Time

The lesson is expected to take around 45 minutes to complete, consisting of 20 min reading and a podcast of 25 min. In addition, there are reflection exercises, which are best done jointly with your fellow students, as well as links to recommended material if you want to know more.

About the lesson: ‘When algorithms decide’

This lesson introduces algorithmic systems, like those you may know from digital services such as Netflix, Spotify, and AI language models such as ChatGPT and Bing Chat. The lesson gives examples of algorithmic systems for ranking content, classification of images, content generation, and decision-making support – and both advantages and disadvantages are presented for reflection and discussion.

The lesson gives insight into how algorithms and artificial intelligence are systems that combine human input, technologies and social structures. It touches on challenges with automated decision-making processes and algorithmic curation of information and AI generated content, and presents a critical approach to the application and use of algorithmic systems in various contexts.

Learning outcomes

When you have completed the lesson, we expect that you:

understand algorithmic curation and automated decision-making processes
can reflect on the advantages and disadvantages of using algorithmic systems in society.

Source

The lesson was produced by the University of Copenhagen as part of its learning resources for digital literacy, 2023.

Course directors:

Christina Neumayer, associate professor, Department of Communication
Camilla Moring, associate professor, Department of Communication

Algorithms and Artificial Intelligence

Let's start by the two key terms in this lesson: Algorithms and artificial intelligence.

Algorithms

There is no clear definition of what an algorithm is, but in its essence it is a set of rules (also called a model) based on data, that then defines a sequences of operations. The impact of algorithms depends on the context it is used in which can range from apps on mobile phones, automated decision making, costumer evaluation in a bank to recommender systems on Netflix or social media. To reflect on the consequences algorithms may have in society, we need to understand them in their context.

Source: Kristian Lumarchive and Rumman Chowdhury (2021, 6 Feb.), MIT Technology Review. What is an “algorithm”? It depends on whom you ask

Artificial intelligence (AI)

Artificial intelligence (AI) is commonly understood as smart machines performing tasks that otherwise require human intelligence. There are various forms of artificial intelligence ranging from image recognition, chat bots (such as GPT-3), facial recognition, surveillance systems to deep fake videos. While narratives of AI often endow computer-powered analysis with human intelligence or beyond that with analytical power that humans cannot perform, AI always serves beliefs and perspectives of people. Similar to algorithms, we need to study AI in the context of its use to understand its larger societal consequences.

Source: Thomas Bolander (2018). Videnskab.dk. Hvad er kunstig intelligens egentlig?

____________________________________________________________

Cases and exercises

There are many definitions for algorithms and and artificial intelligence, and none of them is universally true. To reflect upon the impact algorithmic systems have in society, we need to understand them in context. In the following, we take a closer look at some examples and cases that will allow us to move towards critical reflection. Each case introduces a specific perspective and is followed by a short exercise that will enable you to reflect on consequences of algorithmic systems.

Case#1: Netflix – recommendations and ranking

Netflix recommender system.png

The streaming service Netflix's recommender system is one of the most commonly used examples for how algorithmic systems influence our movie and TV-series consumption. Key to the recommender system is personalisation, so the system recommends content that is adjusted to our own consumption patterns. Based on data from all the other Netflix users who show similar consumption patterns the system outputs personalised recommendations. The recommender system filters based on various themes (e.g., 'Trending', 'Horror') or criteria (e.g., 'romance', 'violence', Danish productions) and calculates the probability that we keep watching on the platform.

For a deeper dive into the recommender system, read David Chong (2020). The Medium. Deep Dive into Netflix’s Recommender System

Exercise: Look at your profile in one of the streaming services you use (e.g., Netflix, Spotify, Amazon Prime, HBO) and think about the recommendations (for films, music, etc.) you get. Compare them with someone else's profile (e.g., fellow student, friend, family member) and discuss the differences.

Case #2: Images and machine learning – classification

Students in class Sam Balye Unsplash.png

Photo by Sam Balye on Unsplash

One of the use areas, where the societal consequences of algorithmic systems perhaps become most apparent is computer vision, a field of artificial intelligence that trains AI to interpret and understand the visual world by accurately identifying and classifying visual objects. In other words, we train the AI to 'see'. Such algorithmic systems often learn based on data that predominantly represent white male perspectives (as engineers and designers of such systems mostly are white men) which leads to consequences such as racial or gender bias. The perhaps most prominent example of such racial bias is the image recognition algorithm in Google Photos classifying and labelling black people as “gorillas.” Instead of investing in better training data and improving the algorithmic system for being able to classify black people, Google fixed the problem by removing "gorilla" and similar labels altogether.

