In 2013, Aleksandr Kogan, a Cambridge University psychology researcher, built a personality-quiz app called thisisyourdigitallife. About 270,000 Facebook users installed it and consented — in vague terms buried deep in a user agreement — to share their data for "academic research." What they did not read, and Facebook's API freely permitted at the time, was that Kogan's app also harvested the profile data of every one of those users' Facebook friends. The final haul: 87 million people's personal information, none of whom had clicked "agree" to anything.
Kogan sold that dataset to a political consultancy called Cambridge Analytica. By 2016 the firm was modeling voter psychology in the United States and United Kingdom, micro-targeting political ads at people who had simply logged in to take a quiz about their personality. The data never traveled with a warning label. The 87 million people had no idea.
For most of human history, privacy was protected by friction. Collecting information about someone required physical effort: knocking on doors, visiting courthouses, filing through paper records. This friction was not a designed feature — it was an accidental byproduct of analog media. It meant that surveillance was expensive, information was localized, and memory was imperfect. Most embarrassing or sensitive facts simply faded with time.
Legal scholars Samuel Warren and Louis Brandeis articulated "the right to be let alone" in an 1890 Harvard Law Review article — a response to the then-new technology of the camera and the gossipy penny press. Their argument: individuals had a natural interest in controlling what others knew about them. That idea underpinned privacy law for a century.
The digital era erased the friction. When Facebook launched in 2004, storing one user's profile cost fractions of a cent. Copying it cost nothing. Analyzing it across millions of people simultaneously became routine by 2010. The old bargain — you share some things publicly, the rest quietly disappears — collapsed entirely. Everything is now recorded, cross-referenced, and retained indefinitely.
Privacy scholars distinguish between different kinds of harm. Informational harm occurs when data about you reaches someone who was not supposed to have it. Decisional harm occurs when that information is used to make a consequential decision about you — a loan denial, a targeted ad designed to manipulate a vote, a medical insurance premium hike. Relational harm occurs when your relationships are damaged because private communications are exposed.
In the Cambridge Analytica case, all three operated simultaneously. Voter profiles were built from friend data (informational harm), used to craft psychologically targeted political messaging (decisional harm), and the revelations about the data collection eroded public trust in Facebook as a platform for genuine social connection (relational harm).
AI systems can now infer information you never shared. In 2013, researchers at MIT demonstrated that 87% of Americans could be uniquely identified using only their ZIP code, birthdate, and sex — three data points available on most voter rolls. Modern AI models can infer sexual orientation from facial images, mental health status from social-media posting patterns, and political affiliation from purchase histories. You do not have to volunteer sensitive facts for them to be known.
AI systems are more useful when they have more data. A medical AI trained on millions of patient records catches diagnoses a solo physician might miss. A navigation AI that knows where millions of drivers are right now routes you around accidents in real time. The ethical challenge of this module is not to reject data collection entirely — it is to ask: who benefits, who bears the risk, and who decided?
You will play the role of a privacy investigator. The AI will give you a realistic scenario involving data collection. Your job is to identify which individually harmless facts, when combined, create a privacy risk — and explain why contextual integrity is violated.
Have at least three exchanges. Push the AI to give you harder cases.
In January 2012, Facebook conducted a now-infamous experiment. Without notifying users, researchers manipulated the emotional content of 689,003 people's News Feeds — some saw more positive posts, some saw more negative ones — to test whether emotions spread through social networks. The results were published in the Proceedings of the National Academy of Sciences in 2014 under the title "Experimental evidence of massive-scale emotional contagion through social networks."
When the experiment became public, the backlash was fierce. Had users consented? Facebook pointed to a line in its 2012 Data Use Policy permitting data use for "internal operations, including troubleshooting, data analysis, testing, research and service improvement." Buried there, the company argued, was consent. Adam Kramer, the lead researcher, later wrote that he was sorry for the distress, acknowledging the experiment "was poorly communicated." No user had been asked. No user had the practical ability to say no.
