Generative AI includes existing and novel text and media-producing technologies derived from human ‘prompts’. The tools use machine learning algorithms and are trained by scraping data (sometimes extraordinarily huge datasets) which are drawn upon to generate outputs in different formats. So called ‘Large Language Models’ (LLMs) (e.g. OpenAI’s ChatGPT, Google’s PaLM 2) work by ‘token prediction’. That is, they use probability to determine most likely next words in given strings of text, based on the prompt input.
OpenAI, the company responsible for both ChatGPT and DALL-E, also employs human moderators as part of the training regime known as Reinforcement Learning from Human Feedback (RLHF). Honed and finely tuned prompts (i.e. ‘prompt engineering’) can be used to set tone, style, audience, depth, breadth and other variables. It is worth noting that these are NOT search engines but increasingly they are being coupled with or integrated into them (GPT-4 is integrated with Microsoft Bing Chat, for example). Because prediction is driving outputs, these text-based Generative AI tools are prone to what is commonly referred to as ‘hallucination’. In other words, from one generation to the next, even given the same prompt, one output may be accurate and another fallacious, though with each iteration notable improvements are evident.
Because any text generated is built on a probability model, the text is not actually lifted or copied from anywhere but derived from the corpus of data upon which it was trained. This renders it - like a human-authored essay purchased from an ‘essay mill’ - undetectable by existing plagiarism software and, at the time of writing, very easily made undetectable to even the most advanced generative AI detection tools. Nevertheless, issues around how these tools are trained and copyright remain very controversial and some organisations have blocked companies sourcing web-based data to train their models.
LLMs can also generate code and are already widely used for code error detection and correction. They can summarise documents, re-write in different formats; generate written text in multiple forms and formats including tables and, with an increasing range of generative AI models available, are able to do much more than text production (e.g. image interpretation, chart production, data analysis, slide production and so on). Whilst the churning of reams of text or code is what is initially eye-catching, it is productivity opportunities that often get overlooked. So, for example, ChatGPT or Google Bard can generate a video summary of a given length and tone based on only the video transcript, or create jargon and definition lists or generate quizzes based on a given text. The natural-looking language and fluency impresses, especially on first use and can easily convince users of apparent intelligence and understanding that are simply not there and, notably in chatbot interfaces, lead to users ascribing human qualities to these tools.
It is important to note that much guidance across the sector will include acknowledgement of limitations that, in such a fast moving - and potentially lucrative - landscape have already been overcome. So, for example, you may read that ChatGPT is not web-connected, that its training data stops in 2021 or that LLMs cannot cite genuine sources. Whilst this remains true of the ‘free’ version (GPT-3.5) the bigger, paid model (GPT-4) and other similar tools are web connected, can generate accurate references and do not have a 2021 limit.
Image generation tools likewise use text-based prompts (and increasingly combined with prompts supplemented with other images such as photographs) to produce ‘original’ artistic or photo-realistic images. Like the LLMs, however, they are dependent upon the training corpus and, as a result, one of the critical issues is that the outputs reflect biases in the training data. For instance, what do you notice about these generated images from the prompt: ‘A biochemistry professor at a UK university’?
Generative AI biases: A ‘Midjourney’ generated image, 30 June 2023, M. Compton
An ongoing and compounding issue, therefore, is that text or images generated will themselves be sources for generative AI data sets at points in the future, potentially buttressing and consolidating biases. For an in-depth consideration of responsible AI education see Bentley et al.s (2023) framework.
The JISC Generative AI primer provides excellent further explanations and details about generative AI with some examples of educational use, and look out for the King's short course on FutureLearn.