Generative AI technologies

Generative AI technologies consist of Large Language Models (LLMs) such as OpenAI’s “ChatGPT”, Google’s “Bard”, Microsoft’s “Bing Chat”, and other similar technologies which enable users to ask questions or tell a story, and the bot will respond with relevant, natural-sounding answers and topics.

The interface of these technologies is designed to simulate a human conversation, creating natural engagement with the bot. These technologies are optimised for conversational dialogue using Reinforcement Learning with Human Feedback (RLHF). Responses produced by these systems can be human-like because they were trained on vast amounts of data written by people. More advanced tools can analyse data and produce a range of data visualisations.

An example is given below:

Image of a prompt and output generated using Google's Bard

An image of a prompt and output generated using Google's Bard.

There are also image generation tools such as Midjourney, DALL-E 2, Stable Diffusion, and Adobe Firefly that work in a similar manner to the text generators but were trained on datasets of images to produce an original image (with variations) following a prompt by a user.

An example is given below:

Image of state-of-the-art paramedic simulation suite generated using Midjourney

Image of a state-of-the-art paramedic simulation suite generated using Midjourney.

There are concerns about the copyright of the training material of both text and image generation tools.

Some of the strengths of these technologies are that they:

• can provide answers to a range of questions, including coding, essays, multiple-choice, and maths/quantitative problems
• continue to improve in terms of their capabilities through updates
• generate unique text-based output each time a user provides a prompt, regardless of how similar those prompts might be
• are assumed to be effective in performing analytical tasks.

Despite strengths, these technologies also have limitations, including that they:

• can produce unreliable information, therefore any content produced requires professional judgement to check appropriateness and accuracy
• return results based on the dataset it has been trained on. In many cases, a given tool may not have been trained on the curriculum relevant to your study. They may also exhibit political and cultural biases
• are prone to “hallucinations” and can make up scientific references to papers that do not exist. This is especially common with LLM bots that can provide you with incorrect facts with assertions, which can easily make you think that those facts or information are true, even though they are not.