What is Generative AI?

Open AI's notorious AI chatbot, ChatGPT, was released in late 2022, inspiring the hype around Generative AI.

Generative AI like ChatGPT can learn from existing artifacts to generate new content that reflect the characteristics of their training data. (Read More)

ChatGPT is trained on a Large Language Model (LLM) containing large amounts of public and some private data, largely dated before 2021.

What are its limitations?

While it may feel like magic, the technology has some inherent limitations. To understand any AI technology, it's important to understand:

a) the data that it was trained on, and 

b) the basics of how the technology is designed.

For ChatGPT specifically, here are the main limitations that users should keep in mind:

• Outdated Information: The LLM that ChatGPT is trained on contains data from before September 2021, and it does not learn from experience. For anyone using the tool to conduct research, they should be aware that it lacks recent information.*
• Hallucinations: LLMs can have something called "hallucinations", which is a response that is unjustified by the training data. In other words, it creates fake, false, or unreasoned responses. This often happens when you ask for a citation, only to find that the resource itself does not exist or contains a dead link.
• Biased or Harmful Information: ChatGPT is trained on a language model of over 3 billion words from the internet. Therefore, it will have been trained on all of the same biases and harmful content that exists on the open web. Users should keep this in mind when evaluating its responses.
• Proxied or Paywalled Content: ChatGPT cannot get past proxied links or paywalls, so it is limited to whatever is freely available online, within the dataset it was trained on. This means that it, for the most part, cannot reach our databases and other library resources. 

- From the Artificial Intelligence Resource Guide by Horn Library at Babson College

Evaluating Factors to Keep in Mind 

• Design: Large Language Models (LLM's) are predictive models, not search engines. They are designed to predict what word is most likely to come next. Their logic is based on statistical reasoning, not subject expertise. 
• Training Data: Consider the data that your tool of choice is trained on. Most LLMs are trained on data scraped from the internet. If this is the case, even if your tool has up-to-date internet access, it is still only as good as what is freely available online.
• Sources: Some generative AI tools will give you the sources that they used to generate the answer to your question, so that you can evaluate those sources (and therefore the answer) for credibility, bias, etc. Most generative AI tools are unable to give you an explanation of how they came to any particular answer. This is due to AI systems' "black box problem", which makes it impossible to determine how an AI reached its decision. When asked to produce its sources, most generative text-based AI's will refer back to their training data.
• Terms and Conditions: We know -- no one reads the terms and conditions. But these tools are still in development, and so the norms around privacy, data protection, and content moderation are still being established. If you're considering using a generative AI tool for a course, to develop your business, or in the workplace, it is worth looking at the fine print. 

- From the Artificial Intelligence Resource Guide by Horn Library at Babson College

ROBOT method