In 2008, Carnegie Mellon researcher Lorrie Faith Cranor calculated that reading every privacy policy a typical American encounters in a year would take 76 work days — roughly 250 hours. That figure has only grown. The policies are written by lawyers, for lawyers, in language deliberately designed to maximize corporate latitude while minimizing user comprehension.
Informed consent — the ethical gold standard borrowed from medical research — requires three elements: the person must understand what they are agreeing to, the agreement must be voluntary, and they must have the capacity to consent. Platform consent violates all three. Users understand almost nothing about algorithmic profiling. Consent is not voluntary when the alternative is social and professional exclusion. And the sheer cognitive overload of modern data agreements tests the limits of practical capacity.
In May 2018, the European Union's General Data Protection Regulation took effect, requiring "freely given, specific, informed and unambiguous" consent for data processing. Companies responded with elaborate cookie-consent banners. Research published in 2019 by Midas Nouwens and colleagues at University College London found that only 11.8% of major UK websites presented consent options in a way that met GDPR's own standards — and dark patterns (interfaces designed to steer users toward "accept all") were ubiquitous. Regulation existed; meaningful consent still did not.
UX researcher Harry Brignull coined the term "dark patterns" in 2010 to describe interface designs that trick users into actions they didn't intend. In the data-consent context, dark patterns are endemic:
Confirmshaming: "No thanks, I don't want better recommendations" forces users to feel foolish for declining. Hidden opt-outs: The "Accept All" button is large and brightly colored; "Manage Preferences" requires navigating three sub-menus. Moving goalposts: Facebook changed its privacy settings interface 12 times between 2005 and 2015, each redesign making it harder to restrict data sharing. Forced bundling: You cannot use Google Maps without location tracking; you cannot use WhatsApp without agreeing to share your phone contacts with Facebook's systems.
These are not accidents. A/B testing has proven, repeatedly, that friction reduces opt-outs. The interfaces are engineered to maximize data collection by minimizing meaningful choice.
Researchers and regulators have proposed several alternatives. Dynamic consent — developed in the biomedical research context — allows participants to update their consent preferences at any time via a persistent online interface, receiving real-time notifications when their data is accessed. Layered consent presents a simple one-paragraph summary up front, with detailed options available for users who want them. Opt-in by default — the GDPR's preferred approach — requires explicit affirmative action before data is collected, rather than requiring users to hunt for opt-out controls.
The challenge is that each of these models, if implemented honestly, would reduce data collection — and therefore revenue. The economic incentive structure consistently works against genuine consent.
There is a difference between notice and consent. A privacy policy is notice — it tells you something is happening. Consent requires that you understand it, that you have a real choice, and that saying no is a viable option. Most platform "consent" is really just notice with a button to acknowledge it.
The AI will present you with excerpts from real-style platform terms and consent interfaces. Your job is to identify the dark patterns present, explain which of the three informed-consent requirements they violate, and then redesign the consent moment to be genuinely ethical.
Have at least three exchanges. Ask for progressively trickier examples.
In late 2019, Hoan Ton-That, a 31-year-old Australian entrepreneur living in New York, quietly approached police departments with an offer: a facial recognition app that could search a database of three billion photographs scraped from Facebook, Instagram, Twitter, LinkedIn, Venmo, and millions of other websites — all without the permission of any of those platforms or the people pictured. Officers would take a photo of a suspect, upload it, and within seconds receive potential matches with links to the originating web pages.
The company was called Clearview AI. By February 2020, when a New York Times investigation by Kashmir Hill exposed the operation, over 600 law enforcement agencies — including the FBI and Departments of Homeland Security — had used the tool. Clearview had never asked for a single person's consent. It had never disclosed its database to those pictured. It had built the world's largest facial recognition system out of photographs people had shared in contexts they believed were social — tagged at a graduation, photographed at a rally, captured at a restaurant — not in a government biometric registry.
Facial recognition operates on biometric data — physical characteristics that cannot be changed. You can change your password. You can get a new credit card number. You cannot replace your face. This makes biometric privacy violations categorically more serious than most data breaches: the damage is permanent, the subject cannot mitigate it, and every future encounter with any face-recognition system using the same database reproduces the original violation.
