ChatGPT: fact, fiction and truth

Lately, many articles appear in newspapers and news items on television about ChatGPT. Because ChatGPT seemingly knows the answer to all questions, the question arises whether we want to use this kind of technology. Or that it is inescapable. This article tries to figure out what the ethical implications of ChatGPT are. And especially what it means for our understanding of facts, fiction and truth.

ChatGPT is based on GTP-3 or GTP-4, so-called Large Language Models. These models can answer in the same natural language based on natural language, spoken or written language. There are also Large Language Models that can generate an image based on a question or assignment formulated in natural language. This is called Generative AI: AI that generates text, images, or videos. But for this discussion, we assume ChatGPT from OpenAI, Microsoft Bing, Replika, Google Bard and many others. You can have a dialogue with these machines and let the artificial intelligence, through questions and answers, carry out textual assignments. You can ask the software to generate a text about the subject you choose.

So how does it work? Based on gigantic amounts of texts, extracted from the internet, artificial intelligence has learned how a language “works”. Which words follow which words. And the machine is now doing that very well. The sentences produced by the software are grammatically and syntactically correct. Good enough to pass as human text.

What is ChatGPT good at?

What ChatGPT is pretty good at is creating fiction. The algorithm is fairly good at making up stories, although the story-line is sometimes friable, the characters are flat and the depth is mediocre. There is now so much fiction being written by ChatGPT that an American fantasy magazine is flooded with stories written by the computer. “It’s a huge hype. Non-writers see it as an easy way to make money, but they will do something different when they fail,” writes Martijn Lindeboom in an article in the Dutch daily paper de Volkskrant.

What is ChatGPT bad at?

But what is the big problem with this language software? First, ChatGPT presents all answers with certainty. Right and wrong answers. The software does not hesitate and can present the biggest nonsense with great persuasiveness. ChatGPT can come up with answers on its own, it hallucinates. The reader of the answers cannot determine whether ChatGPT is writing the truth. For that, the reader has to have knowledge on the topic being written about. But what if the reader doesn’t have that knowledge? People tend to quickly label statements from computers as true. “People tend to take everything the bot says indiscriminately,” writes the Dutch IT journal AG Connect.

All kinds of ethical concerns

Soon, all kinds of ethical objections arose against ChatGPT and other, similar algorithms. Let me briefly list some of the problems that experts in the field of AI and ethics see:

1. Pupils and students have long since discovered ChatGPT as an easy method to write a thesis or a report. Teachers were not prepared for this and are now trying to come up with methods to determine who or what wrote a text. Tools are now being made for this, but perhaps we should carry out writing assignments differently. Back to pen and paper or something else? The same discussion played a role in the rise of search engines we have now become accustomed to it.

2. ChatGPT consumes a lot of electricity and thus indirectly emits a lot of CO₂. This is a general problem in ICT and is getting worse with further digitization. As a colleague sees: ‘If you want to save energy, write your own emails and don’t let an AI do it. Laziness contributes to the climate crisis.

3. To prevent unwanted answers, filters are added to ChatGPT. Certain questions that are politically sensitive, discriminatory or pornographic are filtered out. For this, people are hired in Kenya for $2 per hour. The development of social media and artificial intelligence is accompanied by exploitation. The removal of toxic content remains (underpaid) human work.

4. Large Language Models, of which ChatGPT is one, use huge collections of data. This data may be protected by copyright. But the tools reuse this data without having asked permission. Especially with so-called works of art made by AI, artists are genuinely concerned that their work will be plagiarized.

But I have a bigger objection to using ChatGPT: it violates the truth. When you use ChatGPT to write fiction, a fairy tale or a poem, there’s little to worry about. The stories may be mediocre, but being average may be good enough. Not all prose has to be literature.

But when it comes to describing facts, there is an ethical problem. OpenAI, the builder of ChatGPT, already warns on their website that the answers that the AI gives can be wrong. OpenAI itself provides the following warnings. ChatGPT…

• May occasionally generate incorrect information.

• May occasionally produce harmful instructions or biased content.

