ChatGPT: Hype or the Next Big Thing? Everything you should know
Author Dr Aaron Snoswell
Date 17 April 2023
Since its launch by OpenAI in November 2022, ChatGPT has dominated the headlines. The multi-talented chatbot can generate reports, translate content into different languages and answer a wide range of questions and even write code. Now with Bing’s new chat search features, Google’s ‘Bard’ chat-search, and Microsoft integrating chat features into the office suite of software, it seems this technology has rapidly come to the fore.
But the technology behind ChatGPT is not new. Natural Language Processing (NLP) has been around as long as Artificial Intelligence, since the 1950s.
The difference now is the increase in scale – the size of the models, and the amount of data used to build them – and the processes and technical methods used to build these systems.
What is ChatGPT and how does it work?
ChatGPT is an artificial intelligence dialog agent (or ‘chatbot’) developed by OpenAI and released in November 2022. ChatGPT converses in natural language, and can enable anyone to write computer code, craft poetry, or summarise long documents.
ChatGPT is built using a Large Language Model (LLM). LLM’s are essentially an advanced form of the ‘autocomplete’ technology you might see when SMS-ing or emailing someone. LLMs learn to predict the next character or word in a paragraph, based on patterns in text harvested from the internet. In the case of ChatGPT, humans also provide feedback to refine what kind of topics, language, and tone the LLM will use when responding to users.
The catch is that ChatGPT doesn’t actually know anything. The answers you get may sound plausible and even authoritative, but they might well be entirely wrong.
Dr Aaron Snoswell, research fellow in AI accountability,at the ARC Centre of Excellence for Automated Decision-Making and Society (ADM+S) says that
“You should trust the responses generated by ChatGPT about as much as you would trust a random stranger on the internet.”
OpenAI acknowledges these pitfalls, some of which are easy to spot and some more subtle. “It’s a mistake to be relying on it for anything important right now,” OpenAI Chief Executive Sam Altman tweeted. “We have lots of work to do on robustness and truthfulness.”
Here’s a look at some of the problems with ChatGPT, and what users should be aware of.
Truthfulness (or lack of)
LLMs are approximate and generative in nature. The results that LLMs provide in response to queries can be factually inaccurate, despite often sounding authoritative in nature. These systems do not calculate but are more like ‘stochastic parrots’; that is, they learn to mimic by example. These systems also frequently make things up – which can be a good (think creative writing), or a very bad (e.g. generating incorrect medical advice).
Exacerbation of Biases
LLMs are trained on text from the internet and as a result they reproduce and exacerbate existing cultural stereotypes and biases. ChatGPT has tendencies toward English language, western, american, male viewpoints with left-wing ideologies. As a result LLMs have the potential to generate toxic, hateful content, or explicit content. How best to evaluate and mitigate this issue is an ongoing research problem that we at the ADM+S (and many others) are working on.
Privacy and Data
Systems like ChatGPT cost a lot of money to build, and a lot of money to run. The value for OpenAI, Google, Meta, and Microsoft comes from collecting user data entered into LLMs.
Remember that if you are not paying for a service, you are not the customer (in fact, you’re probably part of the product!) Technology companies may have terms of service that allow them to collect any and all information discussed with a dialog agent to further improve that service.
Equity and Accessibility
Training data for LLMs is scraped from the internet with no regard for copyright. Furthermore, systems like ChatGPT are typically built on the back of crowds of ‘data enrichment workers’ in global majority countries who work in substandard and precarious conditions. Companies will also charge higher costs for access to better LLM tools, creating a disparity of access for users with less resources.
Misuse and Malicious use
LLMs can and will be used maliciously to generate mis- and disinformation and propaganda. These systems are good at generating convincing spam, phishing, and hacking messages and have the ability to find vulnerabilities in existing open-source software and to create code for computer viruses. LLM-based tools also have unique software vulnerabilities – for instance, hackers can hijack a language model’s output through a process known as prompt injection, even using this to steal personal user data.
While the use of ChatGPT raises legitimate concerns about misinformation, plagiarism, copyright and creativity, we should also consider the value of human effort and what it means to communicate as a person.
AI technologies such as ChatGPT are changing the fundamental nature of how we communicate and marginalising the human element out of these interactions. In addition, it is changing the nature and function of the humans that are doing the communication. To quote Joanna Bryson, Professor of Ethics and Technology, from their essay One Day, AI Will Seem as Human as Anyone. What Then?
“Living in a world with technology that mimics humans requires that we get very clear on who it is that we are as humans.” Joanna Bryson.
Dr Aaron Snoswell presented ChatGPT: Hype or the Next Big Thing? At the Hacks/Hackers Brisbane event held 22 March 2023.
Aaron is a computer scientist and research fellow in AI accountability at the ADM+S Centre, based at QUT in Brisbane, Australia. His PhD “Modelling and explaining behaviour with Inverse Reinforcement Learning” was awarded in 2022 from The University of Queensland, and developed new theory and algorithms for Inverse Reinforcement Learning in the maximum conditional entropy and multiple intent settings.
Aaron’s ongoing research is in the development of socio-technical interventions for reducing toxicity in the foundation model machine learning paradigm, looking in particular at the ways sexism and misogyny manifest in large language models. Prior to academia, Aaron worked in industry as a cross-disciplinary mechatronic engineer in doing medical device research and development, pilot and astronaut training, robotics, and software engineering.