23/03/2023

Have you heard of ChatGPT? This is almost a rhetorical question, because who hasn’t heard of the much-hyped text generator? The hype around the language assistant and the underlying language model is huge. People are talking mostly about its benefits and potential applications, but sometimes also about possible misinformation and misuse of the technology to cheat. We take a look at the details that need to be considered in order to use it appropriately.

After the US provider OpenAI released access to the system as a demo version at the end of November 2022, it has made an unprecedented start. After only five days, ChatGPT had more than 1 million users, a figure that took Facebook ten months and Instagram two months to reach. Currently, about 100 million people per month use the service – an unprecedented level of hype in recent digital history.

“Chat Generative Pretrained Transformer” is a language model (Large LanguageModel/LLM), which is based on artificial intelligence or a deep neural network. It has been trained with enormous amounts of data to mimic human language. Thanks to its chat interface, it enables direct interaction. It understands text input, can respond to it in a way that seems natural, and delivers text responses that sound as if they come from humans. The system can answer questions or carry out instructions, for example summarise texts, put content into poem form or provide programming code for specific tasks.

This sets ChatGPT apart from previous AI-based language models, which did not deliver what they promised. In 2016, Microsoft took its Tay chatbot offline after just 24 hours, while Facebook’s Galactica lasted three days last year. Both quickly went off the rails, generating racist or meaningless statements.

ChatGPT turns previous text production on its head and gives an impressive boost to potential AI-based applications. Quite a few text producers feel that their existence is threatened and competitors are now under pressure.
And ChatGPT is evolving rapidly, most recently with the integration of GPT-4, a data model trained with multiples of its predecessor GPT-3.5 and capable of handling image-based input in addition to words. This delivery also marks a watershed: while previous language models were mainly used in research or were integrated into third-party applications, ChatGPT’s chat interface allows for easy interaction and everyday use. No matter whether you’re a student, editor, speechwriter or programmer – anyone can access the system after registering with OpenAI.

It is precisely this ease of delivery, coupled with linguistically very good results, that the translation industry has been familiar with for many years from neural machine translation (NMT). Here, too, there was real hype when providers such as DeepL made the service available free of charge via a web interface and delivered fluent results. And just as with NMT, the question as to whether the systems will soon make humans superfluous is now being asked everywhere with regard to language models.

At first glance, the results of the AI-based language model are impressive. The subtleties become apparent when you take a second look. ChatGPT generates very simple texts that are put together from the input provided and the information found in the training material. The question is usually repeated, and unusually often, the subjunctive is used – indicating that a suggestion rather than facts is being offered. The responses may therefore simply be wrong and, moreover, cannot always be reproduced because the system has no fixed answers to questions but calculates them on the basis of statistical probabilities.

The choice of words is simple, which often doesn’t matter in short texts. With longer texts and more complex topics, however, the clumsy sentence structure is noticeable. If you make more than one request, not only will you see the style repeated, but you will also quickly recognise underlying patterns. On the whole, the results soon repeat themselves.

And while it is certainly fun and useful for private use – birthdays, humorous speeches, Valentine’s cards – to have a text reworded as a poem or in the style of someone famous, these usage scenarios tend not to be of interest in a business context.

To move from hype to productive use, you have to know the limits of the system. First and foremost: ChatGPT does not provide information in real time, but based on the 2021 training material. The system itself points out that it cannot provide up-to-date information, but a two-year gap in information may seriously distort the results in some areas.

Asked which countries lie between France and Portugal, ChatGPT replied in February 2023 that these countries are direct neighbours. Asked about the length of their shared border, the answer was that no common border exists.

The example clearly shows that the factual accuracy of the output probably represents the biggest risk of use. This is because the language model is data-centric AI that draws on an enormous amount of data and texts from all accessible sources on the Internet and from all this generates an answer to questions posed. It is completely unclear where the answer ultimately comes from and whether several sources were combined to produce it. The machine’s answers may therefore – as can be easily seen in the above example relating to world knowledge – never be taken as factually correct, but always only used as information that must be checked and verified. This phenomenon is also well known from machine translation: NMT also adds content or omits words or parts of sentences and does not provide reliable results.

