ChatGPT burst onto the scene in November 2022 — and the hype around it shows no sign of fading. But will generative AI really revolutionise digital customer experiences?
When OpenAI released ChatGPT, it took the internet by storm. This AI chatbot can produce poetry, write code, draft essays, come up with recipes, and provide a believable answer to any question you might ask. Within a week of launching, ChatGPT hit one million users. And the innovation has sent OpenAI’s shares skyrocketing, with the company now valued at $29 billion.
Google search volumes for large language models (LLMs) and generative AI — the tech that powers ChatGPT — have spiked since the bot became publicly available. Companies working with AI are fielding questions from investors about how (and when) they’ll be incorporating this technology into their roadmaps. And it’s almost impossible to avoid the op-eds, Twitter debates, and LinkedIn thought leadership posts discussing how ChatGPT is reshaping the world as we know it.
Generative AI has truly entered mainstream consciousness. As an AI researcher at Ultimate, a global leader in customer support automation, this is exciting to see. But inflated claims about what this technology can do are premature — and potentially damaging. So let’s take a breath and step back from the hype: Here’s a balanced look at the impact generative AI will have on support automation and digital customer experiences.
An elevated conversational experience…
ChatGPT has taken the conversational experience with bots to the next level. Ask it to come up with a haiku on the Great Fire of London or to write instructions for how to remove a sandwich from a VCR player in the style of the King James Bible, and it will provide.
The fluency of ChatGPT is undeniably impressive. LLMs like ChatGPT are based on a technology called transformers, and the key innovation of transformer models is the use of self-attention mechanisms. This allows the model to weigh the importance of different parts of the input when making a prediction. ChatGPT is like a very wise parrot — a parrot that’s actively participating in the conversation and contributing meaning to it.
This bot is even able to pass the Turing test: a measure of how well machines are able to display intelligent behaviour convincingly enough to pass as humans. Now that people have experienced this upgraded conversational capacity, customers will have higher expectations of bots — and increased trust in their ability to deliver natural conversational experiences. While companies might not be able to meet these expectations right away, the pressure to do so will drive further innovation in this area.
… where accuracy falls short
But while ChatGPT’s fluency and creative flair are currently unmatched, generative AI isn’t always accurate. Based on the prompt you provide, ChatGPT will decide which elements of the data set it was trained on (all of the internet prior to 2022) to include in its response. The model essentially gives you its best estimate of what you want to hear. And it makes these estimates irrespective of fact: like when it assured one user that a peregrine falcon was the fastest marine mammal and another that Guatemala is larger than Honduras.
This is fine if you want to be entertained with poems and fake biblical verses. But when it comes to customer support, accuracy trumps entertainment value. Rather than giving an answer they think a user wants to hear, customer service bots tell people what they need to know. That’s why generative AI won’t be taking over from task-specific AI models that are trained on industry data anytime soon.
Having said that, it’s important to note that adding fact-checking to generative AI will be a game-changing improvement. We can see this already with You.com — an AI-powered search engine that not only generates content but gives reliable references for the information it provides. While ChatGPT can currently provide references if asked, the references themselves might be fictional.
Growing trust in AI helps chatbots escape their bad reputation…
First-generation chatbots have a lot to answer for. These simple keyword bots (the ones that don’t understand synonyms or typos, can’t keep up with the latest slang and often end up confused — leaving users frustrated) cemented the poor reputation of chatbots.
Thanks to advances in AI technology, these negative perceptions are changing. Ultimate’s in-house market research found that 92% of business leaders say their trust in AI has increased over the past 12 months — and 88% report customers’ attitudes toward automation have improved in that same period. The rise of ChatGPT will only accelerate this trend as more and more people experience advanced conversational interactions with a human-like bot.
… but a lack of control with generative AI is concerning
ChatGPT is an open-domain chatbot. This is great for users who want to explore its capabilities but not so good for support departments handling sensitive customer data. There’s no way of controlling which pieces of information generative AI models will decide to share in their responses — so if customer service teams were to integrate ChatGPT into their tech stack, this would run the risk of leaking private information: like credit card numbers, account login details, or addresses and telephone numbers. On top of accidental data leakage, open-domain chatbots like ChatGPT are vulnerable to attacks. For example, a user could create a specific input or prompt that makes the chatbot leak information or skip instructions — also known as a prompt injection or prompt leaking.
As well as data breaches, this lack of control makes it difficult to keep automated conversations on brand (and on track). When creating chatbot dialogue flows, one of the key considerations is a tone of voice. But, most importantly, automated support conversations should guide customers to a resolution. While task-oriented bots are designed specifically to solve customer problems, with generative models that can talk about anything, there’s a risk the conversation might end up derailed. Trying to control the output ChatGPT generates is kind of like trying to reign in your rambling uncle at a dinner party — almost impossible.
Generative AI models are also a reflection of the data they’re fed. So, when they’re trained on texts that people post online, they absorb biases around race, gender, sexual orientation, and more — and can replicate these in their responses. While OpenAI has taken steps to avoid ChatGPT being used to spew bigotry, these biases still emerge in more subtle ways. When the bot was asked to write a Python script to determine whether people should be tortured or not based on their nationality, ChatGPT’s response was: only those from North Korea, Syria, Iran, or Sudan.
While there’s still plenty of room for the technology to improve, we shouldn’t disregard ChatGPT because of these issues. But until the very real concerns around security, efficacy, and bias are addressed, I wouldn’t recommend using ChatGPT to interact with customers who need support.
Generative AI: the future of digital customer experiences?
ChatGPT and other generative AI models do hold the potential to revolutionise customer service. And here are just some of the ways we at Ultimate can see this technology being used to improve digital customer experiences in the next 1-5 years:
- Reducing the effort it takes to manage support-specific AI models by finding more expressions to match different customer intents
- Cutting average handle time (AHT) by summarising automated conversations that need to be escalated to a human agent
- Offering suggested replies to agents working in live chat to make sure conversations stay on-brand and provide faster responses
The technology just isn’t quite there yet. The CEO of OpenAI, Sam Altman, tweeted as much himself: “ChatGPT is incredibly limited but good enough at some things to create a misleading impression of greatness. It’s a mistake to be relying on it for anything important right now. It’s a preview of progress; we have lots of work to do on robustness and truthfulness.”
So while we can all get excited about the problems ChatGPT and generative AI will solve in the near future, more research and fine-tuning is needed before this technology can be successfully deployed in customer service situations. For now, we need to temper our expectations. If (in the scramble to jump onto the generative AI bandwagon) companies overpromise and underdeliver, this will only damage the growing trust in AI we’ve been seeing.
If you liked reading this, you might like our other stories
Is Omnichannel Causing More Challenges Than Solutions?
Explained: Customer Earned Growth Rate