L1
·
Quiz
·
Lab
L2
·
Quiz
·
Lab
L3
·
Quiz
·
Lab
L4
·
Quiz
·
Lab
Module Test
Module 3 · Lesson 1

How AI Ad Targeting Works

From behavioral signals to bid auctions — the machine learning pipeline that decides which ad you see next.
How does a platform translate your clicks, pauses, and scrolls into a profile worth billions in advertising revenue?

In 2012, a Target statistician named Andrew Pole built a pregnancy-prediction model using purchase history. The model assigned each customer a "pregnancy score" based on 25 proxy products — unscented lotion, mineral supplements, cotton balls. Target began mailing prenatal coupons to women the algorithm flagged. One Minnesota father complained to a store manager about coupons addressed to his teenage daughter — only to discover, days later, that she was in fact pregnant. The algorithm had inferred the pregnancy before the family knew. This was not social media — but it demonstrated the inference capability that platforms would soon run at billion-user scale.

The Data Collection Layer

Every major social media platform operates a behavioral data collection layer that runs continuously in the background. Facebook's pixel, deployed on millions of third-party websites, sends conversion and browsing signals back to Meta servers even when users are not on Facebook. Instagram tracks which posts a user hovers over and for how long. TikTok's recommendation system monitors watch completion rate — whether you watch a video to 25%, 50%, 75%, or the end — as a primary engagement signal.

These signals feed into user interest graphs: structured databases mapping each account to hundreds or thousands of inferred interest categories. Meta's Ad Manager historically exposed over 1,000 targeting categories to advertisers. A 2021 Markup investigation found that Meta allowed advertisers to target users based on interests including "Jew hater" and other hate-adjacent categories, which Meta removed after publication but which had existed in the system for years — generated automatically by machine learning, not human curation.

Real Case — Facebook Interest Categories (2021)

The Markup published evidence in June 2021 that Meta's ad targeting system had auto-generated interest categories including antisemitic and white-nationalist-adjacent terms. These were created by an unsupervised ML pipeline that clustered user behavior without human review of the resulting category labels. Meta removed over 200 categories after the report, but acknowledged the system could regenerate similar categories without ongoing auditing.

The Auction Mechanism

Once a user profile is built, ad placement occurs via a real-time bidding (RTB) auction that completes in under 100 milliseconds — faster than a human blink. When you load a page or open an app, the platform's ad server sends a bid request to dozens of demand-side platforms (DSPs) simultaneously. Each DSP runs its own ML model predicting your likelihood to click, convert, or engage, and submits a bid price.

Meta and Google run first-price or second-price hybrid auctions. The winning bid is not necessarily the highest dollar amount — platforms weight bids by a predicted relevance score. An advertiser with a lower bid but a highly relevant audience match can outbid a higher-spending competitor. This relevance weighting system, called Ad Rank at Google and Total Value at Meta, means ad algorithms optimize simultaneously for user engagement and advertiser spend — creating structural pressure to show engaging content, including emotionally arousing content.

Lookalike AudiencesA Meta and TikTok feature that takes a seed audience (e.g., existing customers) and uses ML to find users whose behavioral profiles most closely resemble that seed. Advertisers can target people they have never interacted with based on structural similarity to known converters.
Custom AudiencesAdvertisers upload email lists or phone numbers; platforms hash and match them against user accounts, enabling targeting of specific individuals without the advertiser ever knowing which platform account corresponds to which email.
RTB (Real-Time Bidding)The auction infrastructure that selects and prices each ad impression in milliseconds, using ML-based relevance and conversion probability models from multiple competing buyers.

Cambridge Analytica and the Limits of Consent

The 2018 Cambridge Analytica scandal remains the most documented case of social media ad-targeting data being used for political influence at scale. The firm harvested data from approximately 87 million Facebook profiles through a personality quiz app, exploiting Facebook's then-permissive API that allowed apps to collect data not just from consenting users but from all of their friends. Cambridge Analytica used this data to build psychographic profiles — classifying users along the OCEAN model (Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism) — and served micro-targeted political ads during the 2016 US presidential election and the Brexit referendum.

The key technical fact: the targeting did not require Cambridge Analytica to know users' names. The firm uploaded psychographic segments directly into Facebook's Custom Audiences system, which matched profiles to Facebook accounts server-side. Advertisers never needed personally identifiable information — the platform's matching infrastructure did the work.

Key Insight

Modern ad targeting is a two-sided privacy problem: users don't see what data is collected, and advertisers don't see who they're targeting. The platform sits in the middle, matching both sides without full disclosure to either. This architectural opacity is a design choice, not a technical necessity.

