Explore the dilemmas

AI raises questions that don’t always have easy answers. Revisit the dilemmas from the AI Games Café, see where you stand, and explore the research that inspired them.

Explore the 7 dilemmas below!

Dilemma 1: Prediction ethics

A company has developed AI that claims to predict a person’s sexual orientation from a photograph. It performs better than humans in a research experiment. The AI is trained using photographs that people have publicly posted online on a dating website. The company argues: “If people made their photos public through a dating website, we’re allowed to use them.”

Do you agree? Why/why not?

Now consider what happens if this AI technology can be used by countries where homosexuality is prohibited to identify LGBTQ+ people without their consent. Does that change your opinion?

Source:

Dilemma author: Dr. Masuma Shahid, Assistant Professor at EUR in International and EU Law, Researcher AI, Queer Data and LGBTQ+ Rights

Dilemma 2: Inclusive pride representation

Text-to-image generative AI models often show biases and stereotypes in how they depict people of a certain race, gender, culture, or sexual identity. In an era of AI-generated marketing and social media content, these depictions do not only reflect reality but also shape it. Viewing AI-generated images can strengthen a stereotype in the audience.

You work for a city tourism board. An AI tool creates an image for Pride Month. It shows a young, white, slim couple with rainbow clothing and purple hair. The image is attractive and technically flawless, but it does not reflect the city’s diverse LGBTQ+ population.

What do you do? Explain your decision.

Source:

Dilemma author: Dr. Masuma Shahid, Assistant Professor at EUR in International and EU Law, Researcher AI, Queer Data and LGBTQ+ Rights

Dilemma 3: AI (de)growth

Tech development often follows and supports the economic model of exponential growth (at all costs, and above all else). Bigger, better, faster, more functionalities, more data, more users… and more profit! Yet this model requires exponentially more planetary resources. More lithium extracted from mines. More data centres set up. More microchips manufactured and shipped across the world. More local water used to cool servers. Surely, such exponential growth is not sustainable.

Should we build the most powerful AI we can, even if it requires more of the planet’s resources? Why/why not?

If AI keeps growing, what are we willing to give up for it?

 

Easier to understand version: Can AI keep growing forever?

Right now, most technology is built around one idea: growth above all else. Bigger, better, faster, with more features, more data, more users… and more profit! But to keep growing like that, we need more and more resources from our planet. More lithium dug out of the ground. More data centres (a sort of computer warehouses). More computer chips made and shipped all over the world. More water used to keep servers cool. This can’t go on forever and at some point, we will run out of room to keep growing.

Should we build the most powerful AI we can, even if it requires more of the planet’s resources? Why/why not?

Source:

Dilemma author: Adriaan Odendaal, EUR PhD Candidate, Erasmus Initiative Societal Impact of AI (AiPact)

Dilemma 4: Alternative datasets

The current generation of generative AI tools like ChatGPT is trained on historical text-based datasets selected by its developers. The resulting model seems to be able to do everything. It can create poems in the style of Shakespeare or give you advice imitating Sigmund Freud. But datasets are as much about what is excluded as they are about what is included. So much of human knowledge is left out of these AI models, either because they weren’t thought to be important, or they didn’t exist in a readily available text-based form.

You are an AI engineer who must decide which data sets are used during the training process. What do you prioritise when choosing training data resources?

 

Easier to understand version: Who does AI learn from?

Tools like ChatGPT learn by reading huge amounts of text that people chose to feed into them. Because of this, AI can do a lot of impressive things, like write a poem in Shakespeare’s style or explain complex concepts. But this process isn’t just about what is included, but also about what got left out. A big part of human knowledge and experience never made it into these AI models, because the people building them didn’t think it mattered, or because that knowledge was never written down in a way computers could read.

What knowledge would you prioritise if you were to build your own AI model?

Source:

Dilemma author: Adriaan Odendaal, EUR PhD Candidate, Erasmus Initiative Societal Impact of AI (AiPact)

Dilemma 5: AI science simplification

AI gives opportunities to make science more accessible to wider audiences, but that comes with a risk of oversimplifying scientific facts in the explanations given to users. Relying on quick AI summaries can lead to missing important nuances or to a surface-level understanding of complex problems. Research often contains crucial limitations, which an AI model can skip over, showing only part of the picture.

Should we use AI to simplify complex science for wider audiences, knowing it might gloss over nuance or uncertainty, or keep more complex explanations that risk excluding non-experts? Explain your reasoning. Have you experienced a similar situation?

Source:

Dilemma author: Dr. Charlotte Bruns; School of History, Culture, and Communication, Erasmus University Rotterdam

Dilemma 6: Chatbot-driven delusion

“AI psychosis” is an informal term for situations where conversations with AI appear to reinforce beliefs that become increasingly disconnected from reality. AI can contribute to this because it tends to agree with users and build on their ideas rather than question them.

Your flatmate Sam (21) keeps noticing the words “wake up” in a song, an advert, a TikTok video and then on a poster on his way to university. At first, he jokes about the coincidence. But after talking about it with an AI chatbot which helps him look for more patterns, Sam becomes convinced there is a hidden message he needs to uncover.

He has stopped seeing friends, sleeps only four hours a night and spends most of his time investigating the pattern with the chatbot. Sam says he feels great and that things have never made more sense.

Do you think the chatbot contributed to Sam’s behaviour? How would you address this situation?

Source:

Dilemma author: Student assistant team AIDD Convergence

Dilemma 7: Online data privacy

Everyone in Emma’s class is talking about StarMe, a new free app that turns you into a superhero in a short movie. You upload ten photos of your face and record yourself saying one sentence. The app then makes a video where you fly over the city and talk in your own voice. Emma’s best friend has already made one, and it looks amazing.

One evening, her older cousin Liam tells her: “If an app is free, you’re paying with something else. They keep your face and your voice.” Emma opens the app’s rules. They are 38 pages long. Somewhere in the middle, it says the company may use what you upload “to improve our products.”

Should Emma use StarMe? What is she giving the company, and what does she get back? If all her friends are using it, does it change your opinion?

 

Source:

Dilemma author: Student assistant team AIDD Convergence