• Limited knowledge of world and events after 2021.

But ChatGPT does not indicate when this is the case. Or when there is a chance that this is the case. Wrong or bad answers are also presented as if there is no doubt. ChatGPT never hesitates, and bluffs by presenting falsehoods as true. And that is not surprising when you use large parts of the internet as a source of information. Because of the amount of misinformation already online, which is being fed into the training models for major language models, people who use them can also unintentionally spread falsehoods. Not everything on the internet is true, to say the least.

“If you want to know where the information ChatGPT gives comes from, you are completely in the dark.” (Luc Steels, VUB)

Also, ChatGPT does not provide any citations. Classic search engines link to the source documents, and the knowledge system IBM Watson also showed its sources. But ChatGPT’s Deep Learning algorithm can’t provide the sources at all. It huddles together different data into beautiful sentences. The answers are, in a sense, coincidental.

In addition, the answers that ChatGPT gives can also contain prejudices. They may be true, but they are strongly coloured. Where the discussion around artificial intelligence in recent years has mainly revolved around avoiding ‘bias’, it now appears that an AI tool has become available that once again spreads prejudices. This only becomes problematic when texts with prejudices, generated by ChatGPT, are placed back on the internet. Which in turn are used by ChatGPT as source information. Artificial Intelligence can reinforce biases and ChatGPT is no exception.

Fake News & Co.

Since time immemorial, fake news, gossip, falsehoods and misleading information have been created and spread all over the world. This “false” information is spread out of self-interest, political interest, and thirst for power but also out of stupidity and ignorance. Malicious intent is not always present, but the consequences can be malicious. Spreading fake news is however deliberate. Trolls who spread an alternative truth, or simply falsehoods, have political goals. Social media provides these trolls with an easy method to spread their messages. With far-reaching social consequences for democracy and cohesion. “Democracy is not a product, but a common process. And that falters because of AI.” writes Maxim Februari in the Dutch newspaper NRC.

“Generative AI creates a world where anybody can be creating and spreading misinformation. Even if they do not intend to.” (Tim Gordon, Best Practice AI)

Generative AI doesn’t change much in this situation. Generative AI only facilitates the human trait of bending facts to his will, or presenting fiction as facts. Has nothing changed?

Industrialization of falsehoods

What Generative AI can do is facilitate deception on a large scale. Security officers are already afraid that phishing e-mails can be written by such tools in such a way that they can hardly be distinguished from sincere messages. It will also be easier to create fake news. ChatGPT and others are good at fantasizing, so they can also write falsehoods. I think these algorithms are even better at it than describing truths. Even though tech companies are going to charge a subscription fee for such writing assignments, writing an article without checks becomes much faster and cheaper than compiling a truthful article. Journalists and researchers cost money. Writing thorough texts takes time.

We will therefore have to deal more with texts that contain falsehoods. Out of greed, power or ignorance. If we can no longer easily distinguish truths from falsehoods, this will have social consequences. Who else can I trust? What truth is true? Is everything relative and does the truth no longer exist?

Not long ago, culture was stamped by relativism. Truth was “unknowable.” Maybe it existed somewhere ‘out there’, but every idea about it was equally valuable. (Matt Hodges)

It is also important to realize that those who have money and power can now more easily control the truth. Those who have access to artificially intelligent systems, and also control over distribution channels such as social media, can spread their version of the truth, or their own truth, much more easily than others with fewer resources. We’re already seeing this happening on Twitter. This, too, is not new in history. But artificial intelligence does make it a lot easier.

What is truth?

If we can no longer discern the truth in the flood of fake news, we also face a challenge as humans. Perhaps the hype surrounding ChatGPT and others passes because we value human truths higher. That a text written by a human being becomes more valuable than a machine-based text. But we will soon no longer be able to recognize the difference between the two. Malicious people can always make their computer texts pass as human-made. We will become more sceptical of the information we get. At the risk of seeing the real as untrue.

This is a shortened, English version of my article published on Techthics.nl and cvandaag.

Image Creative Commons CC BY-SA 4.0 by Therealjimricks via Wikimedia.