Lesson 1 Quiz

How AI Ad Targeting Works — 4 questions
1. What did Target's 2012 pregnancy-prediction model demonstrate about behavioral inference that would become central to social media advertising?
Correct. The Target case established that ML models inferring sensitive attributes from proxy behaviors — a method now core to social media profile-building — could outpace users' own awareness of their circumstances.
Not quite. The broader significance was the demonstration that behavioral signals could infer undisclosed sensitive facts — the principle that social platforms now use at massive scale.
2. In a real-time bidding auction, what determines which ad wins the placement — not just the highest dollar bid?
Correct. Platforms weight bids by predicted relevance (Ad Rank at Google, Total Value at Meta), meaning a lower bid from a highly matched advertiser can beat a higher bid from a less relevant one.
Incorrect. RTB auctions use a hybrid of bid price and an ML-generated relevance score — creating structural incentives for ads that maximize predicted engagement.
3. What was the key technical mechanism Cambridge Analytica exploited to target users without holding their personal data directly?
Correct. Cambridge Analytica used Facebook's own Custom Audiences tool — the platform performed the identity matching, meaning advertisers could micro-target without ever knowing which account corresponded to which individual.
Incorrect. The firm uploaded audience segments into Meta's Custom Audiences system, which did the matching internally. Advertisers never needed to hold user PII to reach specific individuals.
4. The 2021 Markup investigation found that Meta's ad interest categories had auto-generated offensive targeting options. What does this reveal about how those categories were created?
Correct. Meta's interest category generation used unsupervised learning that clustered behavioral data and auto-labeled clusters — no human reviewed whether those labels were appropriate, which is how harmful categories persisted in the system.
Incorrect. The categories were auto-generated by unsupervised ML that labeled behavioral clusters without human review — demonstrating a systemic gap in how ML-generated outputs are audited before deployment.

Lab 1: Reverse-Engineer a Targeting Profile

AI-assisted exploration — discuss behavioral signals and what they infer

Your Task

You're a digital marketing analyst reviewing how a platform might classify a hypothetical user based on their online behavior. Use the AI assistant to explore which interest categories behavioral signals map to, how advertisers would bid on this profile, and what data-minimization alternatives exist.

Start with: "A user watches cooking videos, browses travel blogs, and recently searched for baby furniture. What interest categories and lookalike audiences would a platform likely assign them?"
Ad Targeting Analyst
L1 Lab
Welcome to Lab 1. I'm your ad targeting analysis assistant. Let's explore how behavioral signals translate into advertising profiles. Describe a user's online behavior and we'll work through what an ad platform would likely infer about them — and what that means for their advertising experience.
Module 3 · Lesson 2

Micro-Targeting and Political Advertising

When ad-tech precision meets democratic discourse — the documented cases that changed platform policy.
What happens when the same tools that sell sneakers are applied to selling candidates and ballot measures — and why do platforms respond so differently to the problem?

During the 2016 US presidential election, the Internet Research Agency — a St. Petersburg-based organization linked to Russian intelligence — ran a coordinated influence operation on Facebook, Instagram, Twitter, and YouTube. The operation created 3,517 Facebook ads and over 80,000 organic posts, spending approximately $100,000 on paid advertising. The ads were micro-targeted by geography, ethnicity, and inferred political affiliation. One campaign targeted Black voters in key swing states with voter-suppression messaging; another targeted conservative Christians with anti-immigration content. The effectiveness was not the dollar amount — it was the precision.

How Political Micro-Targeting Differs from Commercial

Commercial ad targeting optimizes for measurable outcomes: clicks, purchases, app installs. Political targeting optimizes for behavior that is harder to measure — belief change, turnout suppression, or identity reinforcement. This difference matters for platform accountability: a retailer can A/B test conversion rates; a campaign manager A/B tests which message most effectively discourages a demographic from voting.

Facebook's internal research, leaked via whistleblower Frances Haugen in 2021, included a 2019 study on political polarization in the News Feed algorithm. The research found that content optimized for engagement tended to be more partisan and emotionally extreme than content that users said they wanted to see. The algorithm's optimization target — engagement — and the social good — informed democratic participation — were in structural tension.

Documented Case — Ireland Abortion Referendum (2018)

Ahead of Ireland's May 2018 referendum on abortion legalization, both domestic and international campaigns ran targeted Facebook ads. Google and Facebook both banned foreign political ads in Ireland during the campaign after the Irish government expressed concern — but enforcement was inconsistent. A University College Dublin audit found that ads from non-Irish organizations continued to run after the ban, and that the targeting parameters used allowed campaigns to reach users based on inferred political views rather than declared ones.

Dark Posts and Transparency Failures

Dark posts (also called unpublished page posts) were political ads that appeared in targeted users' feeds but were invisible to anyone not in the targeted audience — including journalists, regulators, and opposing campaigns. They left no trace on the advertiser's public page. During the 2016 US election, the Trump and Clinton campaigns both used dark posts extensively; the Trump campaign reportedly ran up to 175,000 different ad variations per day through A/B testing, each visible only to its specific target group.