Just as with machine translation, ChatGPT is a tool that responds to input and delivers output. And for both tools, it can be said that artificial intelligence can deliver good results and, used wisely, can increase productivity in text creation and translation.
However, what is crucial in both areas is what happens to the output, whether and how it is verified and for what tasks the tools are used at the end of the day. There are helpful tips for both steps – input and output.

ChatGPT does not respond to complete questions, but to prompts, i.e. parts of sentences or keywords that are used as input for an AI tool. It is therefore important to formulate precise and clear prompts that the system understands:

• Do not ask open questions
• Ask specific questions and provide context to the question
• Use the shortest possible sentences
• Choose words that are easy to understand
• Avoid technical jargon or slang

According to the provider, ChatGPT has been trained to anonymise or encrypt sensitive information to ensure confidentiality. But of course the data entered is stored and used to further train the model. As with machine translation, it is therefore important not to enter sensitive information to minimise the risk of unauthorised access or misuse.

“A response from the system should never be seen as a final answer, but always as a suggested text that needs to be checked.”

Probably the most important point is to check the facts in texts generated by ChatGPT to verify that they are correct. Here, too, there is a clear parallel with machine translation: while most users are confident enough to check and correct translations in the English they are familiar with, many will be stuck when it comes to Czech texts, for example. Going back to the example we used above, while most users know that ChatGPT has omitted Spain and Andorra as countries between France and Portugal, the length of the shared border between these countries would certainly be something they would need to research and verify.

Since ChatGPT draws its information from all available sources, it is always possible that the results are wrong or inaccurate. A response from the system should therefore never be seen as a final answer, but always as a suggested text that needs to be checked. Just as a machine translation should never be published without being checked first, and should only be used in raw form to gather information.

ChatGPT can optimise or even automate parts of the text creation process. For this, it is helpful to know how and for what purposes it can be used. A selection:

• Briefings
• Pre-formulations
• FAQs
• SEO texts
• Product descriptions and advertising texts
• Summary of existing texts
• Text sections and modules for social media

In addition, ChatGPT can be a source of inspiration and a brainstorming partner. Asking questions when exploring a topic can lead to surprising answers and aspects that you wouldn’t have thought of on your own.

For all the parallels, it is also worth looking at the use of a language model as a machine translation source – because ChatGPT can also do this if requested to do so. Some classic sources of error – such as incorrect links between sentences – are hardly of any relevance in language data models, for example, since here the entire text is regarded as the context. However, initial tests showed that the results are highly dependent on the prompts used and are currently good to moderate compared to generic systems.
In addition, the aspect of time has a role to play: a translation request to ChatGPT currently takes up to thirty times longer than to DeepL, for example. Nevertheless, integration into translation tools is currently underway, meaning that in the future it will be possible to access ChatGPT directly as a translation source.

Currently, more interesting than pure machine translation are application scenarios in the field of terminology extraction, keyword research and the reformulation of existing translations. Here too, the transition from hype to productive use can already be seen, such that the language model is able to usefully support more and more real and meaningful possible uses.

“For the foreseeable future, artificial intelligence is a proven additional tool and welcome assistant, but not a substitute for human intellect.”

As with machine translation, the hype surrounding ChatGPT is totally understandable: impressive results, it’s very easy to use and makes you feel you can solve previously time-consuming tasks at the touch of a button. However, the fear that the systems could make humans superfluous or replace them – translators in the case of NMT and text creators in the case of language models – quickly resonates. But, when used productively, especially in a professional environment, the limitations quickly become apparent.

For the foreseeable future, artificial intelligence is therefore a proven additional tool and welcome assistant, but not a substitute for human intellect. It will be able to significantly increase productivity in some professions and reduce the effort required for simple and routine tasks. It is not yet a threat to technical editing and content creation. Those who like to write should continue to do so.