L1
·
Quiz
·
Lab
L2
·
Quiz
·
Lab
L3
·
Quiz
·
Lab
L4
·
Quiz
·
Lab
Module Test
Module 7 · Lesson 1

The Law Catches Up: GDPR and the Right to Be Forgotten

Europe's landmark privacy regulation gave individuals enforceable rights against AI-driven data systems — and forced companies worldwide to change how they operate.
What can a law actually force a trillion-dollar tech company to do with your data?

In May 2014, the Court of Justice of the European Union ruled on a case brought by a Spanish lawyer named Mario Costeja González. He had asked Google to remove links to a 1998 newspaper notice about a debt he had long since paid. Google refused. The court ruled in his favour — establishing that EU citizens have a right to request delisting of outdated, irrelevant personal information. Google was ordered to comply.

The ruling didn't erase the original newspaper article. It required Google's search index — an AI-driven ranking and retrieval system — to stop surfacing it as a prominent result for his name. The distinction matters: the right applied to the automated system processing the data, not the data's original publication.

What GDPR Actually Is

The General Data Protection Regulation took effect on 25 May 2018 across all EU member states. It is not a voluntary industry standard or a set of recommendations. It is binding law with fines of up to €20 million or 4% of global annual revenue — whichever is higher. For Google, Apple, or Meta, that second figure runs into billions of euros.

GDPR applies to any organisation that processes the personal data of EU residents, regardless of where the organisation is headquartered. A company in California, Singapore, or São Paulo that collects data from someone in Berlin is subject to GDPR. This extraterritorial reach was new and deliberate.

The regulation defines personal data broadly: any information that can identify a living individual directly or indirectly. That includes names, IP addresses, location data, cookie identifiers, biometric data, and inferred characteristics — the kind of outputs AI profiling systems produce.

The Six Rights GDPR Grants You

GDPR created a specific set of individual rights. These are not aspirational goals. They are enforceable entitlements that organisations must honour within defined time limits (generally 30 days).

  • 01Right of Access: You can request a copy of all personal data an organisation holds about you, including how it was obtained, why it is processed, and who it has been shared with.
  • 02Right to Rectification: You can demand that inaccurate data be corrected. If a credit scoring AI has a wrong address or a misclassified behaviour, you can require a fix.
  • 03Right to Erasure ("Right to Be Forgotten"): In specific circumstances — data no longer needed, consent withdrawn, unlawful processing — you can demand deletion.
  • 04Right to Restriction: You can limit how your data is used while a dispute is resolved, preventing deletion while also preventing active use.
  • 05Right to Data Portability: You can receive your data in a machine-readable format and transfer it to another service.
  • 06Right to Object: You can object to processing for direct marketing purposes, and AI systems must stop. For other purposes, you can object and the organisation must demonstrate compelling legitimate grounds to continue.
Article 22: Rights Against Automated Decision-Making

Article 22 is the provision that most directly targets AI. It gives individuals the right not to be subject to a decision based solely on automated processing that produces legal or similarly significant effects. Loan approvals, job application screening, insurance pricing, parole risk scoring — if these are done by an algorithm with no human review, Article 22 may apply.

When automated decision-making is allowed (because you consented, or it is necessary for a contract, or it is permitted by law), the organisation must provide meaningful information about the logic involved, allow you to express your point of view, and ensure human review is available. A black-box algorithm that renders a verdict with no explanation is not compliant.

Real Enforcement · 2023

Ireland's Data Protection Commission fined Meta €1.2 billion in May 2023 under GDPR for transferring EU user data to US servers without adequate protection. It was the largest GDPR fine issued to that point. The case directly implicated the AI systems Meta uses for ad targeting and content ranking — those systems require the data transfers to function.

The Limits of GDPR

GDPR does not apply outside the EU/EEA by default. A US citizen using a US-based service has no GDPR rights. GDPR also has broad exemptions for national security, law enforcement, and journalistic purposes. And enforcement is patchy — complaints to Data Protection Authorities (DPAs) can take years to resolve.

Perhaps most importantly, GDPR says little about AI systems that infer new information about you from seemingly innocuous inputs. If a retailer infers your political views from your purchase history, that inference may not be "personal data about your political views" in the traditional sense — it is a derived output. Courts and regulators are still working out how the law applies to inference engines.

Key Takeaway

GDPR is the most comprehensive privacy law governing AI data practices in existence. It grants six enforceable rights, imposes real financial penalties, applies globally to EU residents' data, and specifically addresses automated decision-making through Article 22. Its limits — no coverage outside the EU, exemptions, slow enforcement — matter as much as its protections.

