AI Blog II: Basics of How AI Systems Work
In order to understand the ethical implications of Artificial Intelligence (AI) use- it’s first helpful to understand the basics of how AI functions. Due to limitations in scope, this blog will focus primarily on Generative AI systems as most commonly used AI tools are generative in nature.
Overview
Generative AI (GenAI) does exactly what you might expect- it generates. GenAIs have the capacity to generate text, images, and even songs after receiving a prompt from a user. In order to generate, GenAI rely on patterns gleaned from the analysis of massive datasets. The process goes something like this:
- Huge amounts of data are fed into a training algorithm based on what kind of generation the tool will be performing. Text generators will use giant sets of text while image generators will use giant sets of images. These data sets are often referred to as training data.
- The data is then analyzed to identify patterns. What kind of system is used to analyze the data will depend on the complexity and cleanliness of the training data used. While simple machine learning is good for cleaner and less complex data sets, neural networks are often used for sets with large amounts of irregularity and nuance-like natural language.
- Using the identified patterns- the model can begin to make predictions. These predictions are the generations that the user receives when they enter a prompt.
While the description above is a very simplified overview of GenAI systems, one thing remains glaringly clear. These systems are not “thinking” but merely predicting using patterns gleaned from access to massive data sets. When a user inputs “What is the capital of France?” into an AI tool, the tool is not thinking back to its 10th grade World Geography class, but rather seeing that in the training data “the capital of France” is nearly always referencing the city of Paris. This is why the quality of the training data is massively important. A GenAI trained on a data set where the capital of France is always referred to as Madrid would have an incorrect response to the prompt.
The Hallucination Problem
A simple look about will illustrate that GenAI models are more than capable of being incorrect. Why would an incorrect response be preferable to a simple “I don’t know”? The answer to that question is complex, but boils down to two main issues:
- AI models don’t know that they don’t know.
- AI models prioritize and are rewarded for thoughtful and confident responses, even when they aren’t right.
Remember that GenAIs cannot think, only put together strings of text based on predictions. They do not have the ability to reflect on knowledge they do or do not have. If “I don’t know” isn’t an answer that appears commonly in the training data, it won’t be an applicable answer either. Consider as well that most AI Models are programmed to prioritize user interaction. The longer an interaction with a user goes on- the more data can be added to the predictive model. The more data a predictive model has, the more accurate it can be. User input on an incorrect answer, too, hones the model further. If you correct an AI tool when it gives you an incorrect answer- you are improving the data set and therefore improving accuracy for future prompts. It’s easy to see that the only loser when an AI tool is incorrect is the user. To the tool- there are only benefits and there is no incentive to admit a lack of knowledge.
Using this overview of AI systems, the next blog will examine the ethical implications and considerations of AI use.