Case 5: AI Act
Class European platform policies
Objectives
Explain why the EU felt it needed an AI Act
Describe the EU decision-making process that led to the AI Act and where
the main conflicts were
Summarise the core structure of the AI Act (scope, risk tiers, governance,
enforcement timeline)
Classifiy simple AI use cases into risk-based categories
Explain how ChatGPT (and similar tools) are regulated under the Act
Critically discuss advantages and drawback of the AI Act
→ especially for digital media and creative industries
Concepts and context
Different AI uses:
Generative AI
In the medical field
HR (to hire people)
Eductional
Creative industries
→ transcription of AV-content, AI generated songs
Governments
→ migration, law enforcement
⇒ inequalities & bias embedded in AI (especially LLM) + however people
often assume it is objective/factual
⇒ which forms of AI use would be so risky that they need to be regulated
(eg law enforcement, surveillance, social scoring)
Case 5: AI Act 1
, Context
AI = omnipresent
→ very difficult to regulate it (+ a lot of overlap/conflicts with other
regulations)
Problem = foundational models (eg ChatGPT) only became popular during
the process of making the Act
→ additional guidelines should be issued in the next years
Machine learning value chain
Difference between AI used in particular services & used for operating
systems and processes
→ are often interlinked, but very different to regulate
Examples of AI in real life
→ Virtual Assistants, autonomous vehicles, healthcare, smart home
devices, finance and banking, robotics, gaming & entertainment, language
translation, education, retail, search engines
→ prominent policy discussions = link between AI and copyright (human-
machine-human), special effects in movies
EU legislation = try to harmonise + not cause fragmentation
→ you don’t want to limit the market, but also mitigate the risk (discussion
of over- vs underregulating)
Fundamental rights risks
Mass biometric surveillance, emotion recognition at work/school
Social scoring (China comparisons)
Predictive policing, discriminatory profiling
Case 5: AI Act 2
, AI systems deployed in high-stakes decisions w/o transparency or
accountability
→ eg hiring, grading, credit, policing
Democratic + media risks
Deepfakes, synthetic media in elections
Disinformation campaigns supercharged by generative AI
Manipulative recommender systems and ‘dark patterns’
Market + innovation risks
Fragmentation: 27 MS inventing their own AI rules
Powerful US and Chinese firms setting de-facto standards
Need for legal certainty so European companies dare to use AI
AI categories
Two broad categories of AI technologies
Artificial narrow intelligence (ANI) Artificial general intelligence (AGI)
weak AI strong AI
designed to perform a wide range of
image and speech recognition systems
intelligent tasks
GenAI = trained on broad set of unlabelled
trained on well-labelled datasets
data with minimal fine-tuning
perform specific tasks & operate within
think abstractly & adapt to new situation
predefined enivronment
From ANI to AGI?
→ achieving machines with human-level intelligence still require these
critical skills
Visual perception
→ current AI struggle with context, color and understanding how to
react to partially hidden objects
Audio perception
→ cannot reliably understand accents, sarcasm and other emotional
speech tones
→ difficulty filtering out unimportant background noise
Case 5: AI Act 3
Class European platform policies
Objectives
Explain why the EU felt it needed an AI Act
Describe the EU decision-making process that led to the AI Act and where
the main conflicts were
Summarise the core structure of the AI Act (scope, risk tiers, governance,
enforcement timeline)
Classifiy simple AI use cases into risk-based categories
Explain how ChatGPT (and similar tools) are regulated under the Act
Critically discuss advantages and drawback of the AI Act
→ especially for digital media and creative industries
Concepts and context
Different AI uses:
Generative AI
In the medical field
HR (to hire people)
Eductional
Creative industries
→ transcription of AV-content, AI generated songs
Governments
→ migration, law enforcement
⇒ inequalities & bias embedded in AI (especially LLM) + however people
often assume it is objective/factual
⇒ which forms of AI use would be so risky that they need to be regulated
(eg law enforcement, surveillance, social scoring)
Case 5: AI Act 1
, Context
AI = omnipresent
→ very difficult to regulate it (+ a lot of overlap/conflicts with other
regulations)
Problem = foundational models (eg ChatGPT) only became popular during
the process of making the Act
→ additional guidelines should be issued in the next years
Machine learning value chain
Difference between AI used in particular services & used for operating
systems and processes
→ are often interlinked, but very different to regulate
Examples of AI in real life
→ Virtual Assistants, autonomous vehicles, healthcare, smart home
devices, finance and banking, robotics, gaming & entertainment, language
translation, education, retail, search engines
→ prominent policy discussions = link between AI and copyright (human-
machine-human), special effects in movies
EU legislation = try to harmonise + not cause fragmentation
→ you don’t want to limit the market, but also mitigate the risk (discussion
of over- vs underregulating)
Fundamental rights risks
Mass biometric surveillance, emotion recognition at work/school
Social scoring (China comparisons)
Predictive policing, discriminatory profiling
Case 5: AI Act 2
, AI systems deployed in high-stakes decisions w/o transparency or
accountability
→ eg hiring, grading, credit, policing
Democratic + media risks
Deepfakes, synthetic media in elections
Disinformation campaigns supercharged by generative AI
Manipulative recommender systems and ‘dark patterns’
Market + innovation risks
Fragmentation: 27 MS inventing their own AI rules
Powerful US and Chinese firms setting de-facto standards
Need for legal certainty so European companies dare to use AI
AI categories
Two broad categories of AI technologies
Artificial narrow intelligence (ANI) Artificial general intelligence (AGI)
weak AI strong AI
designed to perform a wide range of
image and speech recognition systems
intelligent tasks
GenAI = trained on broad set of unlabelled
trained on well-labelled datasets
data with minimal fine-tuning
perform specific tasks & operate within
think abstractly & adapt to new situation
predefined enivronment
From ANI to AGI?
→ achieving machines with human-level intelligence still require these
critical skills
Visual perception
→ current AI struggle with context, color and understanding how to
react to partially hidden objects
Audio perception
→ cannot reliably understand accents, sarcasm and other emotional
speech tones
→ difficulty filtering out unimportant background noise
Case 5: AI Act 3
