By Dr Tristan Jenkinson
Introduction
OpenAI’s ChatGPT application has been a huge talking point in the world of for the best part of the last 10 months. The application was made available for the public in November last year, and just two months later, had already gained 100 million users (see for example this Guardian article).
Having registered an OpenAI account back in June 2022 (because of my interest in the DALL·E 2 system – more on that in a later post), I should have been in a great position to jump in and start playing with ChatGPT swiftly. Sadly, it was a while before I was able to get the chance to really investigate ChatGPT properly. Due to work and personal commitments, it has taken even longer to write up my thoughts and notes on the topic. As a result, I now have a whole series of blog posts to write up.
In this series, I wanted to provide my own perspective on ChatGPT (and Large Language Models in general). I plan to cover a range of topics, starting with covering “What is ChatGPT” (a topic which some may consider redundant, given that over 100 million people already using it), and then going on to discuss areas such as the limitations of ChatGPT, how it has been misused and abused, some further discussion of generative AI (with an aside on audio AI in particular), some legal discussions and concerns, before moving on to discussing the use of ChatGPT and related technology in eDiscovery and legal tech more generally. I then wanted to discuss the investigation of ChatGPT, and if possible, write up a case study on the Mata v Avianca case (which is a fantastic example of how ChatGPT should not be used in legal cases).
With that in mind, lets jump into the main content of the first blog in this series, and talk about what ChatGPT is…
What Is ChatGPT?
ChatGPT is a type of chatbot designed to provide responses to text prompts or conversations.
If you have used a banking (or other services) website recently to try to get something done, you may have experienced chatbots before… A chat session not with a member of staff, but with a computer program. You try and explain what you are trying to accomplish, and the chatbot will provide you with some options to try and resolve your query.
ChatGPT is like this, but far, far more advanced.
The concept is the same, a textual back and forth chat conversation, in normal language, but the responses from ChatGPT are at an almost human level of comprehension and response. It is not dissimilar to having a Microsoft Teams conversation with a very knowledgeable colleague.
To act in this way ChatGPT processes the “natural language” (i.e. everyday language that a human may use, rather than computer commands, prompts, or code) input, and generates (or responds with) natural language that can be easily understood in response.
What’s in a name?
It might be helpful to understand where the name ChatGPT comes from, as this helps to understand a little more about the background of the application.
The Chat part relates to the system being a chatbot, as discussed above. But what does GPT stand for? It stands for Generative Pre-Trained Transformer… which needs a little more breaking down:
- Generative – this means that the system is an AI system which generates something. There are other AI based systems where the output is purely an action, or operation, whereas here we have AI generating data in response to the input it is provided.
- Pre-Trained – this means that ChatGPT has been trained on some large data set. This is how it “learns” how language is structured and gains an understanding (at a statistical level) of the various topics and concepts in the underlying data set on which it is trained. In the case of GPT 4.0 (the architecture under the latest version of ChatGPT in use) the model is understood to have been trained on around 1 Petabyte of data from the internet (see for example this article from dataloop). The data training set for ChatGPT is a topic that will be revisited in a later post.
- Transformer – a transformer is a specific type of AI model. The structure behind this type of model was originally laid out in 2017, and the explanation gets very technical very quickly and (you will be happy to know) is outside the scope of this blog post! Transformers allow the model to consider a whole sentence or paragraph at once (rather than a word or token at a time), meaning that it could then better understand context. This then allows the model to identify key information that is needs to focus on (known as the attention mechanism in AI terms) as it prepares a response. The original paper introducing this concept is “Attention is all you need” by Ashish Vaswani. For a simplified approach to transformers, this towardsdatascience page is a good starting point.
Large Language Models
You may have also heard ChatGPT being described as a Large Language Model. This is a description of a group of transformer based models which designed to be large scale and are trained on large data sets.
Typically, these models are large and complex and do not run on a local machine. Requiring a large amount of computing power. For this reason many large language models, such as ChatGPT are run as applications which link online to a service which does the hard computation on the back end.
This has benefits, such as normal users being able to utilise large language models such as ChatGPT without having to run hugely powerful processing systems in order to do so. However, in contrast there are concerns, such as the data inputs being provided to the party running the processing model, which could have privacy and/or security concerns.
What Can ChatGPT Do?
People are still learning about the capabilities of ChatGPT and what it can be used to do. It might seem that initially, a computer program that will chat with you may not seem like it could be hugely useful… the big consideration is the amount of data from the internet that it has been trained on. As mentioned earlier, ChatGPT is like a Teams chat with a knowledgeable colleague… but now consider that this is a colleague who has memorised a huge chunk of the data availalbe on the internet.
