’m being honest, I don’t think AI is headed in the right direction, I believe it’s taking over for the worse and not the better. It is taking up things like job opportunities, energy, and data. This can be a massive problem in the future. An argument from another groups findings was about the trajectory of AI and how people also think it’s getting worse. It’s scary to look at the many ways robots can project information as well as personal data.

Sources ChatGPT pulled up:

Resumly. “What AI Tools Recruiters Use for Screening: 2025 Guide.” Resumly AI, https://www.resumly.ai/blog/what-ai-tools-recruiters-use-for-screening-2025-guide. Accessed 12 Mar. 2026.

HireGen. “Top 5 AI Recruitment Tools You Should Be Using in 2025.” HireGen, https://hiregen.com/posts/top-5-ai-recruitment-tools-you-should-be-using-in-2025. Accessed 12 Mar. 2026.

Scalar. “The 10 Best AI Recruiting Tools to Supercharge Hiring in 2025.” Scalar, University of Southern California, https://scalar.usc.edu/works/the-10-bestai-recruiting-tools-to-supercharge-hiring-in2025/index. Accessed 12 Mar. 2026.

ChatGPT is strange unless you’re direct. The most strange thing about ChatGPT’s model was that my prompt had to be really specific in order to find academic sources. First I asked it for academic sources related to facial recognition involving policing, environmental studies, or healthcare. It came up with journal articles and one YouTube video, but nothing from a book or database. After my third ask, I was able to filter down 5 from a database and then put those into Notebook LM. Notebook summarized these sources together and claimed that “Facial recognition technology (FRT) in policing creates a conflict between surveillance efficiency and democratic accountability. Public support is often performative; anonymity reveals many citizens privately harbor reservations about biometric tracking. Empirical data shows FRT deployment correlates with increased racial disparities, specifically raising Black arrest rates while decreasing White rates. This stems from automation bias and pre-existing structural inequities. Global regulations remain fragmented; the US lacks the robust accountability frameworks found in the EU, necessitating urgent, transparent impact assessments to protect civil liberties”. Through this lab, we learned that both AI models have strengths, but Chat is not going to excel at what Notebook excels at and vice versa. Chat struggled at finding these sources and struggled even further on giving me great summaries of the sources they provided as it was multiple steps. My prediction about where AI is headed is a continued reliance because it comes up with sources instantly rather than a trip to the library or a database. I learned that Chat says its prompts very confidently and if you do not check it for error, you are using it incorrectly.