Generative AI

This guide provides a basic overview of generative AI, including ethical issues and potential costs, in order to help you make informed choices about using AI tools.

Generative AI refers to AI models that can create new content, such as text, images, code, audio, or video, in response to user prompts. Generative AI uses machine learning techniques to learn the patterns and structures of training data in order to create new outputs with similar characteristics.

Here's a plain language definition from The New York Times:

"Generative A.I.: Technology that creates content — including text, images, video and computer code — by identifying patterns in large quantities of training data, and then creating original material that has similar characteristics. Examples include ChatGPT for text and DALL-E and Midjourney for images." (“Artificial Intelligence Glossary: Neural Networks and Other Terms Explained")

Generative AI models generate content by "statistically analysing the distribution of words, pixels or other elements in the data it has ingested and identifying and repeating common patterns."(1)

Compared to other types of AI, generative AI can be understood as AI that "generates," rather than "discriminates." (2)  AI that "discriminates" is used in classification and prediction tasks, like image recognition.

Generative modeling can be defined as:

"a branch of machine learning that involves training a model to produce new data that is similar to a given dataset."  (3)

This presentation by Ellie Pavlick, Brown University Professor and Google AI Researcher, from the MIT GenAI Summit in 2023 provides a brief overview of generative AI, how it works, and opportunities and risks.

Important to know

• Large language models (LLMs) predict the next likely word in a sequence to generate fluent, plausible-seeming text. They do not "understand" prompts or text in the way another human would.
• AI models are trained on massive datasets often scraped from the internet, which can include social media posts, personal websites, and pirated content. Their training includes annotation and testing by humans to remove offensive or harmful outputs.
• We often don't know exactly what data AI models have been trained on, which means that in most cases, users have no way to know what information an AI tool had access to during development.
• Generative AI often produces inaccurate and biased information, as a result of their datasets, development process, and a lack of true "understanding" of the material. They are complex statistical models, not thinking machines. However, because their outputs appear to mimic human capabilities like language, people may place high levels of trust in these tools without recognizing their limitations.

Generative and discriminative AI may also be used in combination, in order to refine and iteratively improve outputs, in an approach called Generative Adversarial Networks (GANs).

This visual storytelling article by the Financial Times provides a more in-depth exploration of the transformer model, a development key to how large language models function today.

• Generative AI exists because of the transformer. Financial Times. September 12, 2023. By Visual Storytelling Team and Madhumita Murgia.

Sources

1. Guidance for generative AI in education and research. UNESCO. 2023

2. What is Gen AI and How is it Impacting Education? A Presentation by Scott Alfeld, Assistant Professor of Computer Science, Amherst College. Sept 13, 2023. 

3. Foster, David. Generative Deep Learning : Teaching Machines to Paint, Write, Compose, and Play. Second edition., O’Reilly Media, Incorporated, 2023.

4. Pavlick, Ellie. "Getting on the same page about GenAI." YouTube, MIT AI ML Club, March 13, 2023. https://www.youtube.com/watch?v=f5Cm68GzEDE

Text AI / Chatbots

Users provide text or voice prompts to these tools, which are designed to provide fluent, conversational responses. For example: ChatGPT, Google Gemini, Microsoft CoPilot.

Image AI

Users provide descriptions of images or image effects, with options to modify the output in various ways. For example: DALL-E, Midjourney, Stable Diffusion, and Firefly.

Code AI

Users can provide specifications to generate code, or review and check code along specific criteria. For example: GitHub Copilot.

Video, audio AI

Users can provide prompts to generate videos or create video effects. Users can provide text to render in audio, or apply effects to audio, including rendering in different "voices."

These tools are continuing to develop, with new integrations in text-based tools such as voice prompting and text-to-speech and image generation and recognition, providing more modes of input and output. There are also smaller scale models being developed with specialized datasets and uses.

Privacy, security, accessibility

Generative AI tools differ in their levels of security, privacy, and accessibility features. Before you start using a tool, make sure you understand its privacy, security, and accessibility levels, along with associated risks.

Check Amherst IT's generative AI tool ratings and review the recommendations on that page in order to mitigate risks to yourself and others.

Getting started with prompts

A prompt is a starting point or instruction you give to the model to generate specific content. It can be in the form of a statement or a question. Think of it as a first step to getting where you want to go.

Making good prompts

There are lots of different prompt types you can try, but here are some of the basic elements of a good prompt process:

• using simple and clear language
• including context or background information that's relevant to your desired output
• being specific about the format, length, style, etc. of the output that you want
• including examples of the desired type of response or format for outputs
• avoiding prompts that may generate harmful or inappropriate content
• refining and iterating in response to outputs

Sometimes, it can be challenging to get a tool to produce exactly the output you want. They are often sensitive to slight wording changes in a prompt (ex: "fair" instead of "just"), and may produce different outputs to the same prompt over time.

Important: Generative AI tools may operate in a conversational manner, but they do not "think" in a way a person would. If you ask a GenAI tool to explain itself, it is not actually able to do this! It will provide a plausible output that sounds like a good explanation instead. If you find that your tool is not providing good outputs, it's best to start over with a different prompt strategy.