The US Honest Ads Act, first introduced in 2017 and reintroduced in multiple sessions, would require online political ad buyers to disclose targeting parameters and funding sources — analogous to rules that have applied to broadcast political ads since 1971. As of 2024, the act has not passed. Facebook launched its own Ad Library in 2019, which provides some transparency, but it does not reveal targeting parameters — only creative content and rough spend ranges.

Dark PostsUnpublished political ads targeted at specific audience segments; invisible to anyone outside the target group and historically leaving no public record, enabling A/B testing of divisive messaging at scale.
Psychographic TargetingClassifying users by inferred personality traits (e.g., OCEAN model) and serving messages calibrated to those traits — Cambridge Analytica's stated methodology for the 2016 US election campaigns.
Special Ad CategoriesA Meta policy introduced after 2018 that restricts targeting options for housing, employment, credit, and political ads — prohibiting targeting by age, gender, and some demographic proxies in those categories in response to civil rights settlements.

Platform Responses and Their Limits

In October 2019, Twitter banned all political advertising globally. CEO Jack Dorsey framed it as a principled stance: "We believe political message reach should be earned, not bought." The ban covered candidates, political parties, and advocacy ads on contested political issues. Critics noted the ban was easier for Twitter — whose ad revenue was a fraction of Facebook's — and that organic political content on Twitter remained unregulated.

Facebook took the opposite position, explicitly refusing to fact-check political ads and allowing micro-targeting to continue, arguing it was not the platform's role to arbitrate political truth. Google adopted a middle path in 2019: banning targeting of political ads by political affiliation or voter file data, but permitting targeting by age, gender, and geography. None of these policies addressed the core issue: that engagement-optimized algorithmic feeds amplify political content differently than commercial content, regardless of whether it is paid.

Key Tension

Platform policies on political advertising reveal a fundamental conflict of interest: political ad revenue is lucrative, political ad regulation is complex, and the platforms that profit from political micro-targeting are also the ones designing their own transparency rules. Self-regulation in this space has produced inconsistent, incomplete, and easily circumvented systems.

Lesson 2 Quiz

Micro-Targeting and Political Advertising — 4 questions
1. The Internet Research Agency's 2016 Facebook operation spent approximately $100,000. What made it effective despite this relatively small budget?
Correct. The IRA operation's impact came from targeting precision — voter-suppression content for Black voters in swing states, identity-reinforcing content for specific religious demographics — not from total spend.
Incorrect. The operation's effectiveness derived from targeting precision: specific messages reaching specific demographic and geographic segments most susceptible to each message type.
2. What did Frances Haugen's leaked 2019 Facebook research reveal about the relationship between News Feed engagement optimization and political polarization?
Correct. The leaked research showed a structural tension: the engagement metric that drives ad revenue also drives amplification of content that is more extreme and polarizing than users claim to prefer.
Incorrect. The leaked research showed the opposite: engagement-optimized content tended toward more extreme, partisan material — creating a structural conflict between the algorithm's optimization target and democratic information quality.
3. Why were "dark posts" particularly problematic for democratic accountability in political advertising?
Correct. Dark posts' defining accountability problem was their invisibility: targeted content could test divisive messaging or voter-suppression themes with specific audiences while leaving no trace for public review.
Incorrect. The core problem was transparency: dark posts were visible only to their targeted audience, making scrutiny by journalists, regulators, and rival campaigns impossible — a structural accountability gap that broadcast political ad rules were designed to prevent.
4. Twitter's 2019 global political ad ban and Facebook's refusal to restrict political ads represented opposing policy choices. What limitation did both approaches share?
Correct. Ad policy debates focused on paid placements while ignoring that organic algorithmic amplification — driven by the same engagement optimization — shapes political information environments more broadly than paid ads alone.
Incorrect. The shared limitation: both policies addressed paid ads while leaving untouched the deeper issue — that engagement-optimization algorithms amplify politically extreme content in organic feeds regardless of whether it is paid.

Lab 2: Evaluate a Political Ad Transparency Policy

AI-assisted analysis — critique real platform policies on political advertising

Your Task

You're advising a digital rights organization reviewing social media platforms' political advertising policies ahead of a major election. Use the AI assistant to evaluate the strengths and gaps in current transparency approaches — Meta's Ad Library, Twitter's ban, Google's partial restrictions — and explore what a more effective regulatory framework might look like.