Lesson 1 Quiz

GDPR and the Right to Be Forgotten — 4 questions
1. In the González v. Google case, what did the CJEU order?
Correct. The ruling targeted Google's search index — an automated retrieval system — not the original publication. The newspaper article still exists; it just cannot be prominently surfaced by the algorithm for searches of his name.
Not quite. The CJEU's ruling was specifically about Google's search index as an automated processing system, not about the original publication. The right applied to the delisting of results, not deletion of the source material.
2. What is the maximum financial penalty under GDPR?
Correct. The €20 million / 4% global revenue threshold makes GDPR financially meaningful even for the largest technology companies. Meta's 2023 fine of €1.2 billion was calculated on this basis.
Incorrect. GDPR penalties go up to €20 million or 4% of global annual revenue. For large companies, 4% of revenue is the relevant figure — for Meta that ran into billions of euros.
3. Which GDPR article most directly addresses automated decision-making by AI systems?
Correct. Article 22 specifically addresses decisions made solely by automated processing — including AI profiling — that produce legal or similarly significant effects. It requires human review to be available and meaningful explanation of the logic.
Not quite. Article 22 is the provision targeting automated decisions. It gives you the right not to be subject to purely algorithmic decisions with major consequences, and requires organisations to offer human review.
4. A US-based company collects data from EU residents through its website. Does GDPR apply to it?
Correct. GDPR has deliberate extraterritorial reach. Any organisation offering goods or services to EU residents, or monitoring their behaviour, falls under GDPR regardless of where the organisation is based.
Incorrect. GDPR's extraterritorial scope is one of its defining features. It applies to any organisation processing EU residents' data — location, revenue, or office presence in the EU are not the determining factors.

Lab 1: Mapping Your GDPR Rights

Practice applying GDPR rights to real scenarios — discuss with the AI tutor

Scenario-Based Rights Identification

You will describe a scenario involving personal data or AI decision-making, and the AI tutor will help you identify which GDPR rights apply, what the organisation's obligations are, and what limits may prevent you from exercising those rights.

Complete at least 3 exchanges to finish this lab.

Start by describing a situation — for example: "A recruiter told me their software automatically rejected my CV before a human saw it" or "A bank's app gave me a lower credit score than I expected and won't explain why."
GDPR Rights Tutor
AI TUTOR
0 / 3
Welcome to Lab 1. I'm here to help you map GDPR rights to real-world scenarios. Describe a situation — your own, hypothetical, or something you've read about — involving personal data processing or an AI decision that affected someone, and I'll walk you through which rights apply and what the law actually requires.
Module 7 · Lesson 2

The American Patchwork: Privacy Law Without a Federal Standard

The United States has no single federal privacy law. What protects you depends on which state you live in, which industry collected your data, and which Congress happened to pass a law for.
Why does the country that invented most of the AI being used globally have almost no enforceable privacy rights for its own citizens?

In February 2021, Facebook's parent company agreed to pay $650 million to settle a class action lawsuit under the Illinois Biometric Information Privacy Act (BIPA). The lawsuit — Patel v. Facebook — alleged that Facebook's face-recognition "Tag Suggestions" feature had collected biometric identifiers (facial geometry) from photos without obtaining written consent, as BIPA requires.

The settlement covered approximately 1.6 million Illinois users. Users in California, New York, or Texas who were affected by the same feature received nothing — because those states had no equivalent law. The same AI system. The same data collection. Different legal outcomes based entirely on geography.

The Sectoral Approach

Unlike GDPR's comprehensive framework, US federal privacy law is sectoral: different rules apply to different industries. The result is a collection of narrow statutes that often have gaps between them through which AI systems readily pass.

HIPAA · 1996
Health Insurance Portability and Accountability Act
Protects medical records held by covered entities (hospitals, insurers, doctors). Does not cover health data collected by fitness apps, wearables, or consumer AI tools.
FCRA · 1970
Fair Credit Reporting Act
Governs consumer reporting agencies. Gives you the right to see your credit report and dispute errors. Does not cover the raw inputs that credit-scoring AI models use.
COPPA · 1998
Children's Online Privacy Protection Act
Requires parental consent before collecting data from children under 13. Frequently circumvented by platforms that technically prohibit under-13 use but take no steps to verify age.
FERPA · 1974
Family Educational Rights and Privacy Act
Protects education records. Applies to institutions receiving federal funding. Does not cover EdTech companies' own data collection practices in the same classroom.
ECPA · 1986
Electronic Communications Privacy Act
Regulates government access to electronic communications. Written before the web existed; struggles to address cloud storage, AI-analysed metadata, and persistent tracking.
FTC Act · Section 5
Federal Trade Commission Authority
Prohibits "unfair or deceptive acts or practices." The FTC uses this to pursue privacy violations — but only after harm occurs, not proactively, and without specific AI rules.
State Laws Filling the Gap

With no federal standard, states have legislated independently. The result is a growing patchwork that compliance teams at large companies now navigate state by state.

