According to Sydney I. Dobrin, author of Talking about Generative AI: A Guide for Educators (2023), Generative AI works like this:

"A user provides the AI with a prompt asking the AI to create a specific deliverable -- an essay, a song, an image, the solution to a math problem, or so on. The AI then scrubs through all of the data available, looking for patterns and recurring information about the requested task. It then reorganizes that data into a pattern that it deems to answer the prompt. 

We can think of a GenAI as participating in a rudimentary conversation with a user. The user asks a question, then the AI locates information and attempts an answer. This is how most ChatBots function—hence the name “ChatGPT” (Chat Generative Pre-trained Transformer). ChatGPT is just a more complex chatbot, pre-trained to locate data, transform that data, and generate new ways of conveying the data."

According to ChatGPT, some examples of what generative AI tools can do include:

1. Conversational AI: ChatGPT can be used to power chatbots and virtual assistants that can hold natural language conversations with users, answer questions, and perform tasks.

2. Language translation: Generative AI tools can be used to automatically translate text from one language to another, improving communication across language barriers.

3. Imitate genres and styles: ChatGPT can create a scientific treatise, a legal brief, or other genre written in the style of Shakespeare, or at the level of an 8th grader, or ...

4. Text generation: ChatGPT can be used to generate a variety of text, including articles, stories, poetry, and even computer code.

5. Image generation: Generative AI tools can generate images and graphics, including realistic faces, landscapes, and even abstract art.

6. Music generation: Some generative AI tools can generate music, creating unique compositions based on certain parameters or styles.

7. Video generation: Some generative AI tools can create videos from text-based prompts with options to edit and upload media, revise the script, and even format the video for specific platforms.

Some points to keep in mind:

• ChatGPT is trained to sound like a human, not necessarily trained to be factually accurate
• It can plagiarize, and create incorrect or fake citations
• We already have writing assistance such as grammar and style editors, and next work or phrase predictions/suggestions

Because AI tools generate content based on patterns in their training data, they don't "understand" what they're producing in the same way as humans. What an AI tool is capable of generating will depend on the type of tool (e.g., text-generation model or image-generation model) and the set of data it was trained on. ChatGPT, for example, was trained using hundreds of gigabytes of writing on the Internet, but the Internet does not contain the whole world of knowledge. While it can provide extremely helpful explanations (imagine being able to ask Wikipedia virtually any question you want), it sometimes fabricates information as well based on its understanding of what words usually follow others. As the official ChatGPT FAQ puts it, "ChatGPT will occasionally make up facts or 'hallucinate' outputs." These "hallucinations" can include fabricated citations. 

That said, machine learning algorithms are designed to improve performance through experience so that as they process more data over time their responses to user input will become more accurate and effective. Additionally, more academic publishers are licensing their scholarly content for use as training data for GenAI tools, helping to increase the reliability and value of GenAI output. At present, however, there are no standard terms and conditions for such content licensing, which raises significant issues for the authors whose scholarship is being sold, including the absence of compensation, protection, and the choice to opt out.