Start with: "What are the key weaknesses in Meta's Ad Library as a political ad transparency tool, and what information would a truly accountable system need to include?"
Political Ad Policy Analyst
L2 Lab
Welcome to Lab 2. I'm your political advertising policy analyst. Let's evaluate the transparency frameworks that platforms have built — and the gaps that leave political micro-targeting largely unaccountable. What aspect of political ad policy would you like to examine first?
Module 3 · Lesson 3

Discriminatory Targeting and Civil Rights

When algorithmic ad systems reproduce and amplify historical discrimination — documented cases and legal settlements.
If an AI system never explicitly considers race or gender but still delivers housing ads predominantly to white men and job ads predominantly to young women, has discrimination occurred?

In March 2019, the US Department of Housing and Urban Development filed a formal complaint against Facebook, alleging that its ad targeting system enabled housing advertisers to illegally exclude users from seeing housing ads based on race, national origin, sex, disability, and familial status — all protected classes under the Fair Housing Act of 1968. Facebook settled the case in 2019, agreeing to create a separate ad portal for housing, employment, and credit ads with restricted targeting options. Facebook also agreed to pay $5 million and submit to a five-year external audit.

The Proxy Discrimination Problem

The HUD case illustrated a core problem in algorithmic advertising: protected class discrimination does not require explicitly targeting protected classes. Facebook's ad system allowed advertisers to target by zip code — which correlates with race in a racially segregated housing market. It allowed targeting by "ethnic affinity" (a category Meta created), by interests that correlate with religion, and by age ranges. Each individual variable might be legal in isolation; their combination could produce discriminatory outcomes at scale.

A 2019 study by researchers at Northeastern University, published in the ACM FAT* conference, found that even when advertisers made no intentional demographic choices, Facebook's delivery optimization algorithm skewed ad distribution in ways that correlated strongly with race and gender. The algorithm's optimization for click-through rates caused housing ads to be delivered predominantly to white users and job ads in male-dominated fields to skew male — because the algorithm learned from historical engagement patterns that reflected existing discrimination.

Real Settlement — ACLU v. Facebook (2018–2019)

The ACLU, NAACP Legal Defense Fund, and Communications Workers of America jointly sued Facebook in 2018, alleging that its ad targeting tools allowed employers and housing providers to illegally exclude women, older workers, and people of color. The case was settled in 2019: Facebook agreed to end the use of age, gender, and zip code targeting for housing, employment, and credit ads, and to create algorithmic auditing mechanisms — the first time a federal civil rights framework was explicitly applied to an AI ad delivery system.

Delivery Optimization as Discrimination

The Northeastern researchers' finding was technically significant: discriminatory ad delivery occurred at the delivery optimization layer, not the targeting layer. An advertiser using no demographic parameters at all could still receive a skewed audience because the platform's ML model was optimizing for who would click — and who clicks reflects who has historically been included in those opportunities.

This distinction matters for regulation: restricting targeting parameters (what Facebook agreed to do) does not address delivery algorithm discrimination. A 2022 follow-up study by the same research group found that even after Meta's 2019 settlement-mandated restrictions on housing ad targeting, the delivery algorithm continued to show demographic skew in its audience distribution — the problem had moved from the targeting layer to the delivery layer, which remained unaddressed by the settlement.

Proxy DiscriminationUsing variables that correlate with protected characteristics (zip code → race; interests → religion; career stage → gender) to achieve discriminatory ad targeting without explicitly selecting protected class attributes.
Delivery Algorithm BiasDemographic skew introduced by the platform's ad delivery optimization — independent of advertiser targeting choices — because the ML model learns from historical engagement data that reflects existing social inequalities.
Special Ad CategoriesMeta's policy classification requiring housing, employment, and credit advertisers to use a restricted targeting tool that removes some demographic variables — introduced as part of civil rights settlements but shown to be insufficient at the delivery layer.

Facial Recognition and Demographic Inference

A separate but related issue: platforms have historically used computer vision on uploaded photos to infer demographic characteristics, which could feed back into advertising profiles. Facebook's "face recognition" feature, enabled by default until 2021, generated facial signature data for users. While Meta stated this was used only for photo tagging, civil liberties researchers noted that face recognition embeddings could in principle encode demographic proxies including perceived race, age, and gender.

Illinois's Biometric Information Privacy Act (BIPA) — the strongest US biometric data law — resulted in a $650 million settlement between Meta and Illinois users in 2022, the largest BIPA settlement to date and one of the largest privacy class action settlements in US history. The case did not directly address advertising, but it established that biometric data collection without informed consent constitutes actionable harm — a principle with direct relevance to any advertising system that infers demographics from imagery.

Key Insight

The civil rights challenge to algorithmic advertising is harder than the legal settlements suggest. Restricting which targeting parameters advertisers can select does not fix an ML delivery system trained on historically discriminatory engagement data. Meaningful algorithmic fairness in advertising requires auditing the delivery layer — not just the targeting interface — and that has not yet been systematically mandated by law.