2008
Illinois BIPA — first law in the US requiring written consent for biometric data collection. Led directly to the $650 million Facebook settlement and numerous ongoing cases against employers using fingerprint timeclocks.
2018
California Consumer Privacy Act (CCPA) — gave California residents rights to know, delete, and opt out of the sale of personal data. Often described as the closest US equivalent to GDPR, though notably weaker on consent and automated decisions.
2020
California Proposition 24 / CPRA — strengthened CCPA by adding a right to correct inaccurate data, expanding opt-out rights to data sharing (not just sale), and creating the California Privacy Protection Agency as an independent enforcement body.
2021–2024
Virginia, Colorado, Connecticut, Texas, Oregon, Montana and more than a dozen other states passed comprehensive privacy laws — each with different thresholds, exemptions, and enforcement mechanisms.
Why No Federal Law?

Multiple federal privacy bills have been introduced in Congress. The American Data Privacy and Protection Act (ADPPA) passed out of committee in 2022 with bipartisan support but stalled on the floor, partly due to disputes over whether it would preempt stronger state laws like CCPA. California's congressional delegation opposed federal preemption that would weaken Californians' existing protections.

Industry groups generally prefer federal law to a patchwork of 50 state laws — provided the federal standard is weaker than the strongest state laws. Civil liberties groups prefer maintaining stronger state laws. Neither side has built sufficient coalition to move legislation.

The Practical Gap

In practice, Americans' data rights depend heavily on which state they live in. An Illinois resident has stronger biometric protections than a Florida resident. A California resident can opt out of data sales; a resident of most other states cannot. AI systems process data from all 50 states simultaneously — the legal patchwork they navigate is entirely a product of legislative history, not any coherent theory of privacy rights.

BIPA Enforcement: What Real Penalties Look Like

Illinois BIPA creates a private right of action — individuals can sue without waiting for a regulator. Violations carry statutory damages of $1,000 per negligent violation or $5,000 per intentional or reckless violation. Because AI systems can collect biometric data at scale, individual violations aggregate quickly. In addition to Facebook, BIPA cases have been brought against Clearview AI (facial recognition database), various employers using fingerprint timeclocks, and retail stores using gait analysis and facial matching. The Illinois Supreme Court ruled in 2023 in Cothron v. White Castle that each individual biometric scan constitutes a separate violation — a ruling that dramatically increased potential liability for companies using biometric AI.

Lesson 2 Quiz

US Privacy Law Patchwork — 4 questions
1. Why did the $650 million Facebook BIPA settlement only apply to Illinois users?
Correct. BIPA is an Illinois state law. Users in states without equivalent legislation had no legal basis for a claim, even though the same AI feature collected their facial geometry. The same data practice, different legal outcomes based on geography.
Incorrect. The same Face Recognition feature operated nationally. Only Illinois users could claim under BIPA because it is an Illinois law — users in other states lacked equivalent legal protection regardless of the identical data collection.
2. What does "sectoral" mean in the context of US federal privacy law?
Correct. The US has HIPAA for health, FCRA for credit, COPPA for children's data, FERPA for education — each a narrow statute for a specific industry. AI systems that span multiple sectors or operate between them may fall through the gaps entirely.
Not quite. "Sectoral" means each industry has its own specific statute rather than one law covering everyone. The problem is that AI data flows don't respect sector boundaries — a fitness app collecting health data doesn't fall under HIPAA because it isn't a covered entity.
3. What did the Illinois Supreme Court rule in Cothron v. White Castle (2023)?
Correct. This ruling significantly expanded liability for companies using biometric AI — a fingerprint timeclock scanning 100 employees daily could accumulate violations rapidly, as each scan counts separately.
Incorrect. The Cothron ruling held the opposite: each scan is a separate violation, making aggregate liability potentially enormous for companies using biometric systems at scale without BIPA-compliant consent.
4. The American Data Privacy and Protection Act (ADPPA) stalled in Congress partly because of disagreement over what issue?
Correct. California legislators opposed a federal law that would reduce their constituents' existing protections under CCPA/CPRA. Industry generally preferred federal preemption of stronger state laws; privacy advocates opposed it. This coalition problem has blocked federal legislation repeatedly.
Not quite. The central sticking point was federal preemption — whether a federal law would override stronger state privacy laws. California's representatives refused to support a federal standard weaker than their state's existing protections.

Lab 2: Navigating the US Privacy Patchwork

Work through real data-rights scenarios under US law — which statute applies, and where the gaps are

Statute Identification Exercise

You will describe a data collection or AI decision scenario, and the tutor will help you identify which US statute — if any — applies, what rights it creates, and what gaps exist. Many scenarios will fall through the patchwork entirely.

Complete at least 3 exchanges to finish this lab.

