Artificial intelligence is not as neutral as I once thought. Crawford explains that AI can reflect the “characteristics of the dominant voice in their code and in the data they use to learn, and that is male and white.”¹ This stood out to me because AI is created by humans, meaning the biases that already exist in society can become part of the technology itself. Darker skin being treated differently by some AI systems illustrates this problem because the technology is supposed to work for people, yet the information used to create it can already be unequal. Crawford also discusses the Calgang database, which contained inaccuracies and disproportionately included Black and Latinx people.² This shows that data is not simply information. Data can affect real people. Chiang makes a similar point in “ChatGPT Is a Blurry JPEG of the Web,” where he compares ChatGPT to a blurry image that keeps the general picture but loses some of the original details.³ ChatGPT can therefore produce something that sounds correct while still being wrong. Both Crawford and Chiang show that AI should not automatically be trusted simply because it comes from a computer.
The IBM reading helped me understand what is actually happening behind the technology. Before reading it, I was mostly uneducated on how Large Language Models worked. I learned about tokenization, embedding, fine-tuning, and reinforcement learning from human feedback. The idea that stood out to me most was that an LLM does not simply predict the final answer. It predicts one token at a time based on what it learned during training.⁴ This explains why ChatGPT can sound confident even when it is wrong. It is making its best prediction at each step. RLHF also explains why different AI systems can have different personalities because models can be “adjusted to respond in a way that’s more casual, humorous or brand-consistent.”⁵ I found this interesting because I have used different AI systems and noticed that some seem more friendly or serious than others. It is not necessarily because the AI has a personality of its own. It can be the result of how humans trained and adjusted the model.
These readings made me look at AI differently. I used to think of AI as something that simply gives an answer after being asked a question. Now I see it as a system built from human information, human decisions, and human data. That creates a problem because humans are not neutral, and neither is the information they produce. Crawford states that “data is more commonly seen as something to be taken at will, used without restriction, and interpreted without context.”⁶ That is concerning because data can represent real people, and mistakes in that data can affect real people. At the same time, Chiang’s argument shows that AI can take enormous amounts of information and produce something that looks complete even when pieces are missing. AI is therefore useful, but it should not replace human judgment. The technology can help us find information, organize ideas, and understand complicated subjects, but we still have to question what it gives us. AI can make mistakes. AI can carry biases. AI can sound confident while being completely wrong. That makes the responsibility of the person using it just as important as the technology itself.
¹ Kate Crawford, Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence (Yale University Press, 2021).
² Crawford, Atlas of AI.
³ Ted Chiang, “ChatGPT Is a Blurry JPEG of the Web,” The New Yorker, February 9, 2023.
⁴ Cole Stryker, “What Are Large Language Models (LLMs)?” IBM.
⁵ Stryker, “What Are Large Language Models (LLMs)?”
⁶ Crawford, Atlas of AI.
