This Week's News: AI Favors Left-Leaning Political Views
•
9 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.
Last week, attention centred on the recently concluded European Parliament election. DR’s podcast, Prompt, looked at how language models performed in the various candidate tests. Equally interesting is that a study from Københavns Universitet (University of Copenhagen) has tested Meta’s language model, Llama, and found that it also favours left-leaning and pro-European views. The researchers behind the study suggest developing language models that can inform citizens about politics. Such a language model would also be very useful as part of teaching in social studies and history.
It also emerged that Microsoft is expanding its data centres in Sweden while upskilling 250,000 Swedes in artificial intelligence — at a total cost of almost DKK 22 billion.
Tine Wirenfelt Jensen has written a rather interesting article discussing what happens when artificial intelligence becomes part of the guidance and counselling space.
The documentary genre is also evolving, and Gary Hustwit is using artificial intelligence to assemble his new documentary about the composer Brian Eno. That means it is a completely new version each time the film is shown.
Progress in generative artificial intelligence that turns text into video seems to be moving fast at the moment, and there is now a competitor to OpenAI’s Sora. The Chinese technology company Kuaishou has just unveiled KLING, and we have collected a couple of their videos here in the newsletter.
Finally, we have a range of interesting articles and news from Denmark and abroad.
Thank you for reading.
Chatbots’ answers in DR’s EU election candidate test reveal a left-leaning tendency
On DR’s podcast Prompt they examined how different chatbots, including ChatGPT, would answer the candidate questionnaires related to the EU election. The results showed a general tendency towards left-leaning parties such as Alternativet, Radikale and SF.
The experiment was inspired by a Reddit user, and the idea was to see whether chatbots could answer all the questions in DR’s candidate test. Although Google Gemini refused to respond, it was possible to get the other chatbots to complete the test by asking the questions individually. Most answers were either “agree” or “disagree”, some with reasoning, others brief. Henrik Moltke emphasises that the chatbots’ answers do not necessarily reflect a political stance, but rather a weighting from their training data and especially their fine-tuning. This is done to prevent them from giving offensive or unlawful answers.
To read more about the experiment and hear about the chatbots’ political tendencies, you can read DR’s article below.
Turkey in the EU? A TikTok ban? See what ChatGPT answers in DR’s candidate test
DR’s podcast Prompt has had a range of chatbots take the EU election candidate test. The answers point in one direction: to the left.
Meta’s Llama Chat favours left-leaning views
A study from Københavns Universitet has revealed that Meta’s chatbot Llama tends to favour left-leaning and pro-European views. By testing the language model with questions about EU policy, the researchers found that Llama often gives responses that align with political positions from groups such as De Grønne/EFA and S&D rather than centre-right and right-wing parties like EPP and ID.
The study found two main reasons for this skew: firstly, the datasets used to train the model may be biased. Secondly, Meta’s ethical guidelines influence the model’s training to avoid controversial positions.
The researchers also noted that this issue is not limited to Llama. Political bias has been found in several large language models, which can be problematic when they are used for political guidance. With proper training and safeguards, however, such models could potentially provide citizens with better and more nuanced information.
The researchers behind the study, Ilias Chalkidis and Stephanie Brandl, see potential in developing language models that can better inform citizens about politics. They point out that although current models are often biased and misleading, they have significant democratic potential. An improved language model could help voters better understand political issues and parties’ positions, especially in the run-up to elections. The researchers’ vision is that such a model could answer citizens’ questions on topics such as the green transition, parties’ historical actions and voting behaviour.
💡
Comment
Such a language model would also be useful in teaching, as it could help pupils understand political issues. It could be used as a tool to review parties’ positions, historical decisions and their influence on current political debates. In addition, it could give pupils a more nuanced and objective view of political topics, which can strengthen their critical thinking and analytical skills in social studies and history.
Such a language model would also be useful in teaching, as it could help pupils understand political issues. It could be used as a tool to review parties’ positions, historical decisions and their influence on current political debates. In addition, it could give pupils a more nuanced and objective view of political topics, which can strengthen their critical thinking and analytical skills in social studies and history.
Read about the study here:
Major chatbot is a left-leaning EU supporter
The large AI chatbot Llama Chat, which is a popular alternative to ChatGPT, has a political bias in favour of left-leaning and pro-European views. This is shown by a study from Københavns Universitet, where the researchers tested the model on EU policy questions. The researchers also demonstrate how it is possible to correct the skew.
Microsoft invests DKK 22 billion in AI in Sweden
Microsoft announced a major investment of nearly DKK 22 billion (approximately USD 3.2 billion) in AI development in Sweden. This investment includes expanding existing data centres in Sandviken, Gävle and Staffanstorp, which are to be equipped with 20,000 graphics processors.
“We believe that Sweden has the potential to become a leading AI nation,” says Microsoft’s head in Sweden, Sophia Wikander, in a comment to Berlingske Tidende.
