Professional Documents
Culture Documents
OPEN ACCESS
Opinion
This new conversational AI model can be your friend,
philosopher, and guide ... and even your worst enemy
Joyjit Chatterjee1,* and Nina Dethlefs1
1University of Hull, School of Computer Science, Cottingham Road, Hull, HU6 7RX, UK
*Correspondence: j.chatterjee@hull.ac.uk
https://doi.org/10.1016/j.patter.2022.100676
We explore the recently released ChatGPT model, one of the most powerful conversational AI models that
has ever been developed. This opinion provides a perspective on its strengths and weaknesses and a call
to action for the AI community (including academic researchers and industry) to work together on preventing
potential misuse of such powerful AI models in our everyday lives.
Introduction model builds on the GPT-3 model and robust. This allows ChatGPT to combine
Every other day, we hear of powerful new adds a supervised fine-tuning component the qualities of its predecessor GPT (i.e.,
artificial intelligence (AI) models being to it, which makes it learn from human being trained on vast amounts of infor-
released by tech companies. Such feedback and provides it with a metric mation available on the internet)
models are the brain behind multiple for validation. with the ability to hold human-like
products that assist us in our everyday ChatGPT is a conversational AI model conversations.
lives, including robot vacuum cleaners (a chatbot based on natural language
mopping our homes, voice assistants processing and deep learning) that was Our experimental interactions with
like Alexa powering smart homes, smart built as a sibling to InstructGPT,3 a the ChatGPT
wearables tracking our fitness, etc. On less-known model that could deliver re- Like millions of AI enthusiasts around the
November 30, 2022, OpenAI (the famous sponses to simple questions like globe who are currently trying the
AI research lab originally founded by ‘‘explain evolution to a 6 year old’’ in a ChatGPT, we set out to conduct some ex-
Elon Musk, Sam Altman, and others) human-like manner. But what sets the periments with the ChatGPT that blew our
released a new AI model, ChatGPT,1 and ChatGPT apart from any other model minds. This new chatbot was multital-
made it available to the public for free for ever released is its ability to continually ented. First, we asked ChatGPT to
a limited period of research preview. interact with a human naturally and pro- generate programming code for an AI
OpenAI releases new AI models vide stimuli/inputs to the user as well as model (AI generating AI!) for automatically
frequently, and its notable previous inno- ask interesting questions back. This classifying a dataset. Not only did the
vation has been the GPT-3 (Generative makes it possible to have long (and model do this flawlessly, it also asked us
Pre-trained Transformer-3),2 a powerful potentially never-ending) conversations valuable questions on the type of the da-
language model released in 2020 based with the chatbot until the interaction taset (e.g., number of features, context,
on deep learning that is the largest neural eventually dries out. The model admits etc.)—ones that you might expect from a
network ever produced with 175 billion its mistakes, challenges incorrect pre- real-world data scientist. However, it
parameters. The GPT-3 model has mises from the user, and even tries to forgot to provide us with information on
already seen increasing usage in applica- reject inappropriate requests. The devel- pre-processing the original dataset,
tions such as translating language, opers of ChatGPT created the model by which is an important part of AI model
answering questions, and even gener- fine-tuning the previous GPT-3 with a development, and we had to ask it a
ating memes. A major limitation of GPT- large amount of real-world data obtained follow-up question for accomplishing
3 was that it was trained as a fully unsu- from humans (wherein, the annotators this, which it eventually accurately deliv-
pervised model, generating content that played both sides: the end user and the ered on. We even asked the model to
it learned out of vast amounts of informa- AI chatbot). To ensure that the model is automatically write us a new Hindi song,
tion on the internet without any validation not biased and can ask follow-up ques- perform a text-to-speech conversion,
on it, making it provide often uncanny and tions to the users if it is not confident of and melodize the generated speech into
funny responses. The recent release of its responses, the developers also used a song—the model delivered surprisingly
ChatGPT sparked a wave of interest in reinforcement learning from human feed- well here and could generate code to
AI that has never been seen before—this back to optimize the model, wherein hu- automatically convert Hindi text to musi-
model can not only deliver what the man evaluators provided the model with cally sounding songs with a synthesizer.
GPT-3 could do as a machine learning a positive score (reward) when it gener- Refer to Chatterjee4 for the code genera-
model but also interact in a humanly ated realistic responses or a negative tion examples. Whenever the ChatGPT
manner, giving any conversation a sense score (penalty) when it generated un- seems to be confused or less confident
of intelligence, humor, creativity, and canny outputs, based on which the in its responses, it asks the user follow-
emotion. Interestingly, the ChatGPT model was refined and made more up questions before arriving at its final