Lesson 3 Quiz

Discriminatory Targeting and Civil Rights — 4 questions
1. What is "proxy discrimination" in the context of ad targeting, and why is it legally and technically challenging?
Correct. Proxy discrimination uses correlated variables — zip codes that map onto race, interests that map onto religion — to achieve discriminatory targeting without explicitly invoking protected classes, making it harder to detect and regulate.
Incorrect. Proxy discrimination is the use of correlated variables — zip code correlates with race, interest in certain media correlates with religion — to achieve discriminatory ad delivery without explicitly targeting protected characteristics.
2. The Northeastern University study on Facebook ad delivery found that demographic skew occurred even when advertisers selected no demographic targeting. What caused this?
Correct. The delivery layer — not the targeting layer — produced the skew. The ML model learned from historical click patterns shaped by discrimination, and reproduced that skew when optimizing for engagement, regardless of advertiser intent.
Incorrect. The cause was in the delivery optimization layer: the ML model optimized for engagement using historical data that reflected existing discrimination, reproducing demographic skew as a byproduct — even with no advertiser targeting choices at all.
3. A key finding in the 2022 follow-up study on Facebook housing ads was that Meta's 2019 settlement-mandated restrictions were insufficient. Why?
Correct. The settlement targeted the advertiser-facing interface (targeting parameters) while leaving the ML delivery engine untouched. Demographic skew simply shifted from the targeting layer — now restricted — to the delivery layer, which the settlement did not require to be audited or modified.
Incorrect. The settlement's restriction on targeting parameters did not fix discriminatory delivery algorithm behavior. Demographic skew persisted because the underlying ML delivery model continued to optimize from historically biased engagement data.
4. The $650 million Meta BIPA settlement in 2022 related to facial recognition. What principle did it establish with relevance to advertising systems?
Correct. The BIPA settlement established that biometric data collection without informed consent is an actionable harm — relevant to advertising because any system using facial analysis to infer age, gender, or other demographics for targeting would fall under the same principle.
Incorrect. The core principle: biometric data collection without informed consent constitutes actionable harm. If an advertising system infers demographic characteristics from facial imagery to improve targeting, the same legal principle would apply.

Lab 3: Audit an Ad Delivery System for Bias

AI-assisted analysis — explore how delivery algorithm bias is detected and remediated

Your Task

You're a civil rights researcher commissioned to audit a social media platform's housing ad delivery system. Use the AI assistant to design an audit methodology, explore what data you would need, and identify which metrics would reveal delivery-layer discrimination — not just targeting-layer issues.

Start with: "If I want to detect delivery algorithm bias in housing ads — separate from advertiser targeting choices — what data would I need access to, and what would a bias audit methodology look like?"
Algorithmic Fairness Auditor
L3 Lab
Welcome to Lab 3. I'm your algorithmic fairness auditor. We're investigating how to detect and measure discrimination that occurs at the delivery optimization layer of an ad system — distinct from what advertisers choose to target. What aspect of the audit design would you like to start with?
Module 3 · Lesson 4

Regulation, Privacy Laws, and the Future of Targeting

GDPR, CCPA, cookie deprecation, and the post-tracking advertising ecosystem taking shape now.
As privacy regulations tighten and third-party cookies disappear, are platforms building a more ethical advertising system — or just a more opaque one?

In January 2023, the Irish Data Protection Commission — acting as lead EU regulator for Meta under GDPR — fined Meta €390 million for its "consent or pay" model, which required Facebook and Instagram users to either consent to personalized advertising or pay a subscription fee. The European Data Protection Board ruled that Meta's approach did not constitute valid consent under GDPR because users were not given a genuine free choice. The ruling fundamentally challenged the behavioral advertising model that Meta had built its entire revenue stream upon.

GDPR and Its Enforcement Reality

The General Data Protection Regulation, which took effect in May 2018, established several principles directly relevant to ad targeting: lawful basis for processing (consent, legitimate interest, or contract), data minimization (collecting only what is necessary), purpose limitation (using data only for its stated purpose), and the right to object to profiling. Platforms responded by adding consent banners that research consistently showed were designed to make refusal difficult — a practice subsequently classified as an illegal "dark pattern" by EU regulators.

A 2022 study by researchers at MIT and the University of Michigan analyzing 1.5 million cookie consent interfaces found that only 11.8% of consent pop-ups met minimum GDPR requirements. Common violations included pre-ticked consent boxes, making the reject option harder to find than accept, and requiring multiple clicks to opt out vs. one click to accept. The French data regulator CNIL fined Google €150 million and Facebook €60 million in 2022 specifically for making cookie rejection harder than acceptance.