Try a scenario like: "My health insurance company uses an AI to analyse my fitness tracker data to set my premiums" or "My employer uses a fingerprint scanner to track when I start and end work shifts."
US Privacy Law Navigator
AI TUTOR
0 / 3
Welcome to Lab 2. I'll help you navigate the US privacy law patchwork. Describe a scenario involving data collection or an AI system making decisions about people, and we'll work through which federal or state statute might apply — and just as importantly, where the legal gaps are that leave people unprotected.
Module 7 · Lesson 3

Algorithmic Accountability: Fair Lending, Hiring, and the Civil Rights Laws AI Must Navigate

AI systems making consequential decisions about people face existing civil rights law — and regulators are beginning to argue that algorithmic discrimination is still discrimination.
If an AI rejects you for a loan or a job without a human ever reviewing your file, can existing anti-discrimination law reach it?

In March 2019, the US Department of Housing and Urban Development filed a formal complaint against Facebook alleging that its ad-targeting algorithm violated the Fair Housing Act. HUD charged that Facebook's system allowed advertisers to exclude users from seeing housing ads based on race, religion, national origin, sex, familial status, and disability — demographic categories derived from user data and behavioural inference.

Facebook had already settled with the National Fair Housing Alliance and other civil rights groups in March 2019 for a separate but related case, agreeing to overhaul its ad-targeting system for housing, employment, and credit. In August 2019, Facebook settled with HUD, agreeing to build a new system that limited demographic targeting for those ad categories and created a searchable Ad Library for housing-related ads. The case was notable because the discrimination was not programmed intentionally — it emerged from the AI's optimisation logic.

The Civil Rights Laws That Apply to Algorithmic Systems

Several decades-old laws written long before AI existed nonetheless apply to algorithmic decision-making. Regulators and courts are now actively applying them.

  • FHAFair Housing Act (1968): Prohibits discrimination in the sale, rental, and financing of housing based on race, colour, national origin, religion, sex, familial status, or disability. The Facebook ad case showed this applies to AI-driven ad delivery systems.
  • ECOAEqual Credit Opportunity Act (1974): Prohibits credit discrimination based on race, colour, religion, national origin, sex, marital status, age, or because you receive public assistance. AI credit-scoring systems must comply — and in 2023 the CFPB issued guidance requiring explainability for adverse actions.
  • Title VIICivil Rights Act, Title VII (1964): Prohibits employment discrimination based on race, colour, religion, sex, or national origin. The EEOC published guidance in May 2023 confirming that employers remain liable under Title VII when AI tools they use discriminate, even if the AI was provided by a third party.
  • ADAAmericans with Disabilities Act (1990): Prohibits disability discrimination. The EEOC has specifically addressed AI video interview analysis tools that may screen out candidates with disabilities in ways unrelated to job requirements.
Disparate Impact: The Legal Theory That Reaches Unintentional AI Discrimination

US civil rights law recognises two theories of discrimination. Disparate treatment requires proving intent — someone was deliberately treated worse because of a protected characteristic. Disparate impact does not require intent: if a neutral policy or practice has a disproportionately negative effect on a protected group, and there is no business necessity justification, it can be unlawful.

This matters enormously for AI. An algorithm trained on historical data will encode the patterns in that data. If historically Black applicants were denied loans at higher rates — even for illegitimate reasons — a model trained on those approvals will learn to reject similar-looking applicants. The algorithm wasn't told to discriminate; it learned to from the training data. Disparate impact doctrine says that does not matter: the effect counts.

CFPB v. Algorithmic Credit Scoring · Ongoing Enforcement

In 2023, the Consumer Financial Protection Bureau issued guidance making clear that creditors using AI models to make lending decisions cannot cite "too complex to explain" as a reason for adverse action notices. If an AI denies credit, the applicant must receive a specific, accurate reason — not a generic one. This directly challenges black-box AI models that cannot generate human-readable explanations of individual decisions.

EEOC Guidance on AI Hiring Tools · May 2023

The Equal Employment Opportunity Commission released guidance stating that employers who use AI hiring tools from third-party vendors are not insulated from liability under Title VII. If the tool screens out disproportionate numbers of candidates based on a protected characteristic, the employer is liable — and may not be able to use "we just used the vendor's algorithm" as a defence. The vendor may also face liability.

The Explainability Requirement and Adverse Action Notices

Under ECOA and the FCRA, when a creditor takes adverse action based on a credit report or algorithmic score, the applicant must receive a notice listing the specific reasons. Historically, lenders used standardised reason codes. The CFPB's 2023 guidance updated this: AI-generated reasons must be accurate and specific to the individual, not generic. A notice that says "credit score too low" when the actual reason was an AI-detected pattern in purchase history is non-compliant.

This creates a direct tension between commercially deployed black-box AI models and US credit law. Several major lenders have faced CFPB investigations over the adequacy of AI-generated adverse action explanations. The requirement is creating pressure to use more interpretable models or to layer explainability systems on top of complex ones.

