"Dear confirmand. You stand on the threshold of life, and today you are taking an important step on the path to adulthood. You are wise, caring and full of possibilities. Life lies open before you."
Elisabeth Asser reads out a speech written by ChatGPT – smooth and warm, but with no actual author.
Asser, who is an associate professor at the University of Agder (UiA) with a PhD in AI and ethics, warns against using artificial intelligence in this way.
"If you’re going to write a speech, you want to tell someone something. If you use artificial intelligence, you’ve failed to think through what’s really on your mind," she says.
The language model praises you for the result regardless, which means you easily convince yourself that the speech has actually turned out well. This phenomenon is called sycophancy – in other words, the model flatters you in order to be liked.
"It’s lovely to talk to someone who thinks everything you do is fantastic", says Asser.
The contract that’s being broken
What is it actually acceptable to use artificial intelligence for? That is the topic of this episode of the ‘Artificial Intelligence’ podcast, featuring Asser, Morten Goodwin and Maren Thormodsæter Lyngroth.
“We get better at what we do a lot of, and worse at what we do less of. When language models write for us, we become less adept at writing and thinking,” says Asser.
Listen to the episode (in Norwegian) here:
But not all text is the same. There is a big difference between a speech and, for example, a user manual. The difference lies in the unwritten contract between the writer and the reader.
“If you’re going to give me a speech, there’s an implicit understanding that you have something genuine to say. If I find out afterwards that you’ve simply run it through a language model, it becomes obvious that you’ve broken that understanding,” says Asser.
FACTS: Green, amber or red?
This is how Goodwin and Asser rated some typical ways of using language models:
- Green: Using AI as a spellchecker. Asking the model to identify weaknesses in your text without correcting them. Asking for counterarguments to something you have written yourself.
- Amber: Asking for a summary of an article you should have read yourself. Asking the model to find literature that supports or contradicts your argument.
- Red: Using AI to write a first draft of a personal speech. Using it to write a first draft of an academic text. Having AI write an opinion piece for you.
“If, on the other hand, you’re writing an instruction manual, the implicit contract with the reader is quite different. There, the expectation is simply that the manual should help us get the product to work.”
No neutral hammer
But who is ultimately responsible for the result if we hand in something of lower quality after using artificial intelligence? Morten Goodwin uses an image from the building site.
“If a house is crooked because the carpenter has used the tools incorrectly, we don’t blame the hammer,” he says.
Asser agrees with the point, but highlights an important difference:
“The hammer doesn’t tell the carpenter where to put the nail. A language model has far more built-in assumptions. It contains a particular view of the world that you may not be aware of when you use it. It is never entirely neutral,” she says.
The reason lies in how language models are built. They are trained on pretty much all the text available on the internet, and then some. This material can range from blog posts and novels to propaganda and hate speech.
Traditional biases in Western culture — for example, attitudes towards women or people with darker skin — will be carried over into language models to varying degrees.
Green, amber and red Lights
In the podcast, Goodwin and Asser go through a range of examples of AI use and give each one a traffic-light rating based on whether they consider the use acceptable or not. AI as a spellchecker: green. Identifying weaknesses in your own text: green. Summarising an article you should have read yourself: amber.
Glossary:
Language model: An AI system trained on vast amounts of text that essentially works by predicting which word is likely to come next in a sequence of text. ChatGPT is the best-known example.
Generative AI: AI that creates something new, such as text, images or audio. The counterpart is AI that sorts or recognises patterns, such as a system that detects tumours in X-rays.
Sycophancy: When a language model flatters the user and gives them excessive praise. This can make its responses seem more convincing than they actually are.
Using AI to write a first draft of a personal speech gets a red light. If the end product is an academic text, the use of AI gets an even redder one. The researchers believe the worst use of all is to have the technology write opinion pieces for you.
“An opinion piece should express your views on something. If a language model writes it for you, it is no longer your opinion,” says Goodwin.
Shame is not the answer
Both are aware of the extent to which the technology has come to influence us. Goodwin has stopped using the em dash because the long dash has become a telltale sign of text written by artificial intelligence.
“I actually really like the em dash. But I’ve stopped using it,” he says.
Asser, on the other hand, refuses to change the way she uses punctuation.
“When this became an issue, I decided very early on that no one was going to take that em dash away from me.”
Goodwin nevertheless believes that shame is the wrong approach.
“It’s fine to use the technology, at least as long as you’re honest about it. If you’ve had your text language-edited by AI, you can say so,” he says.
Wants AI in the curriculum
In the episode, they also discuss the government’s recommendation that AI should be removed from schools from the 2026–27 academic year. Goodwin is sceptical, pointing out a range of everyday tools that are now incorporating artificial intelligence, from Google Search to the word-processing programme Word.
“If we’re not going to learn how to use AI at school, where are we supposed to learn it?” he asks.
Asser is more positive about putting the brakes on for now, but she wants something else to be added.
“Critical thinking needs to be included in the curriculum at the same time. Being able to think clearly and understand how the technology affects us is also a form of preparedness,” she says.
You can find the podcast Artificial Intelligence with Elisabeth, Maren and Morten wherever you listen to podcasts. The podcast is only in Norwegian.