Illinois recognized this in 2008 with the Biometric Information Privacy Act (BIPA), requiring companies to obtain written consent before collecting fingerprints, retina scans, or facial geometry. In 2021, Clearview settled a BIPA lawsuit, agreeing to limit sales of its technology to private companies in Illinois — though law enforcement use continued elsewhere. In 2022, Facebook's parent company Meta paid $650 million to settle a BIPA class action over its Tag Suggestions feature, which had used facial recognition to identify users in uploaded photos without their consent.
In 2019, the NIST (National Institute of Standards and Technology) tested 189 facial recognition algorithms and found that the best systems were 99.5% accurate on high-quality images of cooperative subjects. But in real-world conditions — low-resolution surveillance footage, varied lighting, partial occlusion — error rates climbed dramatically. Critically, many systems showed significantly higher false-positive rates for darker-skinned faces, women, and elderly individuals. Robert Williams, a Black man in Detroit, was wrongly arrested in 2020 based on a faulty facial recognition match. He was detained for 30 hours before the case collapsed. He was the first documented US case. He was not the last.
The Clearview model illustrates a structural shift: AI enables private companies to build surveillance infrastructure that previously only nation-states could maintain. China's system — combining mandatory national ID, ubiquitous cameras, and state-controlled AI — is often cited as the extreme case. But the components of comparable capacity now exist in the private sector in liberal democracies: Amazon Ring cameras create a privately-owned street-level surveillance network; Amazon Rekognition offers facial recognition APIs to any paying customer; Palantir aggregates law enforcement databases across jurisdictions.
The legal frameworks governing these systems lag dramatically. The US has no federal facial recognition law. The EU's AI Act (formally adopted in 2024) bans real-time remote biometric identification in public spaces by law enforcement — with exceptions for terrorism, missing children, and serious criminal investigation — but enforcement architecture is still being built.
Privacy is not only about preventing individual harm. It is about preserving the social conditions for free thought, dissent, and democratic participation. When people know — or suspect — that their face is being scanned at a political rally, a union meeting, or a mosque, many choose not to attend. This chilling effect on lawful behavior is a documented consequence of mass surveillance, studied extensively in the context of NSA bulk data collection after Edward Snowden's 2013 disclosures. The facial recognition layer makes the chilling effect more immediate: unlike phone metadata, it applies to the physical world without any device interaction at all.
San Francisco banned government use of facial recognition technology in 2019, followed by Boston, Portland, and other cities. The bans are not permanent scientific judgments — they are temporary governance tools, creating space for democratic deliberation about whether and how such systems should be used. Is a technology moratorium the right response when accuracy is inadequate and governance is absent? Or does it delay benefits — like finding missing persons — that could save lives?
The AI will describe a realistic facial recognition deployment scenario — a city transit system, a stadium security operation, a retail loss-prevention program. Your job is to evaluate it: What are the accuracy risks? Who bears the harm if the system errs? Is consent possible? Does it produce a chilling effect? What governance rules should apply?
Have at least three exchanges. Request a new scenario once you've thoroughly analyzed the first.
On May 25, 2018, the EU's General Data Protection Regulation took effect. Within hours, Max Schrems — an Austrian lawyer who had spent years fighting Facebook's data practices through European courts — filed four complaints against Google, Facebook, Instagram, and WhatsApp. The total claimed damages: 3.9 billion euros. His argument: that the platforms' "forced consent" — accept our terms or don't use the service — was not legally valid consent under GDPR's requirements. The cases would drag through regulatory bodies for years, but they established a new reality: data protection law had teeth, and someone was willing to use them.
By 2022, the Irish Data Protection Commission — the lead regulator for most major US tech companies whose European headquarters sit in Dublin — had issued fines totaling over €900 million against Meta alone. In 2023, the DPC fined Meta a record €1.2 billion for transferring European users' data to US servers in violation of GDPR rules on international data transfers. The era of consequence-free data exploitation was, at least in Europe, ending.
