Made available online to the public at the end of November: ChatGPT, a new version of an AI application that can produce text. The solution has attracted a lot of attention in the media in recent weeks. Can Google, copywriters and customer contact employees go home?
TechUpdate is Ziptone's section on emerging technology relevant to customer engagement. We briefly explain what it is, how it works, why it is relevant, what the pitfalls are and we ask the reader's opinion: Hot or Not?
What is it
ChatGPT is a 'chatbot' that can produce different types of texts in different ways. ChatGPT is an example of so-called generative AI: a form of artificial intelligence that can independently produce new original content (image, sound, text).
ChatGPT was developed by OpenAI. This is a non-profit enterprise in which parties such as Microsoft, Elon Musk, Peter Thiel, Amazon Web Services (AWS), Infosys have invested more than 1 billion dollars. The company also develops other AI applications. Several successive versions of ChatGPT have been launched in recent years, the latest version being ChatGPT4. This version has been temporarily made available online and there is also one Dutch version.
How it works
ChatGPT can draw up structured and coherent texts based on entered commands (such as a question). You can choose from different types of text, such as a pitch or an article. ChatGPT is based on extensive language models and works based on pattern recognition and predictions of the next word in a sentence. Both mechanisms use large amounts of text from the Internet.
It is therefore not an 'intelligent' or creative system, but a system that produces a grammatically correct and substantively structured text based on probability. During development, the system was further refined based on feedback from end users.
According to the developers, the underlying model is aimed at being able to conduct a dialogue. The tool should be able to ask follow-up questions, admit its own mistakes, recognize and correct inaccuracies. The developers themselves explicitly state that the tool is not yet perfect, makes mistakes and suffers from bias. However, given OpenAI's funding, there's still a lot to be expected from this application.
Why is ChatGPT relevant
When sufficient attention is paid to specific training, this solution can eventually become part of customer service processes. This training should include reference texts about a company or about specific processes or products. The question is whether this will also be possible in a commercial version. In addition, checks are necessary to prevent undesirable content (incorrect or with a bias) (see below under 'pay attention').
For example, coherent and well-structured texts can serve as the basis for a voicebot, which opens the way for the use of automated written or spoken responses in a dialogue. Such an application can also help generate answers in self-service environments.
The power of ChatGPT is currently, however contradictory, not being able to generate (new) texts. But it is in the enormous processing capacity of resources (the computing power) that ensures that the system places words one after the other that can be said with great certainty that they are a logical addition. The system pays attention to word choice, sentence structure, tone of voice and structure of texts, but literally bases itself on existing text material.
Just pay attention
1. When training ChatGPT, copyrighted source material has also been used. This raises discussions about copyright violations and the risk of plagiarism. This is a particular objection because no sources are mentioned, unless you literally ask for this in your assignment. In addition, it is not yet clear where the copyrights lie when AI produces content.
Sam Altman, OpenAI CEO, about ChatGPT (Source: Twitter)
2. One of the pitfalls for AI chatbots is presenting misinformation as fact. ChatGPT is said to be quite capable of self-censorship. For example, the system can refuse unethical questions and the system would warn when requests ('How Columbus discovered America in 2025') lead to a fictitious text. However, experts have now found out that these qualities in the current version are far from guaranteed and that there is considerable bias. As mentioned, these shortcomings are also explicitly mentioned by the developers and the CEO of OpenAI themselves. ChatGPT is unable to understand texts, but reacts and imitates based on probability calculation.
3. The more specific the assignment to compose a certain text (with specific elements), the greater the chance that ChatGPT fictional paragraphs in the text. The end result is a text that may be factually incorrect or misleading in whole or in part. The fact that it is not clear how a text is created is also problematic. Source references are not used so that the reader can better assess the content.
4. The above objections mean that the solution is currently unsuitable for automated applications. The texts that ChatGPT produces must be verified by humans if you want to use them for formal communication. For Stack Overflow (an online platform where software developers exchange questions and answers) this was the reason to ban the use of the tool within their own community.
Loading...In January 2023 we will take a closer look at the opportunities that a solution like ChatGPT can offer in a customer service setting.