Imagine using Google to run searches, but rather than giving you links to websites (that may, or may not have the answer) ChatGPT can just give you the answers. Or, it can give you the answers, with references to sites where you can go and check the answers.
ChatGPT can also write content, give it a subject and it can write an article for you. ChatGPT can write in a number of different styles, or imitate existing authors. For example, I asked ChatGPT to create a synopsis in the style of William Gibson, I’ve included the full prompt at the bottom of this article.
But it isn’t just for content creation, you can provide it written content and use it to critique the content – many are using ChatGPT to provide guidance on content like articles and CVs.
I’ll be discussing some of the uses of ChatGPT and Large Language Models in a later post, but in the meantime, sites like Forbes have lists of some of the uses for ChatGPT.
Prompt Engineering
The above gives some very high level ideas about what you may be able to do with ChatGPT. One of the benefits of ChatGPT using natural language is you can login and say “Write me an article about X”. ChatGPT will respond and create an article for you, however, there may be a better way to provide your request to the system.
Prompt engineering is the emerging skill of creating prompts in such a way that you get an optimum response from generative AI systems such as ChatGPT. This is an extension of the old adage “Garbage in means garbage out”. In theory, the better the prompt, the better the results may be.
Already there are a large number of approaches that can be employed, such as using different types of prompt. For example demanding different outputs, such as open ended responses, step by step instructions, or even an output in code. You can also assign roles to an AI, ask for responses in a particular style, or use other different approaches, such as providing examples, or using chained or nested prompts.
This continues to be an evolving area, and is likely to become a commonly sought after skill. Prompt engineering may be something that employers will expect to see on CVs in the future, especially in the legaltech world.
Summary
Ultimately, ChatGPT is a very powerful tool, which has a huge variety of uses that users can sign up for and use for free. There are however, some significant limitations which users should be aware of. That is one of the topics that I will cover in part two, coming soon.
In the meantime, I’ll leave you with the prompt and resulting plot synopsis that ChatGPT put together for a William Gibsoneque story, including the title that ChatGPT created. Bear in mind, this synopsis could then be fed back into ChatGPT to build a step by step detailed plot synopsis to plan out how to write the novel.
Neon Diaspora
Prompt:
Act like a new author, create a synopsis of a new, original science fiction novel written in the style of William Gibson
ChatGPT:
Title: Neon Diaspora
Synopsis:
In the sprawling neon-lit metropolis of New Shanghai, where advanced technology intertwines with the gritty underbelly of a cyberpunk society, a disillusioned data-hacker named Rook finds herself caught in a web of conspiracies that span the darkest corners of the digital realm and the physical world.
As a freelance operative on the fringes of society, Rook is haunted by the memories of her past and seeks solace in the virtual planes of the Deepnet—a hidden network beyond the reach of conventional surveillance. But when a mysterious package, encoded with an enigmatic algorithm, falls into her hands, Rook becomes the target of powerful forces who will stop at nothing to possess the secrets within.
Driven by a relentless pursuit of truth, Rook embarks on a perilous journey through the neon-soaked streets of New Shanghai. With the help of an enigmatic information broker known only as Jinx, Rook navigates the city’s tangled underworld of corporate espionage, renegade hackers, and transhumanist subcultures.
As she uncovers the layers of deception woven into the fabric of society, Rook discovers that the lines between human and machine have blurred beyond recognition. The technology that once promised liberation has become a weapon, controlled by a shadowy cabal known as the Syndicate. Their agenda threatens to reshape the destiny of humanity itself.
As the world teeters on the edge of chaos, Rook must confront her own demons while deciphering the secrets hidden within the encoded package. With every step, she uncovers a vast conspiracy that spans multiple dimensions, merging artificial intelligence, quantum computing, and the nature of consciousness itself.
In a race against time, Rook must ally herself with unlikely companions and confront the harrowing truth that her destiny is intertwined with the fate of the city she calls home. In the neon-lit heart of New Shanghai, where reality and virtuality converge, she must harness her skills as a hacker, uncover her true identity, and embrace the power within to ignite a revolution that could reshape the course of humanity.
Neon Diaspora is a mind-bending, high-octane cyberpunk thriller that delves into the murky depths of technological advancement, the boundaries of human existence, and the eternal struggle for freedom in a world ruled by the relentless march of progress.
The eDiscovery Channel 