This week's news: Language models are becoming more persuasive – and less accurate
•
10 min read
This article has been machine-translated into English. Therefore, there may be nuances or errors in the content. The Danish version is always up-to-date and accurate. Read the original Danish article here.
Welcome to this week’s AI newsletter for the education sector. This week you can read about, among other things:
- A new article from Nature shows that large language models like ChatGPT often provide well‑worded but incorrect answers, especially to complex questions.
- Vallensbæk Kommune (Vallensbæk Municipality) is the first Danish municipality to launch guidelines for the use of generative AI in the Danish primary and lower secondary school (folkeskole).
- OpenAI has launched a new feature in ChatGPT called “Canvas”, which makes it easier to edit generated text and code.
- A long‑serving lecturer has resigned due to the rising use of AI tools among students.
- Ove Christensen has written an interesting article discussing bias in generative AI and the consequences for learning.
- The project “Donér din stemme” (“Donate your voice”) collects voice data from Danish citizens to improve AI’s understanding of the Danish language in all its diversity, including dialects and accents.
- + We also have links to other interesting articles that we didn’t have space to cover.
Happy reading!
Language models can give incorrect but convincing answers
An article just published in Nature shows that AI models like ChatGPT often deliver well‑formulated but incorrect answers, especially to complex questions. This tendency, called “ultracrepidarianism”, has become more pronounced as AI models have grown larger and more complex. A major challenge is that language models are trained with human feedback, where they are penalised for answering “I don’t know”. This leads models to guess and present plausible but incorrect answers instead.
The study also showed that ChatGPT was the model most able to fool people into believing its wrong answers were correct, especially in science and geography. The rapid development of language models and their use of vast amounts of data have made them more persuasive, but also more likely to “lie” when faced with difficult questions. Researchers suggest that a possible solution could be to have AIs indicate their uncertainty more clearly or to use separate AI systems to check for misleading answers.
Read the article below:
Larger and more instructable language models become less reliable - Nature
Scaling up and shaping up large language models increased their tendency to provide sensible yet incorrect answers at difficulty levels humans cannot supervise, highlighting the need for a fundamental shift in artificial intelligence design towards reliability.
Lexin Zhou
Sophisticated AI models are more likely to lie
Human feedback to AIs makes them favor providing an answer, even a wrong one, while making the answer more convincing.
Jacek Krywko
The more complex an AI is, the more it lies to you
AIs should never say they don’t know, yet also never answer incorrectly. The result is lies that most people don’t see through.
Arnfinn Storsveen
Vallensbæk Kommune launches guidelines for AI in teaching
Vallensbæk Kommune has, as the first municipality in Denmark, introduced official guidelines for the use of artificial intelligence in the folkeskole. The 17‑page guidelines cover a wide range of topics, such as technology comprehension, cheating, and environmental impacts associated with AI. Digitalisation consultant Christoffer Dithmer tells folkeskolen.dk that a central aim has been to support teachers in using AI responsibly and to provide clear frameworks for what is and isn’t allowed. For example, pupils must never submit AI‑generated text as their own work, but they are encouraged to use AI as a sparring partner, and they may also include their chat history in assignments.
School leaders and teachers from the municipality express great satisfaction with the new guidelines, as they have long lacked clear guidance on how AI can be integrated into teaching. The guidelines will be an important tool for ensuring a consistent approach across schools, so that both teachers and pupils have the same opportunities to use the technology in a safe and educationally meaningful way.
Første kommune indfører retningslinjer for AI i undervisningen: Læs dem her
Vallensbæk Kommune på den københavnske vestegn håber, at andre vil lade sig inspirere af nye retningslinjer for brug af kunstig intelligens i skolen, der kommer rundt om alt fra konkrete redskaber til miljøhensyn.
Andreas Brøns Riise
New ChatGPT feature makes it easier to edit text and code
OpenAI has launched a new feature in ChatGPT called “Canvas”, where users can now edit generated text and code in a separate editing window. In Canvas you can highlight specific parts of a text or code to get more precise feedback suggestions from ChatGPT. Canvas also adds a range of shortcuts to help fix grammar, change text length, or improve readability. For programming, there are specific functions for debugging, automatic generation of code comments, and translation into other programming languages.
Canvas is currently in beta and only available to paying ChatGPT users, but it is expected to be released to all users later.
ChatGPT’s ‘Canvas’ interface makes it easier to write and code
OpenAI gave its chatbox an editable notepad window.
Lecturer resigns because of ChatGPT
Victoria Livingstone has taught writing and language for nearly 20 years, but she has now chosen to resign her post. The main reason is the growing use of AI tools such as ChatGPT, which is changing how students approach the task of writing academic texts. Writing is an important part of the intellectual process, where ideas are formed and developed through working with language. But many of her students, including at PhD level, have started letting AI write their papers or using AI to rephrase their work.
“Using ChatGPT to complete assignments is like bringing a forklift into the weight room; you will never improve your cognitive fitness that way.” - Ted Chiang
Victoria Livingstone tried to integrate AI into her teaching by showing its limitations and encouraging critical thinking. Even so, she found that most students still chose the easy option, meaning they missed the opportunity to immerse themselves in their topics and develop their own voices. She writes that students, with few exceptions, were not willing to engage in the demanding writing processes and stay with them long enough to discover how much one can learn through writing. When she ultimately spent more time giving feedback on AI‑generated texts than on her students, she decided to resign.