The Core Principle

Algorithmic discrimination is still discrimination. Existing civil rights law does not have an "AI exception." Regulators at HUD, CFPB, EEOC, and FTC have all signalled active enforcement. The legal risk for organisations deploying AI in housing, credit, and employment without bias auditing and explainability is no longer theoretical.

Lesson 3 Quiz

Algorithmic Accountability and Civil Rights Law — 4 questions
1. The HUD complaint against Facebook's ad algorithm alleged discrimination under which law?
Correct. HUD's 2019 complaint alleged that Facebook's targeted advertising AI violated the Fair Housing Act by allowing advertisers to exclude users from housing ads based on protected characteristics. Facebook settled and overhauled its housing, employment, and credit ad targeting systems.
Not quite. The case was brought under the Fair Housing Act, which prohibits discrimination in housing-related transactions including advertising. HUD filed the formal complaint; Facebook had already settled with civil rights organisations in a parallel case.
2. What is "disparate impact" and why does it matter for AI systems?
Correct. Disparate impact doctrine is crucial for AI accountability because AI systems trained on biased historical data will reproduce those biases — without any intent. The legal question is not whether the AI "meant" to discriminate but whether its outputs disproportionately harm protected groups without sufficient justification.
Incorrect. Disparate impact does not require intent — that is what makes it so significant for AI. An algorithm can encode historical discrimination purely by learning from past decisions. Disparate impact doctrine can reach that outcome even though no one programmed bias into the system deliberately.
3. What did the CFPB's 2023 guidance require regarding AI-generated adverse action notices?
Correct. The CFPB guidance updated existing ECOA and FCRA requirements: black-box AI models cannot generate generic adverse action reasons. Notices must explain the actual, specific reasons for the individual decision — creating pressure to use interpretable models or add explainability layers.
Not quite. The CFPB guidance says AI-generated reasons must be accurate and specific to the individual. "Too complex to explain" is not acceptable. This creates real compliance pressure for lenders using opaque models that cannot generate meaningful individual explanations.
4. The EEOC's May 2023 guidance on AI hiring tools stated that employers:
Correct. The EEOC confirmed that "we used a vendor's algorithm" is not a Title VII defence. Employers bear responsibility for the discriminatory effects of tools they deploy. This placed significant compliance obligations on organisations using third-party AI HR tools.
Incorrect. The EEOC explicitly rejected the vendor-insulation argument. An employer cannot avoid Title VII liability by outsourcing hiring decisions to an AI tool — they remain responsible for discriminatory outcomes regardless of who built the system.

Lab 3: Algorithmic Bias and Civil Rights Law

Explore how civil rights statutes apply to AI discrimination scenarios

Bias Scenario Analysis

You will describe a scenario involving an AI system that may be producing discriminatory outcomes, and the tutor will help you identify which civil rights statute applies, whether disparate impact or disparate treatment is the relevant theory, and what enforcement options might exist.

Complete at least 3 exchanges to finish this lab.

Try: "A company's AI résumé screener was trained on ten years of historically male-dominated hires and now screens out women at higher rates" — or describe a different AI discrimination scenario and we'll work through the legal analysis together.
Civil Rights and AI Tutor
AI TUTOR
0 / 3
Welcome to Lab 3. I'll help you apply civil rights law to AI discrimination scenarios. Tell me about an AI system that might be producing discriminatory outcomes — in hiring, lending, housing, or another context — and we'll work through the legal analysis: which law applies, which theory of discrimination is relevant, and what the affected person could potentially do about it.
Module 7 · Lesson 4

The EU AI Act and the Next Generation of AI-Specific Law

The world's first comprehensive AI regulation assigns legal risk levels to AI systems and outright bans several uses — and it will affect AI deployed globally, not just within Europe.
What happens when regulators stop trying to fit AI into existing laws and write entirely new rules specifically for AI systems?

On 13 March 2024, the European Parliament voted 523 to 46 to adopt the EU Artificial Intelligence Act — the first binding comprehensive AI regulation in the world. It entered into force on 1 August 2024, with most provisions phasing in over two years. The Act's passage followed three years of negotiation, significant redrafting after generative AI systems like ChatGPT appeared mid-process, and intense lobbying from both technology companies and civil society organisations.

The Act uses a risk-based architecture: it assigns AI systems to risk categories and applies different requirements based on those categories. Some systems are banned outright. Others face heavy obligations. Most face light-touch rules. And a small class of systems face requirements tied specifically to transparency — the right to know you are interacting with an AI.