Real Enforcement — Irish DPC vs. Meta (2022–2023)

The Irish DPC issued a €265 million fine to Meta in November 2022 for a 2021 data scraping incident involving 533 million users' phone numbers and personal data. In January 2023, an additional €390 million fine addressed the legal basis for processing data for advertising. Combined with a separate €17 million fine, Meta faced over €1.3 billion in GDPR fines from Ireland alone by May 2023 — the largest GDPR penalty in the regulation's history to that point, covering Meta's systemic approach to consent for behavioral advertising.

Cookie Deprecation and Privacy Sandbox

Third-party cookies — the tracking mechanism that allowed advertisers to follow users across websites and build cross-site behavioral profiles — have been deprecated by major browsers. Firefox and Safari blocked third-party cookies by default years before Google. Chrome, which holds over 60% of browser market share, began phasing them out in 2024. Google's replacement, called the Privacy Sandbox, uses an on-device API called Topics that groups users into broad interest categories locally (on the user's device) rather than sharing individual identifiers with advertisers.

Privacy advocates have criticized Privacy Sandbox as insufficient: users' interest categories are still shared with advertisers and the system still enables targeting, albeit less granular. Advertisers have criticized it as too restrictive for effective campaigns. The UK's Competition and Markets Authority opened an investigation into Privacy Sandbox in 2021, concerned that Google's dominant position in both browser and ad markets meant the replacement system would favor Google's own advertising infrastructure over competitors.

GDPR Lawful BasisThe legal justification required for processing personal data under EU law — consent, legitimate interest, contract performance, legal obligation, vital interests, or public task. Behavioral advertising has struggled to establish a valid lawful basis that survives regulatory scrutiny.
Dark Patterns (Consent)Interface design that manipulates users into consenting to data collection — pre-ticked boxes, hidden reject options, misleading framing — classified as illegal under GDPR and increasingly enforced by EU data protection authorities.
Contextual AdvertisingServing ads based on the content of the page being viewed rather than the behavioral history of the user — the pre-behavioral-targeting model that is seeing renewed interest as cross-site tracking becomes harder.

The EU Digital Services Act and Platform Accountability

The EU Digital Services Act (DSA), fully effective from February 2024, introduced new obligations for very large online platforms with over 45 million EU users. Relevant to advertising: platforms must maintain a public, searchable repository of all ads shown on the platform, including who paid for them and which audiences they targeted. Targeting based on sensitive personal data — health, religion, political views, sexual orientation — is prohibited. Targeting of minors is banned entirely.

The DSA also requires large platforms to conduct annual risk assessments of their recommender systems and advertising practices, with results submitted to EU regulators. X (formerly Twitter) was the first platform to receive a DSA non-compliance finding, in 2024, related to its advertising transparency repository and content moderation practices. Fines under DSA can reach 6% of global annual revenue — for a company like Meta, potentially exceeding $7 billion.

Looking Forward

The trajectory of ad tech regulation points toward a world where behavioral surveillance-based advertising faces increasing legal and technical constraints. The likely outcome is not an end to targeting, but a shift: from individual-level behavioral profiles to contextual signals, cohort-level interest groups, and first-party data relationships between users and platforms. Whether this represents genuine privacy improvement or the same economic extraction model with better marketing is the central open question.

Lesson 4 Quiz

Regulation, Privacy Laws, and the Future of Targeting — 4 questions
1. The Irish DPC's €390 million fine against Meta in January 2023 challenged the "consent or pay" model. Why did the EDPB rule that this did not constitute valid GDPR consent?
Correct. GDPR requires consent to be freely given. Conditioning a social service on either surrendering data or paying money is coercive — the choice between "consent or pay" fails the free choice standard because financially or socially constrained users have no real alternative.
Incorrect. The ruling addressed the voluntariness of consent: when users must choose between paying or surrendering their data, they are not exercising a genuine free choice — GDPR's foundational requirement for valid consent.
2. The 2022 study on 1.5 million cookie consent interfaces found that only 11.8% met minimum GDPR requirements. What were the most common violations?
Correct. These are classic consent dark patterns: asymmetric design that makes acceptance easy and rejection burdensome, pre-selecting consent by default, and burying the refusal option — all subsequently subject to regulatory enforcement actions by CNIL and other EU authorities.
Incorrect. The primary violations were dark patterns: pre-ticked acceptance, harder-to-find rejection options, and requiring more user effort to refuse than to consent — design choices that systematically undermine the voluntariness of consent.
3. Critics of Google's Privacy Sandbox argued it was insufficient as a privacy protection. What was the UK Competition and Markets Authority's separate, competition-focused concern?
Correct. The CMA's concern was structural: Google controls Chrome (60%+ browser share) and dominates the ad exchange and DSP markets. Designing the cookie replacement gives Google the ability to create standards that work better for its own ad products than for competing ad networks.
Incorrect. The CMA's concern was about competitive advantage: Google simultaneously controls the dominant browser and the dominant advertising infrastructure, giving it unilateral power to define cookie replacement standards in ways that could disadvantage rival ad technology companies.
4. Under the EU Digital Services Act (DSA), what new advertising-specific obligations do very large platforms face?
Correct. The DSA's advertising provisions represent the most comprehensive transparency and restriction framework yet enacted: mandatory public ad libraries with targeting disclosure, bans on sensitive category and minor targeting, and required risk assessments — with fines up to 6% of global revenue.
Incorrect. The DSA requires: a public searchable repository of all ads with targeting parameters; prohibition on targeting using sensitive personal data and on targeting minors; and annual systemic risk assessments of recommender and advertising systems submitted to regulators.

