To make their assignments, my students increasingly use chatgpt or similar generative artificial intelligence (AI). The development of this is so fast that examination committees and teachers’ teams at universities are only slowly starting to get a picture of the place that AI must have in academic education.
Consensus is starting to arise about what AI students can and cannot help. Checking spelling, writing a piece of computer code, AI can help with that. Then AI is a referee who can tell what is correct and what is not. But for me there is a clear limit: the use of generative AI for writing texts should be prohibited. Generative artificial intelligence is designed to write an average text as well as possible. On average, that is boring, flat and predictable. And training for boring medium, is not a university for that.
To understand why Chatgpt will never be able to produce a creative text or an original idea, we have to dissect what such Large Language Model actually does. In the core, the model searches for the most likely word that follows the words that have already been written. You can see the operation of it nicely in chatapps on your phone. If you ‘heartily’ in ‘heartily’, give your phone as follow -up suggestions ‘congratulations’, ‘congratulations’, or an emoji with a party hat. With that you probably have the right continuation of that sentence – the one time that you heartily dismiss a condolence per text message.
The essence of generative language models is that they always choose the most likely continuation. That prevents just about all linguistic slips, and also part of the substantive mistakes. But it also eliminates the chance that a story will be given an unexpected twist. You will never be positively surprised by an original turn in a chatgpt text. In the urge to exclude errors in a text, it also closes the door for a positive outlier.
Steering for probability is the reason that AI does not have election programs can keep apart. A language model has no substantive knowledge of topics. If you then ask such a model what the VVD and the PVV have for views on migration, then a language model halfway through a sentence about asylum seekers simply does not know whether VVD or PVV should be there. The chance of one of the two options is about the same. Both parties write similar texts about asylum. And instead of admitting a language model that it is not able to indicate the difference, it then chooses one of them. It is as if you like several main courses on the menu. Yet you will really only choose one. The final choice is still created in blind panic. But where you can then think for minutes whether you should not have chosen the other dish, the algorithm of a language model after such an arbitrary choice is already busy with the rest of his story.
To formulate new ideas, you will have to work yourself
Music service Spotify nowadays also stops songs that are made with AI in playlists of listeners. That is useful, because Spotify does not have to pay any rights to an artist about that. But those AI songs are all incredibly soulless and mainstream. Here too, the inability is to make something creative, because AI is simply designed to give the most likely follow -up to a melody or chord diagram. AI will therefore never be the following Beethoven, Stravinsky, Queen or Radiohead.
The problem with the use of AI in academic education is precisely the tendency of language models to make the average texts as possible. Chatgpt feeds the six -culture. We can expect more from students than texts that strive for medium. We would not accept it from a conservatory student if those compositions made by a computer.
I was reminded of the Feynman Lectures on Physics from the brilliant physicist Richard Feynman. He explains that you can always reconstruct an existing physical formula through two other formulas combine. Although you have not remembered all formulas, you can always find them by combining them. But that is not the use of a good physicist, says Feynman. A good physicist makes combinations of formulas that has never made anyone else. Only that leads to new discoveries. To learn how to approach that, you really have to understand the connections in nature.
That’s how it is with good writing. AI is perfectly able to write texts that are an extension of what has already been written. But to formulate new ideas and get them clear on paper, you will have to work yourself. Writing is not about after-monkeys or reproduce. Writing is a creative process where originality and surprise are the core. It is an essential academic skill that you will have to learn yourself. Outsourcing to AI is therefore not belonging at a university.