The Risk Architecture: Four Levels
Level 1 · Unacceptable Risk
Prohibited Systems
Social scoring by governments; real-time biometric surveillance in public spaces (with narrow law-enforcement exceptions); emotion recognition in workplaces and schools; subliminal manipulation; exploitation of vulnerable groups. These are banned outright — no compliance path exists.
Level 2 · High Risk
Heavily Regulated
AI in biometric identification, critical infrastructure, education, employment (hiring/firing), essential services (credit, benefits), law enforcement, migration, justice administration. Must register in EU database, undergo conformity assessment, maintain logs, ensure human oversight, and meet accuracy/robustness standards.
Level 3 · Limited Risk
Transparency Obligations
Chatbots, deepfakes, AI-generated content. Providers must disclose that users are interacting with an AI. Deepfake images/videos must be labelled as artificially generated. This is the tier most consumer AI applications fall into.
Level 4 · Minimal Risk
Voluntary Codes of Conduct
AI spam filters, AI-powered video games, AI recommendations in most contexts. No mandatory requirements; voluntary codes of conduct are encouraged. The vast majority of existing AI applications fall here.
What the Ban on Social Scoring Actually Means

Article 5 of the AI Act prohibits "the placing on the market, putting into service or use" of AI systems that evaluate or classify people based on their social behaviour over a period of time, where that scoring leads to detrimental treatment in unrelated contexts. This is a direct legal response to China's social credit systems — but it also applies to private-sector scoring systems that aggregate behavioural data to assign trustworthiness ratings used in making decisions about housing, services, or employment.

The prohibition is comprehensive: it does not matter whether a European or non-European company developed the system, or whether the system operates online or offline. If it targets EU residents, the ban applies.

Real-Time Biometric Surveillance in Public Spaces

The AI Act prohibits real-time remote biometric identification (facial recognition) in publicly accessible spaces for law enforcement purposes, with three narrow exceptions: targeted search for a missing child, imminent threat of terrorist attack, or prosecution of serious crime where a court or independent authority has authorised it. Post-hoc (recorded) biometric identification for serious crime investigation is permitted with judicial authorisation.

This directly responds to the rapid expansion of facial recognition in policing. In the UK, the Metropolitan Police and South Wales Police have faced legal challenges over their use of live facial recognition. Under the AI Act, EU law enforcement agencies face strict conditions; commercial deployment for crowd surveillance is banned outright.

Fines Under the EU AI Act

Violations involving prohibited systems: up to €35 million or 7% of global annual turnover. Violations involving high-risk AI requirements: up to €15 million or 3% of turnover. Providing incorrect information to regulators: up to €7.5 million or 1.5% of turnover. These exceed even GDPR penalties for the most serious categories.

General-Purpose AI Models (GPAI) — The ChatGPT Problem

The original EU AI Act drafts focused on specific-use AI systems. The rapid rise of large language models — ChatGPT, Gemini, Claude — created a new problem: foundation models that are not built for a specific risk-level use case but can be deployed in any context, including high-risk ones. After significant debate, the Act added rules for GPAI models:

All GPAI providers must publish technical documentation, comply with copyright law, provide a detailed summary of training data. GPAI models with systemic risk (defined by training compute threshold: over 10²⁵ FLOPs, currently reaching only the largest frontier models) face additional requirements: adversarial testing ("red-teaming"), incident reporting to the European Commission, and cybersecurity protections. This is the first time legislation has placed specific compliance burdens on the largest foundation model developers.

Timeline and Global Impact

The Act's provisions phase in gradually: prohibited practices from February 2025; high-risk AI requirements for systems in Annex I from August 2026; full implementation by August 2027. As with GDPR, the extraterritorial reach is significant: any AI system used within the EU, regardless of where its developer or deployer is based, is subject to the Act. This is expected to create a "Brussels effect" similar to GDPR — where global companies build their systems to the highest (EU) standard rather than maintaining different versions for different jurisdictions.

Key Takeaway

The EU AI Act is the first law written specifically for AI rather than adapted from existing frameworks. Its risk-based architecture bans the most dangerous uses outright, imposes heavy obligations on consequential applications, and creates transparency requirements for consumer AI. Its global reach — through the Brussels effect and extraterritorial scope — means it will shape AI development far beyond Europe's borders.

