2022 was a watershed year for AI with the release of two key products: DALL-E 2 image generation and ChatGPT. 2023 looks set to continue that trend.
ChatGPT (Generative Pre-trained Transformer), a chatbot from OpenAI burst onto the scene with fanfare in November 2022 and brought in one million users within just five days1. The hype machine ran white-hot with people on Twitter and Reddit hailing it is revolutionary for its ability to generate detailed human-like responses for almost any topic or question. I resisted the temptation to ask ChatGPT to write an article extolling its benefits so this is a human perspective on the where I think it can add value, potential business ideas using it, and what the limitations are to its capabilities.
What is ChatGPT?
ChatGPT is a natural language processing tool based on neural networks in GPT 3.5 from OpenAI. It facilitates human-like answers and interactions in question or conversational style. Usage is currently open to the public free of charge whilst ChatGPT is in a beta phase, however the service is currently under heavy utilisation and being rate limited, presumably to contain operational costs and show some mechanical sympathy to infrastructure behind the scenes.
One of the key features of ChatGPT is its ability to generate responses that are consistent in the thread of a conversation.
Users interact with the service using a web browser; however, it will eventually be accessible via APIs to integrate with other applications and accessible as ChatGPT Professional which include no unavailability windows, no throttling, and an “unlimited” number of messages with ChatGPT.
Figure 1 - ChatGPT Interface
How does ChatGPT work?
Large Language Models, such as GPT-3, are neural networks that are trained on vast amounts of text from the internet. The creators of ChatGPT have taken this a step further in GPT-3.5 by also using Reinforcement Learning from Human Feedback (RLHF) to fine-tune and produce output which is more consistent with human expectations. OpenAI stated2 “We trained an initial model using supervised fine-tuning: human AI trainers provided conversations in which they played both sides—the user and an AI assistant. We gave the trainers access to model-written suggestions to help them compose their responses.”
Users also provide feedback that can be used to further improve the model by upvoting or downvoting the responses they receive from ChatGPT.
The models were trained in collaboration with Microsoft on their Azure supercomputing infrastructure and the service used by the public is hosted on Azure.
Figure 2 - Model training using Reinforcement Learning from Human Feedback (RLHF)
ChatGPT and new business ideas
The language model in ChatGPT can assist with tasks such as composing emails, writing essays, and developing code to automate mundane daily tasks and enable faster decision making. Some product ideas that may emerge using it:
- Intelligent customer support – Online help lacks the ability to be truly adaptive to customer support requests. Anything complicated often requires escalation to a human resulting in slow responses, an inability to resolve an issue, or lost sales. ChatGPT could provide more targeted support, freeing up support staff to focus on complex or unique activities.
- Workflow optimisation – Straight-through processing is the holy grail in banking and finance. ChatGPT could be used to optimise workflows and decision making in mortgage application processing or customer onboarding (for example, ask ChatGPT “what is missing from this mortgage application”) to generate dynamic input forms and automate responses to customers at each step of the process. It could also make decisions on workflow if key criteria are met.
- Content generation – Generating quality content for marketing purposes is time-consuming. Content also needs to be delivered at a recurring frequency to ensure that it is relevant, contemporaneous, and audiences are engaged. ChatGPT could generate ideas for content and marketing copy which forms the base of articles that are reviewed or refined by a human to ensure it is complete and on-brand.
- Customer engagement – Websites and social media lack the ability to be completely dynamic and typically use static content or content that is pre-generated and served using limited personalisation. ChatGPT could be used to deliver hyper-personalised content, generated on the fly based on how a user found the site (i.e., the search terms used) or through their path navigating around website or app as pages are rendered.
- Software application generation – Application development requires a considerable amount of boilerplate code and setup to ensure the right foundations. The no-code and low-code movement has tried to reduce this friction, but the tools still require familiarisation and mastery to be used effectively. ChatGPT could be used to provide application code generation through voice or text input. For example, “generate an iOS application using Swift. It should use tab navigation and have the following views: login, transaction history, transaction details, user preferences, and chat. The application must also provide a dark mode”. Other features could be prompted for, based on the input provided at each stage. Experimentation and innovation could happen at a faster pace and lower cost.
Limitations and the dark side of ChatGPT
ChatGPT currently has limitations which make it unsuitable for certain use cases.
- Timeliness of information - The models used are trained on past information and are not completely up to date. This means ChatGPT is unlikely to be able to create a meaningful recommendation on anything new or novel since the latest trained model and it cannot currently advise on breaking news.
- Cognitive and political bias – There are cautions from researchers against “problematic biases, such as disproportionate false positives when discussing groups that are frequently the target of hate,” and “counterfactual bias towards certain demographic attributes.” indicating that humans involved in training it may not have been diverse in background and demographics.
- Weak positioning - When asked to make a recommendation between two options, it sometimes fails to pick a side, defaulting along the lines of “it is a matter of preference, […] it is hard to say what is better”. This is of limited use for people who need to decide quickly.
There are several areas where the use of ChatGPT is also causing concern.
- Copyright – ChatGPT has been trained on content taken from the internet, much of it without explicit permission from the authors. The technology can be used create derivative works without compensating or citing the original creator. Serious questions are also being asked about who should be considered the author of content generated by ChatGPT for copyright purposes, and whether such content should be entitled to the same IP protections as human creators.
- Education – Academics have cited concern that ChatGPT is able to create intelligent sounding responses to user prompts including homework assignments and exam-style questions. Some have also stated that responses generated by it would result in pass marks. Given the sophistication of content produced by ChatGPT, it will take equally as intelligent technology to determine whether academic work was computer generated.
- Scientific research – According to an article in nature.com3, researchers cannot always differentiate between AI-generated and original abstracts.
- Harmful and nefarious use – ChatGPT attempts to refuse answering inappropriate requests, however it will sometimes respond to harmful instructions. Content moderation is being used to block certain types of unsafe content although this is not fail-safe.
- Impact to employment and learning – some have suggested that ChatGPT will affect demand for knowledge workers4 and that over reliance on it will lead to a decline in critical thinking and problem-solving skills.
The boundaries of ethical and acceptable use of AI to assist with academic and scientific writing are yet to be determined. Critical to the acceptance of the technology in a scientific context, is accuracy of information.
Concluding thoughts
OpenAI was founded in 2015 as a non-profit research organisation with the overarching goal of achieving safe and socially beneficial artificial intelligence, however, the costs involved and pressure to turn a profit on AI products means that significant investment from outsiders and commercialisation is required to unleash further potential in ChatGPT.
On 23 January, Microsoft confirmed it is making a multibillion-dollar investment into OpenAI with CEO Sataya Nadella stating5 “we formed our partnership with OpenAI around a shared ambition to responsibly advance cutting-edge AI research and democratize AI as a new technology platform” and “developers and organizations across industries will have access to the best AI infrastructure, models, and toolchain with Azure to build and run their applications”.
Despite the recent challenges faced by the technology sector, generative AI as reflected by OpenAI’s ChatGPT that Microsoft will incorporate into Office, Bing, and its cloud service Azure, is pointing to the road ahead.
It will be interesting to see the productivity boosts and new business opportunities that emerge from the use of ChatGPT and the next evolution of this technology, GPT-4.
1 https://twitter.com/gdb/status/1599683104142430208
2 https://openai.com/blog/chatgpt/
3 https://www.nature.com/articles/d41586-023-00056-7
4 https://www.nytimes.com/2022/12/06/opinion/chatgpt-ai-skilled-jobs-automation.html
5 https://blogs.microsoft.com/blog/2023/01/23/microsoftandopenaiextendpartnership/