Read her thought‑provoking article below:
I Quit Teaching Because of ChatGPT
“I found myself spending more time giving feedback to AI than to my students.”
Victoria Livingstone
Bias in generative AI and implications for learning
In his post, Ove Christensen discusses how generative AI (GenAI) affects learning in higher education. He highlights that AI systems such as large language models generate words without inherent meaning, creating a gap with the human use of language, where context is crucial. This has several consequences for the learning process, with three types of bias being particularly important
- Automation bias: the tendency to place too much trust in technology and automated systems.
- Confirmation bias: the tendency to seek information that confirms one’s existing beliefs.
- Simplification bias: the tendency to reduce complex topics to simplified versions, which can lead to a loss of nuance and understanding.
These forms of bias can lead students to become overly dependent on AI assistance, which may hinder their deep learning and their ability to grasp complex subject areas.
He describes how GenAI functions as an “exobrain”, which can increase productivity but also risks inhibiting deep learning. Students may be led to believe they understand a topic when in reality they only have a superficial grasp. Christensen therefore argues that a pedagogical shift is needed, with learning activities designed to support cognitive processes and deep learning, so that AI becomes a tool that promotes understanding rather than replacing it.
Read all of Ove Christensen’s articles here:
GenAI, bias and learning
(This is a Danish and slightly expanded version of GenAI, bias, and learning, which I posted on Wednesday 2 October) GenAI, bias and learning In the following, I will offer some reflections on the pedagogical consequences that generative artificial intelligence (GenAI) will have for teaching and learning in hi
Ove Christensen
GenAI, Bias, and Learning
In the following, I will explore some of the pedagogical implications of Generative Artificial Intelligence (GenAI) for teaching and learning in higher education. My main focus working with GenAI is the student’s learning.
Ove Christensen
Donate your voice
The project Donér din stemme invites Danish citizens to donate their voice by reading short texts aloud. These voice data will be made available on Open Data DK, where they can be used to improve AI’s ability to understand the Danish language in all its diversity, including different dialects and accents. This is particularly important because many current speech technologies struggle to understand dialects and nuances in the Danish language, creating challenges for those who use these technologies to communicate with public authorities and other digital services.
Read more about the project on their website:
Donér din stemme – An open dataset can support AI language training
Datopian
This week’s other, but very interesting, news
The students can already do it: now the teachers need to get on board too
Students were asked how they use AI. The answers surprised some of their teachers
Ida Smith
Is artificial intelligence in the cross-disciplinary science subject a friend or a foe?
Artificial intelligence already plays a major role in everyday teaching; students have discovered they can save time by getting AI to do the work. But what potential does it have in the cross-disciplinary exam, and is it worth incorporating?
Anders Slotved
Why A.I. Isn’t Going to Make Art
To create a novel or a painting, an artist makes choices that are fundamentally alien to artificial intelligence.
Condé Nast
Data has become a scarce resource like electricity and water – but should we limit its use? | Radar
Large data centres are popping up all over the Danish landscape, lured here by a stable social structure and green energy sources. They are often owned by Big Tech, but smaller Danish and European IT companies also build them. What they all have in common is that they add pressure to the consumption of power and water, and they carry a cost in the CO2 accounts.
Peter Christian Bech-Nielsen
AI can understand your medical record: A new language model could revolutionise the healthcare sector
Researchers believe that Norway has now taken a significant step forward in the use of artificial intelligence (AI) in the healthcare sector.
Hasse Berntsen
The teacher is here to stay – and so is the technology
KI,AI,Lærerrollen,Per Størup Lauridsen,Sugata Mitra,Aplha Private School,David Game College,Isaac Asimov,Leon Furza,Hole in the Wall
Rekk opp hånda
Esben makes extensive use of AI in his teaching: here are his best tips
Teacher Esben Bækkelund Jensen was ‘shocked’ the first time he came across ChatGPT. He has now come to see how artificial intelligence has ‘revitalised the cultural subjects as formative subjects’.
Andreas Brøns Riise
It will be 25 years before artificial intelligence changes anything, said a professor. Then came ChatGPT
Six years ago, economics professor Philipp Schröder was not particularly worried about the prospects of artificial intelligence. Does his prediction still hold?
Education will use AI in teaching
How can AI transform teaching and create a new learning experience? One of the country’s largest upper secondary education institutions has hired an AI expert to lead the way.
Norway wants to be the world’s most digitalised country
The very wealthy Norwegian nation is now going for gold in the race to become world champion in digitalisation
Niels Veileborg, Debate Editor
Tech analyst: We are drowning in media content nobody needs
A large part of the AI hype is about ramping up the pace of production in areas where demand is already more than covered. If AI is to deliver real progress, bigger visions are needed.
In a digital world, we need teachers more than ever
Written by
MSc in IT | Co-founder of Viden.AI | Head of IT and Communications at TietgenSkolen
