Newsletter: Artificial intelligence - cheating or learning?

Here is newsletter no. 2 for week 45, where we round up the stories there wasn’t room for in the first newsletter.

• VidenSkaber has held a workshop at Aabenraa Statsskole, where upper secondary school (gymnasium) students, together with experts, discussed how to use generative AI wisely in an educational context and avoid cheating themselves.
• A BBC article describes how a student, Hannah, regrets using generative AI to cheat, as it led to an investigation and left her feeling ashamed.
• Several of the leading universities in the UK are seeing a dramatic rise in cases where students use AI to cheat in exams.
• Jeppe Klitgaard Stricker writes in an article about how AI challenges critical thinking in higher education and argues that the education system needs to be rethought.
• AI tutoring delivers, according to a Harvard study, better learning outcomes than traditional active learning in the sciences.
• But there is also criticism of the hype around AI tutoring and its learning impact. Associate Professor Mikala Hansbøl questions the results of the Harvard study and warns that the hype around AI-based learning tools may lead to superficial expectations of technology’s role in education.
• Researchers from Stanford University have developed Tutor CoPilot – an AI assistant that guides tutors in real time with suggestions for pedagogical strategies.
• New article about STORM – a system developed to help language models write well-structured articles based on reliable sources in the style of Wikipedia.

Meanwhile, Claus has just published an article on shared knowledge and the challenges that arise in an age of generative artificial intelligence. This is particularly pertinent given that some students regard language models as a valid source of knowledge without themselves exercising source criticism.

Read it here:

Shared knowledge and philosophy of science: challenges in an age of generative artificial intelligence

As generative artificial intelligence, based on large language models such as ChatGPT, Gemini, Claude or Copilot, becomes an increasingly integrated part of everyday life in educational institutions, our traditional understanding of knowledge—and especially the concept of “shared knowledge”—is challenged. This development raises fundamental questions within the philosophy of science and

Viden.AIClaus Scheuer-Larsen

Artificial intelligence – cheating or learning?

What is the sensible use of generative artificial intelligence in an educational context? How do you avoid cheating yourself when using AI as a support tool? Watch below to hear what upper secondary school (gymnasium) students at Aabenraa Statsskole learned from a workshop on artificial intelligence and learning.

In September 2024, the outreach initiative VidenSkaber of Videnskabernes Selskab (the Royal Danish Academy of Sciences and Letters) ran a workshop for a 3rd-year class at Aabenraa Statsskole together with Ane Qvortrup, Professor of Education at SDU (University of Southern Denmark), Magnus Høholt Kaspersen, Postdoc at Videnscenter for teknologiforståelse (the Knowledge Centre for Technology Understanding), and a group of students from Cognitive Science at Aarhus Universitet (Aarhus University).

VidenSkaber is supported by the Carlsbergfondet, Lundbeckfonden and Novo Nordisk Fonden.

Read more at videnskaber.dk, where you can also find podcasts, articles and videos on VidenSkaber’s theme “how is artificial intelligence changing our society?”

VidenSkaber – How do we know?

In VidenSkaber, we explore everyday and current issues and ask the question: How do we know? VidenSkaber is a meeting place. Here we bring citizens and researchers together to discuss the major issues of everyday life. We create and share knowledge when, in workshops, debates and communication, we insist on asking: How do we know? And we do it together with some of Denmark’s finest researchers across scientific disciplines.

VidenSkaber logo

When artificial intelligence becomes a temptation

The BBC recently published an article about a student named Hannah who regrets having used artificial intelligence to cheat. She shares her experience of using generative AI to finish an assignment under the pressure of illness and tight deadlines. Although AI helped her submit the task, her use was detected through AI detection, which led to an investigation for academic misconduct. She avoided severe penalties, but the experience left her with deep regret and shame.

Hannah’s story highlights the complex challenges of integrating AI into higher education, where the technology can be both a valuable tool and a potentially problematic temptation.

Read the article here:

‘I used AI to cheat at uni and regret it’ - BBC News

Hannah says she is lucky she wasn’t thrown out of university after being caught using AI.

BBC NewsBen Moore

Sharp rise in AI-related cheating at UK universities

Several leading universities in the UK report an explosive increase in cases of students using artificial intelligence (AI) such as ChatGPT to cheat. According to new data from Times Higher Education, for example, the University of Sheffield saw a rise from just 6 suspected cases in 2022–23 to 92 in 2023–24. Other universities, such as Queen Mary University and the University of Glasgow, have seen similarly sharp increases and sanctioned almost all suspected cases with penalties such as warnings or grade reductions.

Data from 24 Russell Group universities, however, point to large differences in the handling of AI cheating. While some institutions have detailed breakdowns of AI-related cheating, others do not record these cases at all or leave the handling to individual faculties. Some institutions, such as the University of Southampton, have no data distinguishing AI cheating from other academic misconduct, which, according to expert Thomas Lancaster, underscores uneven practice and a lack of traceability.

According to researchers, universities face challenges in defining and detecting AI cheating, as it can be hard to prove. Michael Veale from UCL points out that centralising procedures could make cases harder to process, as assessing AI cheating often requires subject-matter expertise. Des Fitzgerald from University College Cork argues that the problem stems from the tech sector’s rapid roll-out of these tools and believes that traditional plagiarism policies are not sufficient to address AI-related challenges in assessment environments. He recommends a broader re-evaluation of assessment forms and a national AI policy for higher education, which is already underway in Ireland.

