Introduction
Imagine you arrive at school one day, and your teacher has a new challenge for the class. She tells you that she has a specific objective in mind, but she’s not going to tell you what it is. Instead, she asks you to try different things. Each time you do something, she’ll either give you a thumbs-up if you're getting closer to her goal, or a thumbs-down if you're way off. With no instructions or clues, you have to figure out the right strategy just from trial and error.
Your friend pulls out a book, and gets a thumbs down. You walk toward the whiteboard, and that gets a thumbs up! It must have something to do with writing on the board! Eventually your class figures out she wanted you to draw a picture of the original American flag with only 13 stars on it. It was great to figure out the goal, but that was a strange lesson!
Explanation
This may not be the best way to run a classroom, but it highlights how computers learn to create things like text or images. Computers guided by Artificial Intelligence, or AI, start with trial and error, producing different outputs to see what works. Just like you would adjust based on your teacher’s feedback, the AI gradually figures out patterns from its experiences. The output improves over time, even though the AI doesn’t fully understand the purpose behind its tasks. Before it’s properly trained, the AI’s results can seem random or off-target, but they get better with more feedback.
Definition
Generative AI refers to artificial intelligence that can produce new content like writing, artwork, or music by learning from large sets of existing examples. The word “generative” means that it is creating something that doesn’t already exist. Instead, it is basing the creation on information people have given it. As it processes data, the AI gets better at giving people what they want.
How It Works
In 2022, a company called OpenAI revolutionized how people interact with computers with the public release of ChatGPT. This app, which is an example of a “large language model” or LLM, uses computers trained on huge amounts of internet text to craft text that sounds like a human may have written it. This may seem like extremely modern technology, but chatbots like ChatGPT have actually been around since the 1960s. One of the earliest was called ELIZA. It could run different pre-programmed scripts, such as one in which the computer acted like a doctor asking a person questions about health.
Generative AI chatbots, like ChatGPT, work by analyzing enormous amounts of existing data from sources like books, websites, and other online content. These systems are trained using this data, allowing them to learn patterns in language and relationships between words. When you ask ChatGPT a question or give it a prompt, it uses what it has learned to generate a response based on these patterns. The AI doesn't “understand” the content like a human would, but it predicts what words or sentences should come next based on what it's seen in its training data. For tasks that have plenty of data—like answering common questions or writing about well-known events—it usually does a good job. However, it can struggle when asked to handle more unusual tasks or less-written-about topics.
If given a task with little to no data, such as completing a sentence about something highly unrealistic, ChatGPT still attempts to generate a response based on similar patterns. For example, if you ask it to write a story about a spider that ate an entire town, it probably wouldn’t have much relevant data to use. Since there probably are no example stories about that scenario, the AI may produce sentences that seem bizarre or inaccurate. In these cases, the model has to rely on more general patterns of language and storytelling, which can lead to responses that feel unreal or nonsensical. This is because the AI doesn’t have the context or common sense that humans use to interpret unusual or imaginative tasks.
Image generators, like those used in creating AI art, work similarly but with visual data. These models are trained on countless images and learn to generate new ones based on patterns in color, shape, and composition. Like ChatGPT, they can produce highly realistic images when tasked with something familiar. However, they are often off the mark when asked to generate something unfamiliar or abstract because, like text-based models, they rely on existing data and have no inherent understanding of the content. Whether creating images or text, these models can sometimes produce results that seem far from the intended goal, especially when the task is complex, ambiguous, or highly specific.
So What?
Understanding how generative AI works, and why it can produce strange or inaccurate results, is important, especially as this technology becomes more integrated into everyday life. One major reason to care is that AI systems, if not carefully trained, can reinforce harmful biases. There’s a well-known story of a company that used AI for help with hiring people. The AI learned from historical data that included a lot of managers that had names that sounded like white males (such as “Robert Johnson.”) Trained on this data, the AI prioritized white male-sounding candidates, even when it wasn’t relevant to the candidates’ qualifications. The company hadn’t told the AI to focus on names at all, but it still learned that detail. These kinds of issues can have real-world consequences, reinforcing inequality or making poor decisions that impact people’s lives.
This issue matters because there is a good chance that you will increasingly use AI tools for schoolwork and other tasks. Because of its unsupervised learning, AI often produces poor or irrelevant answers. A response may be confusing or wrong, and the AI wouldn’t know the difference. It is essential to critically evaluate the AI's output, rather than blindly trusting the information it generates. Especially now that AI answers are an integrated part of more and more online information, including Google searches.
On the positive side, generative AI can be a lot of fun and incredibly useful for certain tasks, such as brainstorming or generating ideas. If you're stuck on a project or need to think outside the box, AI can offer a range of suggestions that might spark new ways of thinking. It’s also great for tasks like helping you organize your thoughts or drafting early versions of writing assignments. However, it’s important to remember that generative AI works by learning through patterns and adjusting over time. Just like in the scenario where your teacher gave you a challenge without telling you the goal, you may need to give it directions to provide what you want. AI is a tool, not a substitute for critical thinking. It can help generate ideas, but it’s up to you to refine those ideas and ensure that the final output is meaningful and accurate.