For the full story, read James Vincent (2018, Jan 12). The Verge. Google ‘fixed’ its racist algorithm by removing gorillas from its image-labeling tech.

For more about how politics are built into AI systems based on visual data, read Kate Crawford and Trevor Paglen (2019). Excavating AI: The Politics of Images in Machine Learning Training Sets.

Exercise: Do a Google image search with different keywords and reflect on the results. Which biases and stereotypes are reproduced in the images? (e.g., if you search for 'professor', 'expats' vs. 'refugees')

Case #3: Artistic research with AI – creation

Screenshot 2023-01-30 at 14.52.07.png

Image from Rine Rodin, The Wise and The Mad (2021), Performance for video. 18 minute performance.
Digital. Color. Sound. Courtesy of the artist. Viewable at AIPerformance.space.

We can also create with algorithmic systems. AI has consequences for creative processes and conversely, we can learn about the consequences of AI through art. Mirabelle Jones is an artist and PhD researcher at the department of Computer Science at University of Copenhagen, and they curate the AI Performance Space which "brings together artists of different practices, backgrounds, identities, locations and methods to collectively investigate the influence artificial intelligence can have on the creative process and consider how creative practices can in turn develop new insights and approaches to artificial intelligence."

Exercise: Experiment with Crayion to create your own visual art with AI.

Case #4: Algorithmic decision making in the public sector

What happens when algorithmic systems take decisions in the public sector instead of humans? Algorithmic systems are increasingly used across the public sector including criminal justice, education, or provision of welfare benefits. This can have consequences for privacy of individuals, as these decisions are often made based on fine-grained personal data; accountability, as it is unclear who is held accountable when the systems fail and work incorrectly (e.g., wrongly classify someone as a criminal); or for democratic processes (e.g., social media bots in elections).

If you want to know more about algorithmic decision making in the Nordics, read Algorithmic Decision Making Nordic lead by Assoc.Prof. Stine Lomborg at the Department of Communication at Unversity of Copenhagen.

For more about the consequences of algorithmic systems on democracy in Denmark, have a look at the examples from the Algorithms, Data & Democracy project lead by Sine Nørholm Just, Roskilde University.

Exercise: Read the news and look for stories where algorithmic systems take human decisions. Reflect about the consequences for the specific sector in which automated decision making takes place.

Case #5: ChatGPT and other AI language models – When algorithms express themselves as humans

ChatGPT - screenshot.png

What does it mean that we now have machines that can express themselves so convincingly that one might easily believe the words are coming from a living and thinking human? Will we lose the ability to devise arguments and express ourselves when computers can do the work for us? And what does it mean for Danish as a small language and culture that the largest and most advanced language models are developed in the USA?

Listen to an interview with Tanya Karoli Christensen, professor of Danish language and linguistics at the Department of Nordic Studies and Linguistics. The interview was recorded in December 2023 and is available in Danish. For English version, please read this transcript.

If you need an explanation of what AI language models are and how they work, check out the lesson ‘The Technology behind ChatGPT and Large Language Models’.

Exercise: Discuss with your fellow students what impact large language models will have on your field of study and your future profession.

____________________________________________________________

Learn more

Hear about implications of the human-like chatbot, ChatGPT (in Danish), DR Genstart: En menneskelig maskine, 6 January 2023

Read about Nordic Perspectives on Algorithmic Systems, NOS-HS network lead by Assoc. Prof. Airi Lampinen, Stockholm University

Notes from the first workshop on metaphors and concepts
Notes from the second workshop on methods and approaches for studying algorithmic systems

Read more about recommender systems, IEEE Information Theory Society, How Recommender Systems Work

Hear more about the politics of AI. Atlas of AI, Kate Crawford in conversation with Judy Wajcman

Learn about your interactions with AI through playful mobile Apps. Ridiculous Software Project lead by Prof. Miguel Sicart, Center of Digital Play, IT University of Copenhagen, Apps available for free (iPad and iPhone):

____________________________________________________________