Lesson 4 Quiz

EU AI Act — 4 questions
1. Under the EU AI Act, AI systems used for real-time facial recognition in public spaces by law enforcement are:
Correct. The Act bans real-time biometric identification in public spaces for law enforcement as the default, with only three narrow exceptions requiring prior judicial or independent authority approval. Commercial deployment for crowd surveillance is banned without any exceptions.
Not quite. The Act creates a near-ban with three narrow exceptions: searching for a missing child, preventing an imminent terrorist attack, or prosecuting a serious crime — each requiring prior judicial authorisation. General law enforcement surveillance is prohibited.
2. A company uses an AI system that assigns ongoing trustworthiness scores to people based on their social behaviour, used to determine eligibility for services. Under the EU AI Act, this is:
Correct. The social scoring prohibition under Article 5 applies to both public and private actors. Scoring systems that aggregate social behaviour data to make decisions about people in unrelated contexts — whether operated by a government or a private company — are banned without a compliance path.
Incorrect. Social scoring by any actor — public or private — falls under the prohibited practices in Article 5. There is no compliance tier for this: it is simply banned. The exemption for private companies does not exist in the text.
3. What additional obligations do "GPAI models with systemic risk" face under the EU AI Act?
Correct. The systemic risk tier — currently applying to the very largest frontier models above 10²⁵ FLOPs of training compute — adds red-teaming requirements, mandatory incident reporting, and cybersecurity obligations on top of the standard GPAI requirements that all foundation model providers must meet.
Not quite. GPAI models with systemic risk face three additional requirements beyond standard GPAI obligations: adversarial testing (red-teaming) before deployment, reporting serious incidents to the European Commission, and cybersecurity protections. Open-sourcing is not required.
4. What is the "Brussels effect" in the context of the EU AI Act?
Correct. The Brussels effect describes how EU regulation effectively becomes global standard: companies building AI systems to comply with the EU AI Act tend to apply those same standards everywhere, rather than engineering different products for different jurisdictions. GDPR demonstrated this pattern; the AI Act is expected to do the same.
Not quite. The Brussels effect is a well-documented economic and regulatory phenomenon: because EU standards are often the most stringent and the EU market is too large to ignore, global companies tend to build to EU requirements — effectively making EU regulation the de facto global standard without any formal international agreement.

Lab 4: EU AI Act Risk Classification

Classify AI systems under the Act's risk tiers and reason through compliance obligations

Risk Tier Classification Practice

You will describe an AI system and its use case, and the tutor will help you classify it under the EU AI Act's risk architecture: prohibited, high-risk, limited-risk, or minimal-risk. Then you'll work through what compliance looks like at that tier — or why the system cannot be deployed at all.

Complete at least 3 exchanges to finish this lab.

Start with a system like: "An AI that analyses employee productivity by monitoring keystrokes and time on task" or "A chatbot used by a government agency to screen benefit applications" — or propose any AI system you want to classify.
EU AI Act Classifier
AI TUTOR
0 / 3
Welcome to Lab 4. I'll help you classify AI systems under the EU AI Act's four-tier risk architecture. Describe an AI system and its use case — who deploys it, who it affects, what decisions it makes or assists — and we'll work through whether it is prohibited, high-risk, limited-risk, or minimal-risk, and what the compliance obligations or outright ban means in practice.