The investment also entails that 250,000 Swedes will undertake Microsoft’s AI training over the next three years. The training will be carried out in collaboration with universities and the Swedish public employment service, where Linköpings Universitet will play a major role through a newly established knowledge council (kunskapsråd).
Microsoft Invests Billions to Make Sweden an AI Powerhouse
Microsoft is investing billions to train Swedes in generative AI.
Camilla Jessen
Press invitation: The Prime Minister holds a press conference together with Microsoft
On Monday 3 June, Prime Minister Ulf Kristersson visits Microsoft Sweden’s head office in Stockholm. In connection with the visit, a joint…
Microsoft in a huge multi-billion initiative: 250,000 Swedes are to be trained in AI
Read more here.
Morten Høyrup
When artificial intelligence becomes part of the guidance and counselling space
Tine Wirenfelt Jensen from Syddansk Universitet (University of Southern Denmark) has written an article for the journal Danish Journal of University Pedagogy (DUT), which discusses how generative artificial intelligence (GAI) can be integrated into supervisory environments. The focus is on the potentials and challenges that GAI brings, including ethical considerations, data protection, and the need to ensure that AI is used in a way that supports supervisors’ work without replacing them.
The article underlines the importance of having clear guidelines and a critical approach to the use of GAI in supervisory contexts.
Read the full article here:
View of: When artificial intelligence becomes part of the supervision space
Documentary about Brian Eno changes with every viewing
Gary Hustwit has created a groundbreaking documentary about the British composer Brian Eno, in which the content changes with each screening using advanced algorithms. This means that every time you watch the film, it adapts to current events, the weather and trends. This is done by giving the algorithm access to a library of pre‑recorded clips and allowing it to change the order to create new sequences. As a result of these changes, the viewer gets a unique experience each time instead of a fixed narrative structure.
It is a new way of thinking about film as a medium, which is otherwise very linear and fixed. This approach can help create dynamic learning materials that adapt to students’ interests and needs.
The BBC has posted a short video about the new documentary on its website, which is worth watching as the entire concept is explained. The BBC item can be used for classroom discussion about artificial intelligence and the documentary film genre.
Watch the movie that rewrites itself
Filmmaker Gary Hustwit has created a documentary which can rewrite itself before every screening.
An ever-adapting film that rewrites itself every time it’s screened
Gary Hustwit’s latest documentary possesses the unique ability to rewrite itself prior to each screening, thanks to advanced algorithms.
Gairika Mitra
New AI video generator challenges OpenAI’s Sora
The Chinese technology company Kuaishou has unveiled KLING, a new AI video generator that could potentially compete with OpenAI’s Sora.
Kling can generate videos up to two minutes long at 1080p resolution and 30 frames per second, and it simulates realistic physical movements. The model also offers advanced 3D face and body reconstruction for improved expressions and movements in the videos.
💡
The development of faster and more user‑friendly AI video generators can give educators new tools to create engaging teaching materials, which can improve the learning experience for students.
It is not yet possible to test Sora yourself, but KLING is currently available as a public demo in China. Although it is uncertain whether KLING will ever become available outside China, its availability could push OpenAI to release Sora sooner. Competition from other companies such as Runway, Pika Labs and Google’s new Veo model may also push development forward.
Below, we have collected some of the best video clips that show KLING’s capabilities:
KLING is the latest AI video generator that could rival OpenAI’s Sora
Chinese tech company Kuaishou has unveiled KLING, a new video generation model. Based on the demos, it could rival OpenAI’s Sora.
Matthias Bastian
This week’s news
New study: Students lack instruction in artificial intelligence
Students are not being taught for the realities of the labour market when they do not learn about artificial intelligence, argues a researcher, who urges KU (University of Copenhagen) to make the new technology part of students’ everyday lives.
Martin Juhl
Is OpenAI Compatible With School?
What does May 2024 tell us about OpenAI and K-12 education?
Tom Mullaney
Between heaven and algorithm: Does artificial intelligence have a place in the Church of Denmark (Folkekirken)?
Artificial intelligence is no longer science fiction. But does it belong in Folkekirken – and if so, how? These are some of the key questions and dilemmas that leading experts in AI will focus on at Folkekirken’s Digitaliseringskonference this autumn.
More schools monitor students’ screens during exams
Text_ Dorthe Kirkgaard Nielsen Photo_ (Archive) Carsten Bundgaard
Vincent F. Hendricks: Generation Z has a strong vision for a world with artificial intelligence
“The Gen Z vision for the world recognises that AI is here to stay – and thus must be both harnessed and restrained at the same time.” Opinion by Vincent F. Hendricks MMXXIV is the year when roughly half of the world’s population, according to a tally by Time Magazine, goes to the polls. In Bangladesh over Bhutan to Belarus, from Uzbekistan [...]
Written by
MSc in IT | Co-founder of Viden.AI | Head of IT and Communications at TietgenSkolen