Privacy law today operates at three levels. At the international level, the EU's GDPR is the most comprehensive, establishing data subjects' rights — access, erasure ("right to be forgotten"), portability, and objection to automated decision-making. The Council of Europe's Convention 108+ extends similar principles to non-EU members. The OECD Privacy Guidelines, while non-binding, influence national frameworks globally.
At the national level, the US remains fragmented: sector-specific laws (HIPAA for health data, COPPA for children's data, FERPA for education records) coexist with the California Consumer Privacy Act (CCPA, 2020) and its successor the California Privacy Rights Act (CPRA, 2023), which extended GDPR-style rights to California residents. As of 2024, 19 US states have passed comprehensive consumer privacy laws, though no federal framework exists.
At the organizational level, corporate privacy programs vary enormously — from genuinely privacy-conscious firms that build data minimization into their architectures, to those that implement the minimum required to avoid liability while maximizing data extraction.
In 2014, the Court of Justice of the European Union ruled in Google Spain v. AEPD that individuals have the right to request removal of search results linking to personal information that is "inadequate, irrelevant, or no longer relevant." This "right to erasure" is now codified in GDPR Article 17. By 2023, Google had received over 5 million erasure requests and complied with approximately 46% of them. Critics argue the right is under-enforced; others argue it enables suppression of legitimate journalism. The tension between privacy and public interest in accurate information is unresolved.
Canadian privacy regulator Ann Cavoukian developed the Privacy by Design (PbD) framework in the 1990s, and it is now codified in GDPR Article 25. PbD's seven principles require that privacy protection be built into systems from the start — not bolted on after launch, not traded against other design goals, but embedded as the default condition. The principles include: data minimization (collect only what is necessary), purpose limitation (use data only for its stated purpose), default privacy (the most private setting is the automatic one), and full functionality (privacy should not degrade service).
DuckDuckGo's search engine is a functional example: it provides comparable search results to Google without storing user IP addresses, search histories, or building behavioral profiles. Apple's App Tracking Transparency framework, introduced in iOS 14.5 in 2021, requires apps to obtain explicit permission before cross-app tracking — a change that cost Meta an estimated $10 billion in 2022 revenue by eliminating much of its behavioral targeting capability. Privacy by design, implemented at infrastructure scale, produces real economic consequences for surveillance capitalism.
Individual privacy protection operates at the technical, legal, and social levels. Technically: privacy-respecting browsers (Firefox, Brave), tracker-blocking extensions (uBlock Origin), end-to-end encrypted messaging (Signal), and VPNs reduce — though do not eliminate — passive data collection. Legally: GDPR and CCPA give residents of covered jurisdictions rights to access, delete, and opt out of sale of their data — rights exercisable through companies' formal request mechanisms. Socially: supporting organizations like the Electronic Frontier Foundation, the Privacy Rights Clearinghouse, or the ACLU's Privacy Project contributes to advocacy for systemic change.
The honest caveat: individual protective action is necessary but insufficient. A single user blocking trackers while everyone around them remains unprotected does not fix the system. Privacy is a collective infrastructure problem, not a personal hygiene problem. The companies that profit from data collection have teams of engineers, lawyers, and lobbyists working to maintain the status quo. Individual choices matter, but structural reform matters more.
The core unsolved problem: the entities that collect and profit from personal data are not the entities that bear the consequences of privacy violations. A data broker sells your location history to an abusive ex-partner's private investigator and faces no direct liability. A facial recognition company provides a false match that leads to wrongful arrest and may face only civil BIPA claims. Until the economic and legal incentive structure changes — until the costs of privacy violations land on those who caused them — technological capability will continue to outrun ethical constraint.
You are a privacy architect advising a product team. The AI will describe a product or service that currently has significant privacy problems — drawn from real documented cases. You will apply Privacy by Design principles to redesign it: proposing data minimization measures, default settings, consent mechanisms, and accountability structures.
Have at least three exchanges. Push for concrete, implementable design decisions, not abstract principles.