ChatGPT

Kenton Reynolds, Writer
@uxconnections

Over the last few months, an incredibly intelligent AI called ChatGPT has taken the internet by storm. What will this open chatbot mean for the future?

All you need is an email and password to create an account at www.chat.openai.com/chat in order to access a seemingly endless supply of knowledge. The chatbot is happy to answer questions on pretty much any topic (I certainly couldn’t get anything past it!) with surprising eloquence in its replies. As well as this, it can be used for conversation too. Showing the extent of its helpfulness, here are a few ideas it gave me on what to include in an article about itself.

The training method to create such an advanced chatbot is important in distinguishing it from other competitors. OpenAI have used a combination of Supervised and Reinforced Learning with the latter carrying the most weight in its success. The developers use human feedback to reinforce learning which is important to minimise errors such as discriminatory, biased, incorrect or otherwise harmful statements. Reinforcement Learning from Human Feedback (RLHF) is carried out in three steps. First, the Supervised Fine Tuning Step uses a baseline model to create a fairly small demonstration data set. Then a set of human participants vote on preferred responses from a relatively large number of outputs Fine Tuning Step outputs. These preferred responses are then used to retrain and improve the system which can then be plugged into step 1 by generating a series of outputs. Such intense cyclical fine tuning has created ChatGPT with an almost personal feel, it does not feel as cold or impersonal Siri, Alexa or other online chatbots.

The platform is open to anyone with no payment or download required. You just need to create an account with OpenAI on their website. From there you are free to enter any questions or queries which it will answer (almost) flawlessly. If you do not like the response it gives you then you can press regenerate or ask the question again and will be met by a new, unique answer.

Such powerful technology being open to anyone has raised a few ethical concerns. Primarily is the obvious problem that new technology seems to be ceaselessly greeted by – unemployment. The apparent intelligence of the system is such that it is beginning to put many jobs at risk if it continues to be developed and becomes more reliable. Usually, we talk about technology threatening low-paid, unskilled labour however, with ChatGPT, the conversation seems to be revolving around more intellectually-centric roles including researchers, coders, writers and more. Although there seems to be an overarching sense of doom mongering in his speech, Jordan Peterson spoke on ChatGPT predicting that it will make universities obsolete in the next five years. He recalls that the system was able to create an extra chapter for his book in “about 3 seconds” and it wasn’t obvious to Peterson that he did not write it himself. There is also a mention of an engineer at Twitter who asks ChatGPT to list what he should do in a typical day at his work and to write the accompanying code for the tasks it had given which it performed successfully with working code.

Beyond this, the body of text and content that the AI is trained on raises concerns as if there are biases and inconsistencies that could lead to problems down the line for the system. Furthermore, the training data is ChatGPT’s most significant limitation. With the most recent components of its corpus coming from 2021, any questions on recent events cannot be answered and any new developments and findings cannot be kept up with. Though working to base the system on real-time information is in process, it is significantly more difficult to create.

In summary, ChatGPT is still to be taken with a pinch of salt. It is not complete and, although it becomes more intelligent constantly, the corpus of text can not be relied on to always produce a perfect answer to questions. Academic use is a concern – my exams at University have “Do not use AI resources such as ChatGPT” on the cover page now – and AI testing is creeping into common use in the same way as checking for plagiarism so it may not be the solution to homework just yet! The takeup on the platform has been enormous with it being regularly inactive recently due to being at capacity with requests. This is a fantastic tool if used in the right way as a supplementary resource and not something to base your work on.