This article was originally published in GovInsider on 23 February 2023.
Within just a few weeks, the capabilities of ChatGPT astounded the world. But it may not change the world as dramatically as the Internet, or even the iPhone did, writes William Webb, Chief Technology Officer of Access Partnership.
At first glance, ChatGPT appears almost scarily capable. It generally returns sensible text in a well written form. It can pass exams. Initially, it is hard to tell its output from that of a human.
But having looked at about a dozen more open-ended searches, I think I can tell. It is the lack of insight and the lack of opinion that tends to set it apart, and the lack of nuance of its response. As one commentator said, “darned thing refused to state a thesis and support it. It couldn’t analyse or take a stand.”
Take, for example, the issue as to whether chatbots contravene intellectual property rights by copying and repurposing material written by others. When asked about this ChatGPT responded:
It depends on the specific implementation and purpose of the chatbot. Generally speaking, if a chatbot is designed to engage in activities that would infringe on intellectual property rights (IPR), such as distributing copyrighted materials without permission or selling counterfeit products, then the chatbot could be considered to contravene IPR. However, if the chatbot is designed for legitimate purposes and does not engage in activities that infringe on IPR, then it would not contravene IPR. Ultimately, it is up to the creators and users of chatbots to ensure that they are complying with relevant laws and regulations related to IPR.
This seems to me a classic response from a chatbot – “on the one hand…on the other”. And it says pretty much nothing that’s not obvious. Who would want to read an article composed entirely of such blandness? Far better would have been links to a range of opinion pieces from leading thinkers.
How ChatGPT really works
Let’s step back and ask ourselves what is going on here. A chatbot works by ingesting a “large language model” (LLM) – basically a vast amount of text lifted from the Internet and other sources. It then looks for patterns. It might look for what word most commonly follows “twinkle twinkle” and then when it sees those input words it responds with “star”. It also looks at patterns of words and uses the most frequent patterns to form sentences. It is nothing more than a next word prediction system based on what seems most likely given the data it is trained on.
What is astonishing is that it can come up with such human like responses once the data set gets sufficiently large. But let’s be clear, it is not reasoning, it does not “understand” the question the way humans do. What it is effectively saying in response to a query is “here is the pattern of words in my vast LLM that occurs most often in relation to the words in your enquiry”.
Done well, a response is the wisdom of the crowds – the averaged answer from thousands or millions of us to related questions in the past. But crowds do not generate insights, new ideas or anything or much value.
ChatGPT vs chatbots
The current rush to integrate chatbots into search engines makes some sense. Effectively, a chatbot removes the intermediate stage when a search engine returns links and a user selects the most attractive one. For some users, that is a benefit. For others, it takes away one’s ability to filter information based on its source, date of publication and other factors.
And it isn’t new – smart speakers have been delivering a direct answer for many years. A search engine that delivers both a chatbot simple response and a set of links might be the best of both worlds.
Much of the current interest in ChatGPT probably stems from the fact that there appears to have been a step-change in capability. A couple of years ago, computer-generated prose was awful. Now it is grammatically sound and superficially human-like.
That has surprised many, including me. It suddenly raises the questions of (1) is this new ability transformational in any sphere and (2) will chatbots continue to develop at this amazing speed, and if so, into what?
Chatbots, AI and human interaction
Where chatbots are useful is in humanising interactions with machines. Where we want to do something relatively routine, like book a hotel, a chatbot can make this much simpler than the current approach of navigating travel comparison websites and clicking through to make a booking. That can improve the experience of “self-service” interaction and might enable automation in other areas currently just too awkward to automate. That would advance existing trends of the last 20 years or more.
Chatbots might also be useful in delivering pro-forma template material that can be modified. That is clearly true in computer programming where they can produce (or more likely copy) code for tasks that are common across multiple programmes, such as gathering user input. It might also help when working on something like a retirement speech where it can generate a structure with common messages as the basis for subsequent individualisation.
Just don’t actually give the chatbot-generated speech verbatim if you want guests to remember what you said.
Might chatbots continue their rapid advancement into something better that does have an opinion – like something that could write this article? Many commentators are sceptical.
As mentioned above, chatbots train themselves on large language sets and take material from things written by humans. As a result, they struggle to be anything more than backwards-looking. They can summarise what has been said by others, and repurpose it in apparently new ways, but they cannot generate new ideas.
Public sector implications
Stories of inappropriate, offensive or damaging responses from chatbots are already flooding the media. While those designing chatbots will undoubtedly resolve some of these issues, the way that chatbots work makes it hard to, for example, stop offensive responses.
The result will likely be calls for regulation that are even more strident than those that call for regulating social media. It will be very challenging for regulators to work out how to enable innovation by allowing chatbots to evolve while protecting consumers from harm as soon as possible – as will be the need for global rather than local responses. Working out how to balance the enabling of innovation by allowing chatbots to evolve with protecting consumers from harm as soon as possible will be very challenging for regulators, as will be the need for global rather than national responses.
On the other hand, chatbots might be useful for governments. They might enable much greater automation of online interaction, making digital self-service more fulfilling and effective. They might be able to help draft government output, providing first passes of much material, which is often of sufficiently similar format and content to prior material that their wisdom of the crowds approach can be useful.
The challenge here will be to convince citizens that the use of chatbots is legitimate and helpful, and not governments abdicating responsibility.
Challenges and limitations
Chatbots effectively give us the wisdom of the crowd – often right but rarely insightful.
There are other barriers. Chatbots are running out of training materials, and they are consuming ever-more computing resources. These factors will be a drag on their future evolution. And there may be backlash against copyright, against inaccuracy and against material that looks true but for which the genesis is dubious.
There are a few moments in the history of technology when it is clear that the world has changed. The introduction of the iPhone was one. But perhaps the biggest transformation of modern times – the Internet – grew unnoticed by most for decades and it only became clear over many years that it could be transformational.
ChatGPT looks initially like an “iPhone moment” as it surges out of nowhere. But that speed of arrival makes it easy to over-predict its influence. It is an amazing achievement that will enhance our interaction with machines. It is here to stay. But I doubt it will materially change the world.
And I hope that it won’t change it for the worse rather than the better.