Chat GPT: “A small step for research but a giant leap for utility”

Artificial intelligence is on many people’s lips thanks to Chat GPT. But is the new technology a threat or an opportunity? Or is it both? AI researcher Fredrik Heintz sees great opportunities but also challenges.

Chat GPT – the powerful AI-driven chatbot revolutionising communication. Trained by OpenAI, Chat GPT offers fast and reliable answers to questions about a broad spectrum of subjects. With its natural language processing ability and enormous data set, Chat GPT is the ultimate tool for simplifying communication and increasing productivity.

The above paragraph was written by Chat GPT when prompted to write an introductory text about itself. Maybe not a masterpiece, but still understandable. Is it tempting for a writer to have the chatbot write the rest of the text? Well, that is not going to happen… this time.

The beta version of Chat GPT was released in November 2022. It is, as already mentioned, a chatbot run by artificial intelligence – AI. The results generated are in many cases astonishingly good and have fascinated many in such a short time. After only just over a week, Chat GPT had more than a million users.

With earlier versions of chatbots there was no doubt that your questions were being answered by a computer. Its limitations when it comes to natural language processing and its inability to adapt its replies to the context have been obvious. But Chat GPT blurs the lines between human and machine, for better or worse.

Fredrik Heintz is one of Sweden’s most prominent researchers in AI and the coordinator of the TAILOR project, which aims to build the capacity to provide the scientific foundations for trustworthy AI across the EU. 

“There are quite a lot of interesting things in the texts it produces, and it manages to extract a lot from the source material, which could provide new insights. At the same time, it can’t tell the difference between nonsense and what makes sense,” he says.

Chat GPT is trained with deep neural networks on enormous data sets to provide answers that are as general and accurate as possible. The model has processed approximately 500 billion concepts from books, web archives and Wikipedia. To fine-tune its answers, it is trained by humans. It also uses self-learning to get better and better – a working method long used in AI research.

“It’s not really because of any new scientific breakthroughs that Chat GPT works so well. It’s the scale that matters. The more training data, the better the results. I’d say that it’s a small step for research but a giant leap for the utility of AI,” says Fredrik Heintz.

And the utility of Chat GPT seems almost infinite. It can do anything from being a human-like chatbot to coding software, composing music and writing short stories in the style of Joyce Carol Oates or Edgar Allan Poe. 

“Human culture largely consists of text. Chat GPT has been given so much attention because the language it generates has come so much closer to us. Maybe a little too close,” says Fredrik Heintz.

But in addition to writing creative stories, innovative poetry and emotional lyrics it can also write essays, scientific articles and reports. This has made some of us think about how to relate to facts and knowledge. 

Currently existing AI is trained on existing texts and images once created by humans. Most of this material can be found on the web. This is also where the AI generated material will end up. But what happens when we find ourselves in the precarious situation that AI is trained using its own texts? 

Some say that knowledge risks becoming a floating concept when no-one has the tools needed to fact-check AI generated content on the web. But Fredrik Heintz is not that pessimistic. 

“It will be possible to generate new knowledge and find new perspectives on things we already know. Chat GPT can simply provide us humans with new insights,” he says. 

He does, however, see other problems with Chat GPT than the possibility of generating large amounts of content in a short time. This in combination with social media advertising could result in a deluge of tailored messages in our flows. 

He also sees a risk that AI trained on existing online material will be biased against certain cultures, genders and disabilities. Open AI has built-in filters against e.g., homophobia, racism, sexism and ableism. But these filters are currently easy to bypass.

It is therefore important to lay the foundations for reliable and human-centred AI right now. Researchers in the EU TAILOR project have specified three criteria for achieving trustworthy AI: it must comply with laws and regulations, comply with a set of ethical principles and its implementation must be robust and safe. In Fredrik Heintz’ opinion, this entails huge challenges, mainly as concerns compliance with the ethical principles.

“Take fair distribution, for instance – does this mean equal distribution among individuals or that everyone is given the resources they need to get to the same level? These are major long-term issues that will take time to solve. After all, the definition of what is fair has been debated by philosophers and thinkers for hundreds of years.”

Facts: Chat Generative Pre-trained Transformer or Chat GPT is a language model-based AI developed by the US company Open AI.