The year began under the sign of fascination with ChatGPT, which now allows a wide population to participate in the benefits of artificial intelligence. However, merely asking questions will not yield any real results. It will be most beneficial for people who are experts in a certain field to achieve even better results, and we can only hope that it will not cause an even greater flood of poor and meaningless content on the internet.
As the global world grapples with many very tangible troubles that have not occurred for decades, technology is relentlessly moving forward, and this year began with fascination with ChatGPT, a tool that is currently being ‘played with’ by a large population, from young to old, exploring the capabilities and possibilities of this ‘wonder’. What is it about? GPT-3 (Generative Pre-trained Transformer 3) is the most advanced language model developed by OpenAI. It is a machine learning model trained on a vast amount of textual data from the internet, such as various web pages, books, newspaper and scientific articles, etc., which enables it to generate new text in various styles and formats that is very similar to human writing. The application of this model in the form of the ChatGPT tool allows for easy and simple writing of various texts: journalistic articles, student assignments, theses and dissertations, professional and scientific papers, manuals, textbooks, books, poems, programming code, and many other types of written forms.
—
—
Since the GPT-3 model is ‘trained’ on data available until 2021 and does not have real-time internet access, it cannot provide current answers to questions about topics related to current events, but regardless of that, it is very useful. The usability of such tools will only increase as they develop.
ChatGPT is a tool
It is crucial to ask the right questions with an understanding of what is possible for AI. It is like a version of very abstract programming. If we look at ‘classical’ programming, it largely involves writing precise, pre-known instructions that have pre-known effects on the computer. For example, when we program, we usually need to import some data (packages) and some predefined instructions (libraries), and then write accurate and appropriate commands that we assemble into a logical whole that will perform part of the overall functionality of the code. These commands, in a meaningful combination with other commands, e.g., commands for loops, inputs, outputs, using variables in certain ways, etc., ultimately yield meaningful code that performs some functionality or more for us humans. This is essentially similar to how we should approach using the ChatGPT tool, but at a much higher level of abstraction. With ChatGPT, you do not have commands; instead, using natural language that we all understand, we give instructions as we would to another person to do something for us.
An example would be giving instructions to a children’s book writer to generate content, and we who give the instructions are those who have the idea of that children’s story in our heads, just not knowing how to articulate it nicely or not having the time or needing help from someone who will shape our thoughts nicely and present them for us to decide if that is what we wanted or if something needs to be refined, and we will give that person instructions for that refinement. We will give instructions until we are satisfied with the result. The same applies to ChatGPT, which acts as our assistant.
A great plagiarist
An example of a simple ‘instruction’: ‘Write me a 500-word introduction on the topic of artificial intelligence in the style of a thesis.’ And after that, we can refine the result with additional ‘instructions’ to include more details or to emphasize certain elements more depending on what our goal is.
—
—
Also, if we want to use an existing text that someone else has written in our work, without it being obvious that it is plagiarism, we can give ChatGPT an ‘instruction’ to rephrase the text (again in some style, with some focus, with some variations, intentional mistakes, the form of the sentences themselves, etc.). You can test this yourself, and you will see that it works. Not only can it ‘fool’ tools that check how much of a text is plagiarized, but also tools that assess how much of a text was written by artificial intelligence and how much by a human. So we should not place too much hope in tools that will fight against artificial intelligence because we must keep in mind that these tools are not fighting against artificial intelligence, but against the person who uses artificial intelligence. Therefore, it is a futile, unnecessary, and lost battle from the start. AI ≠ human + AI; human + AI > AI.
Example of giving an instruction
If you ask ChatGPT to write a text about Mount Velebit, it will do so, but by giving such a task, you will not get a Pulitzer Prize-winning text. To be able to get the most out of using tools like ChatGPT, it is essential to know how to give instructions. Simply asking questions is not enough to achieve the greatest effect. Of course, ChatGPT can somewhat answer a question, but it is much more important to ‘prepare’ before we tell it to generate text.
This is generally what ChatGPT sees and expects: ‘Write me [type of text] [size of text] about [topic] in the style of [style]. Write me a 1000-word essay about planet Earth in the style of Carl Sagan. (Write me a 1000 word essay about planet Earth in the style of Carl Sagan).‘
In addition to that ‘formula’, it is important to approach using ChatGPT almost strategically. It is crucial to ask the right questions (to seek something that is unknown or specific, rare, etc.), to request text of certain characteristics (effective, emotional, demagogic, populist), to clearly define limitations (amount of text, what must be included in the text, which definitions to use, what it may write, and what it may not…), to specify the type of text (prose, poetry, scientific article, student paper…), writing style (like a famous person, funny, conversational, melancholic, boring…), level of complexity, references, concrete data, context, comparisons, and various other elements, so to speak, the reality in which that text needs to be generated. This is, of course, a matter of practice and using ChatGPT.
Moral questions
Of course, there are many more possibilities depending on whether it is a specific case such as drafting contracts, writing email messages, summaries, reports, etc., but these are the main steps on how to approach using ChatGPT to generate better texts that we will not only be satisfied with, but no one will know that they were generated by an artificial intelligence-based tool. It is essential that we clearly know what we want, and that means that such tools will be most beneficial for people who are experts in a certain field, as they will be able to achieve even better results. We can only hope that such tools will not cause an even greater flood of objectively poor and meaningless content on the internet that people waste time and attention on.
In the future, when artificial intelligence-based tools will have a much greater impact on, say, shaping public opinion and various social processes, moral questions will be key. When we look at what kind of content is most consumed on the internet, it is clear that there is a huge danger of falling into bias (which is already part of such tools), censorship, manipulation, populism, and propaganda, which can easily lead us back into a ‘dead end’ of history. This is especially important in ‘training’ such tools that cannot distinguish truth from lies, good from evil, essential from non-essential, meaningful from meaningless, and many other concepts that even we humans do not distinguish and understand. So far, very little attention has been paid to these issues.