Lab 4: Design a Privacy-Respecting Ad System

AI-assisted design — explore what ethical, regulation-compliant advertising infrastructure looks like

Your Task

You're advising a new social platform that wants to build an advertising system that is both commercially viable and compliant with GDPR and the DSA — without relying on behavioral surveillance. Use the AI assistant to explore contextual advertising, first-party data approaches, and what trade-offs the platform would need to accept.

Start with: "If a social platform wanted to generate advertising revenue without building behavioral profiles or cross-site tracking, what technically viable alternatives exist and what would each one sacrifice in targeting precision?"
Privacy-First Ad Architecture
L4 Lab
Welcome to Lab 4. I'm your privacy-first advertising architecture consultant. We're designing an ad system that generates real revenue while meeting GDPR and DSA requirements — without behavioral surveillance. Let's think through the technical, commercial, and ethical trade-offs. What would you like to explore first?

Module 3 Test

Targeted Advertising Systems — 15 questions · 80% to pass
1. Real-time bidding (RTB) auctions complete in under 100 milliseconds. What two factors determine the winning ad?
Correct. RTB winners are determined by bid price weighted by a predicted relevance score — Ad Rank (Google) or Total Value (Meta) — creating structural pressure for engagement-maximizing content.
Incorrect. RTB auctions weight bid price against an ML-predicted relevance/quality score, meaning a lower-bidding but highly relevant advertiser can outbid a higher spender with a poor match.
2. Facebook's pixel, deployed on third-party websites, sends behavioral signals back to Meta. When does this data collection occur?
Correct. The Facebook pixel fires on third-party websites regardless of whether the user is actively on Facebook, enabling cross-site behavioral tracking that feeds into ad profile building across the web.
Incorrect. The pixel collects and transmits data whenever a user visits a pixel-equipped site — Facebook login status is irrelevant. This is the mechanism enabling cross-site behavioral surveillance.
3. Cambridge Analytica harvested data from approximately how many Facebook profiles, and through what technical mechanism?
Correct. The app that consenting users installed could access not just their data but all of their friends' data under Facebook's then-permissive API — reaching 87 million profiles from a much smaller set of direct app users.
Incorrect. Cambridge Analytica used a quiz app that exploited Facebook's API, which allowed apps to collect data from consenting users and all their friends — yielding approximately 87 million profiles.
4. What is a "lookalike audience" in social media advertising?
Correct. Lookalike audiences use ML to find users structurally similar to a seed group — enabling advertisers to reach new users they have no prior relationship with based on behavioral profile similarity to known converters.
Incorrect. Lookalike audiences are ML-generated: the platform finds users whose behavioral signatures most closely match a seed audience, reaching previously unknown individuals based on profile similarity.
5. The Internet Research Agency's 2016 Facebook operation ran voter-suppression ads targeting Black voters. What does this illustrate about the relationship between targeting precision and budget size?
Correct. The IRA spent approximately $100,000 — a tiny political advertising budget — but achieved significant documented reach by using precision targeting to deliver specific messages to specific high-value demographic segments.
Incorrect. The IRA case demonstrates the opposite: precision targeting decouples impact from budget. A $100,000 operation with surgical demographic targeting can outperform million-dollar broad-reach campaigns for specific influence objectives.
6. Frances Haugen's leaked Facebook research showed what structural tension in the News Feed algorithm?
Correct. The leaked research established that engagement optimization and democratic information quality were in structural conflict: what drives engagement tends to be more extreme than what users report wanting, creating a feedback loop toward polarization.
Incorrect. The leaked research showed engagement optimization amplified partisan, emotionally arousing content — more extreme than what users said they preferred — creating a structural conflict between the algorithm's optimization metric and informational quality.
7. Twitter banned all political advertising in 2019. What limitation did this policy share with Facebook's opposite approach of refusing to restrict political ads?
Correct. Both policies addressed paid placements only. The deeper mechanism — algorithmic amplification of politically extreme content in organic feeds driven by engagement optimization — remained untouched by either approach.
Incorrect. Both policies' shared limitation: they regulated paid political ads while leaving untouched the organic amplification layer, where engagement-optimized algorithms distribute political content at scale with no ad policy constraints.
8. The HUD v. Facebook complaint (2019) alleged that housing advertisers could illegally exclude users from seeing ads based on protected class attributes. How did advertisers achieve this without explicitly selecting protected classes?