Student AI cheating cases soar at UK universities

Figures reveal dramatic rise in AI-related misconduct at Russell Group universities, with further questions raised by sector’s ‘patchy record-keeping’ and inconsistent approach to detection

Times Higher Education (THE)Jack Grove

Why domain-specific knowledge matters more than ever

As generative AI offers quick answers to almost anything, the ability to understand and assess information becomes crucial. In higher education, critical thinking is already a central goal, but the advent of AI challenges how we understand and practise this competence. The article argues that critical thinking is not only about spotting misinformation; it also requires deep domain-specific knowledge. This expertise enables us to evaluate AI-generated content and distinguish between valuable insight and potentially misleading information.

Jeppe Klitgaard Stricker suggests that the education system should be rethought in light of AI’s role, rather than merely adapting existing structures. Institutions should experiment with AI in teaching and develop solutions through practice and continuous learning over time. The challenge is to convince students that in-depth knowledge is still crucial—even at a time when AI can quickly deliver ready-made answers.

Critical Thinking in the Age of AI

Why Domain Expertise Matters More Now, Not Less

The Future of Higher EducationJeppe Klitgaard Stricker

AI tutoring delivers better learning outcomes than active learning in science

A study conducted at Harvard University compares AI-based tutoring with active learning in physics teaching for university students. The results show that students who used the AI tutor achieved more than double the learning gains in less time compared with students who took part in active classroom instruction. The study was conducted as a crossover experiment, in which groups of students experienced both modes of instruction and were tested before and after each lesson.

The researchers behind the study point out that AI tutors can be used to introduce new topics, enabling teachers to focus on advanced skills such as problem-solving and critical thinking during class time.

AI Tutors Double Rates of Learning in Less Learning Time

Inside Harvard’s new groundbreaking study

Dr Phil's Newsletter, Powered by DOMS™️ AIDr Philippa Hardman

Critique of the hype around AI tutoring and its impact on learning

Although the Harvard study highlights promising results for the use of AI tutors in physics teaching, substantial criticism has arisen regarding the study’s conclusions and method. Associate Professor Mikala Hansbøl questions whether the reported results are representative of the learning field more generally, and warns that the hype around AI-based learning tools can lead to superficial and misleading expectations about the role of technology in education. Her critique highlights the need to look beyond the immediate results and instead examine how theoretical perspectives on learning and pedagogy can contribute to a more nuanced understanding of AI’s potential in the teaching context.

For Mikala Hansbøl, the article is an example of how AI hype can create inflated expectations without sufficient scientific substance. She urges her network to avoid the uncritical dissemination of non–peer-reviewed articles and stresses the need for AI and learning researchers to work with greater theoretical depth and a more nuanced understanding of learning processes.

”AI Tutoring Outperforms Active Learning”

I call hype & bullshit!

Mikala’s SubstackMikala Hansbøl

AI assistance makes tutors more effective

Stanford researchers have developed Tutor CoPilot – an AI assistant that raises the quality of mathematics teaching. The system guides tutors in real time with suggestions for pedagogical strategies, so they ask the right questions and help students find the solution themselves.

Students whose tutor used Tutor CoPilot achieved mastery 66% of the time, compared with 62% without AI support. For inexperienced tutors, the effect was particularly pronounced, with a 9% increase in success rate. Tutor CoPilot also improved tutors’ language, leading them to ask more guiding questions rather than giving the answers.

"AI should support, not replace, teachers," the researchers emphasise. With AI as a pedagogical sparring partner, the vital teacher–student relationship is preserved, while teaching becomes more targeted and students become more independent. Tutor CoPilot illustrates the potential of AI tools as accessible supports that help educators foster more engaged and independent students.

What Happens When an AI Assistant Helps the Tutor, Instead of the Student

A randomized controlled trial from Stanford University examines the efficacy of an AI-powered tutoring assistant.

Education WeekLauraine Langreo

Scholarly articles

Assisting in Writing Wikipedia-like Articles From Scratch with Large Language Models

The article above introduces STORM, a system developed to help large language models write well-structured and well-grounded articles based on reliable sources, in the style of Wikipedia. The STORM system focuses specifically on the complex planning phase, in which the topic is researched and a comprehensive outline is produced. This phase involves gathering diverse perspectives and simulating conversations to create deeper insight into the topic.

You can try STORM yourself below by clicking the image (login required).

Assisting in Writing Wikipedia-like Articles From Scratch with Large Language Models

We study how to apply large language models to write grounded and organized long-form articles from scratch, with comparable breadth and depth to Wikipedia pages. This underexplored problem poses new challenges at the pre-writing stage, including how to research the topic and prepare an outline prior to writing. We propose STORM, a writing system for the Synthesis of Topic Outlines through Retrieval and Multi-perspective Question Asking. STORM models the pre-writing stage by (1) discovering diverse perspectives in researching the given topic, (2) simulating conversations where writers carrying different perspectives pose questions to a topic expert grounded on trusted Internet sources, (3) curating the collected information to create an outline. For evaluation, we curate FreshWiki, a dataset of recent high-quality Wikipedia articles, and formulate outline assessments to evaluate the pre-writing stage. We further gather feedback from experienced Wikipedia editors. Compared to articles generated by an outline-driven retrieval-augmented baseline, more of STORM’s articles are deemed to be organized (by a 25% absolute increase) and broad in coverage (by 10%). The expert feedback also helps identify new challenges for generating grounded long articles, such as source bias transfer and over-association of unrelated facts.

arXiv.orgYijia Shao