Module 7 Test

Your Rights: What the Law Actually Says — 15 questions · Pass at 80%
1. Mario Costeja González's 2014 case established the legal principle that AI-driven search indexes:
Correct. The CJEU ruling established that the right to be forgotten applies to automated retrieval systems — search indexes — not just to original publications. Delisting, not deletion of the source, was the remedy.
Incorrect. The González ruling held that search indexes as automated processing systems can be required to delist personal information even when the original article remains published. The right targeted the AI-driven index, not the source content.
2. Under GDPR, an individual's right to object to automated decision-making is found in:
Correct. Article 22 is the GDPR provision specifically addressing automated decision-making. It grants the right not to be subject to purely algorithmic decisions with significant effects, and requires human review, explanation of logic, and ability to contest outcomes.
Not correct. Article 22 is the relevant provision. It directly addresses decisions made solely by automated processing — including AI profiling — that produce legal or similarly significant effects on individuals.
3. GDPR's extraterritorial scope means:
Correct. The determining factor is whether an organisation processes EU residents' data — not where the organisation is headquartered or operates. This is what gives GDPR global reach and created compliance obligations for US, Asian, and other non-European companies.
Incorrect. GDPR applies based on the data subject's location, not the organisation's location. A company in any country that processes EU residents' data must comply — there is no opt-out mechanism.
4. The $650 million Facebook BIPA settlement covered only Illinois users because:
Correct. The same AI feature collected identical data from users nationwide, but legal protections existed only where state law created them. This geographic inequity is a defining characteristic of the US sectoral approach to privacy law.
Incorrect. The settlement's limitation to Illinois users reflects the absence of equivalent biometric privacy laws in other states. The AI system operated identically everywhere; the legal outcomes differed based entirely on geography and state legislation.
5. Illinois BIPA's private right of action means:
Correct. BIPA's private right of action has made it the most litigated biometric privacy law in the US. Individuals — and class action plaintiffs — can sue directly, creating significant financial exposure for companies using biometric AI at scale.
Not quite. BIPA creates a private right of action: individuals can sue without waiting for the Attorney General or any other regulator to act. This has driven enormous litigation, including the Facebook settlement and the Cothron v. White Castle ruling.
6. The "sectoral approach" to US federal privacy law creates a problem for AI because:
Correct. A fitness app collecting health data isn't a HIPAA "covered entity." An AI HR tool isn't subject to FCRA. A consumer profiling AI operates between multiple sectors' statutes. The patchwork leaves large gaps that general-purpose AI systems readily exploit.
Incorrect. The sectoral approach creates gaps because AI doesn't respect industry boundaries. A health-adjacent fitness app, a credit-adjacent employment tool, or a general data broker operates between the narrow statutes without clearly falling under any one of them.
7. The Illinois Supreme Court's ruling in Cothron v. White Castle (2023) held that:
Correct. This ruling dramatically expanded potential liability — an employer scanning fingerprints of 200 employees twice daily creates hundreds of violations daily. For companies using biometric AI without compliant consent procedures, the financial exposure can become enormous rapidly.
Incorrect. Cothron held that each scan is a separate violation — not that damages are capped or that there are temporal exemptions. This ruling significantly increased the financial risk for any company collecting biometric data at scale without BIPA-compliant consent.
8. HUD's 2019 complaint against Facebook alleged discrimination in which area?
Correct. HUD alleged that Facebook's ad algorithm allowed advertisers to use protected characteristics to exclude users from seeing housing ads — a Fair Housing Act violation. The discrimination was not programmed intentionally but emerged from the AI's optimisation of ad delivery.
Not quite. HUD's complaint specifically targeted housing ad targeting under the Fair Housing Act. The case demonstrated that algorithmic discrimination in housing — even when unintentional — can violate civil rights law.
9. Under the ECOA and CFPB 2023 guidance, when an AI denies credit, the applicant must receive:
Correct. The CFPB updated existing ECOA requirements for the AI era: adverse action reasons must be specific to the individual's actual circumstances, not generic. "Insufficient credit history" when the real reason was an AI-detected spending pattern is non-compliant.
Incorrect. CFPB's 2023 guidance requires that AI-generated adverse action notices contain accurate, specific, individual reasons — not generic boilerplate. This creates compliance pressure for lenders using opaque AI models that cannot generate individualized explanations.
10. The EEOC's May 2023 AI hiring guidance held that employers:
Correct. The EEOC guidance confirmed that there is no AI vendor exception to Title VII. If a tool produces discriminatory outcomes, the employer who deployed it is liable — regardless of who built the system or what the vendor claimed about its objectivity.
Incorrect. The EEOC specifically rejected the vendor insulation argument. Employers cannot outsource civil rights liability to AI vendors. They remain responsible for discriminatory outcomes from tools they deploy in their hiring processes.
11. Under the EU AI Act, which of the following AI applications is classified as a prohibited (unacceptable risk) practice?
Correct. Social scoring — assigning trustworthiness scores based on social behaviour that affect people's access to services in unrelated contexts — is explicitly prohibited under Article 5 of the EU AI Act. No compliance path exists; deployment is simply unlawful.
Not correct. The social scoring system falls into the prohibited tier. A benefits chatbot would be assessed for its specific use context; a mortgage screener is high-risk; a streaming recommendation is minimal-risk. Social scoring by any actor — public or private — is outright banned.
12. The EU AI Act classifies AI used in employment decisions (hiring and firing) as:
Correct. Employment decisions fall in the high-risk category under the AI Act. This means mandatory registration in the EU database, conformity assessment before deployment, maintaining activity logs, ensuring human oversight, and meeting accuracy and robustness standards.
Incorrect. AI in employment is classified as high-risk — not prohibited, but subject to the most demanding compliance requirements short of outright prohibition. Developers and deployers must meet conformity assessment, logging, human oversight, and accuracy standards.
13. GPAI models with "systemic risk" under the EU AI Act are currently identified by:
Correct. The systemic risk designation is tied to training compute: 10²⁵ FLOPs. Currently this reaches only the very largest frontier models. These face additional red-teaming, incident reporting, and cybersecurity requirements on top of standard GPAI obligations.
Not quite. The EU AI Act uses a compute-based threshold — 10²⁵ FLOPs of training compute — to identify systemically risky GPAI models. This ensures the heaviest requirements fall on the most capable and potentially most dangerous systems.
14. "Disparate impact" as a legal theory in US civil rights law is important for AI accountability because:
Correct. AI systems trained on biased historical data reproduce historical inequities without anyone programming in discriminatory intent. Disparate impact doctrine reaches those outcomes — the legal question is the effect, not the algorithm's "intent."
Incorrect. Disparate impact doctrine requires no proof of intent. This is its central importance for AI: systems that learn from biased data will produce biased outputs without deliberate design, and disparate impact can reach those outputs across housing, credit, employment, and other contexts.
15. The "Brussels effect" in the context of the EU AI Act refers to:
Correct. Because the EU market is too large to ignore and maintaining separate product versions is costly, global companies often build to the highest applicable standard — which tends to be the EU's. GDPR demonstrated this dynamic; the AI Act is expected to replicate it for AI governance globally.
Incorrect. The Brussels effect describes a real market dynamic — not a treaty or funding mechanism. When companies build AI systems to EU standards because of market access needs, those standards effectively become global even without formal international agreements.