Correct. Proxy discrimination: zip code correlates with race in a segregated housing market; "ethnic affinity" was a Meta-created category; interest combinations correlated with religion. No explicit protected class targeting was required.
Incorrect. Proxy discrimination: variables like zip code (correlated with race in segregated markets), Meta's "ethnic affinity" categories, and interest combinations correlated with religion allowed advertisers to achieve discriminatory exclusion without explicitly selecting protected characteristics.
9. Northeastern University researchers found delivery algorithm bias in Facebook housing ads even when advertisers selected no demographic targeting. What was the mechanism?
Correct. Delivery algorithm bias: the ML model learned who was likely to click on housing ads from historical data shaped by discrimination, then optimized delivery toward those users — skewing results without any advertiser demographic input.
Incorrect. The delivery ML model learned from historically biased engagement data and reproduced that bias when optimizing who saw the ads — a discrimination mechanism entirely within the platform's infrastructure, independent of advertiser choices.
10. Meta's 2022 $650 million BIPA settlement in Illinois related to facial recognition. What principle relevant to advertising did this settlement establish?
Correct. The BIPA settlement established that unauthorized biometric collection is actionable harm. Any advertising system using facial analysis to infer age, gender, or ethnicity for targeting would face the same legal exposure.
Incorrect. The principle established: biometric data collection without informed consent constitutes actionable harm. This applies to any advertising system that uses facial imagery analysis to infer demographic attributes for targeting purposes.
11. The Irish DPC fined Meta €390 million in January 2023 for its "consent or pay" model. What specific GDPR requirement did this violate?
Correct. GDPR consent must be freely given. A binary choice between surrendering behavioral data or paying a subscription fee does not constitute free choice — it coerces consent through financial or social pressure.
Incorrect. GDPR's consent validity requirement: consent must be freely given. Requiring users to either accept data processing or pay a fee removes genuine choice, making the consent invalid regardless of how clearly it is presented.
12. The MIT/University of Michigan study of 1.5 million cookie consent interfaces found only 11.8% met GDPR minimums. Which design pattern was most commonly cited as a violation?
Correct. Consent dark patterns: pre-ticked acceptance, reject options buried behind extra clicks, and accept buttons highlighted while reject options were visually de-emphasized — all subsequently subject to enforcement actions by CNIL and other EU authorities.
Incorrect. The dominant violation was asymmetric interface design — making acceptance easier than refusal through pre-ticked boxes, hidden or multi-step rejection paths, and visual design that emphasized accept over reject.
13. Google's Privacy Sandbox Topics API replaced third-party cookies. What does "on-device" processing mean in this context, and why do privacy advocates still criticize it?
Correct. Topics classifies browsing locally (a privacy improvement over cross-site cookie tracking) but still transmits interest category labels to advertisers — meaning targeting continues, just at coarser granularity. Critics argue this is insufficient privacy protection.
Incorrect. On-device means interest categories are computed locally rather than tracked cross-site — but the resulting interest labels are still transmitted to advertisers for targeting purposes. Privacy advocates argue the improvement in privacy is insufficient since targeting persists.
14. Under the EU Digital Services Act, what is prohibited in social media advertising that was previously permitted?
Correct. The DSA's advertising prohibitions are categorical: sensitive personal data targeting (health, religion, political affiliation, sexual orientation) and minor targeting are banned outright — not merely restricted — for very large platforms.
Incorrect. The DSA specifically prohibits targeting using sensitive personal data categories and any targeting of minors — representing categorical bans on practices that had previously been commercially common on social platforms.
15. The 2022 follow-up study on Meta's housing ad delivery found that the 2019 civil rights settlement was insufficient. What does this reveal about the relationship between targeting restrictions and delivery algorithm fairness?
Correct. This is the fundamental insight of the discrimination-in-delivery research: discriminatory outcomes can persist even with restricted targeting because the ML delivery model — trained on historically unequal engagement data — continues to reproduce demographic skew at the distribution layer.
Incorrect. The research reveals a fundamental gap: civil rights enforcement focused on the targeting interface (what advertisers select) while leaving the delivery algorithm (how the ML model distributes ads) unaudited and unreformed. The discrimination moved layers, not disappeared.