Understand how existing data privacy laws apply to AI.
- ✓Every AI phase that touches personal data needs a lawful basis; training and deployment are assessed separately.
- ✓Legitimate interests is the usual basis for training, but only after a purpose, necessity and balancing test, and it brings a right to object.
- ✓Reusing data for a new AI purpose requires a compatibility assessment first.
- ✓Transparency includes telling people about AI use and giving meaningful information about the logic of significant automated decisions.
- ✓US law leans on notice and opt-outs, with the FTC policing misleading or retroactive changes to terms.
- Lawful basis:
- One of the six GDPR grounds (consent, contract, legal obligation, vital interests, public task, legitimate interests) that must apply before personal data is processed.
- Legitimate interests assessment (LIA):
- A documented three-step test of purpose, necessity and balancing that a controller runs before relying on legitimate interests.
- Purpose limitation:
- The principle that personal data collected for specified purposes must not be reused in ways incompatible with them.
- Compatibility assessment:
- The Article 6(4) analysis of link, context, nature of data, consequences and safeguards used to decide whether a new purpose fits the original one.
- Right to object:
- The GDPR right (Article 21) to object to processing based on legitimate interests or public task, which the controller must honor unless it has compelling overriding grounds.
Exam tips: 'The data was already public' or 'we already hold it' never ends the analysis. Pick the option that checks lawful basis and purpose compatibility first. · Employee consent is a red flag because of the power imbalance; look for another basis plus safeguards. · Training and deployment are separate processing activities. An option that lets one lawful basis cover both is usually wrong. · Match the jurisdiction: EU means a lawful basis before processing; US means notice plus opt-out rights, with the FTC policing misleading practices.
- ✓Minimization means adequate, relevant and necessary for the purpose: a balance, not a blind cut.
- ✓Storage limitation: identifiable data goes when the purpose is done.
- ✓GDPR Article 25 codifies privacy by design and by default, building on Cavoukian's seven principles.
- ✓Privacy-enhancing techniques reduce risk, but only real anonymization takes data out of scope.
- ✓Models can memorize, so minimization continues after training.
- Data minimization:
- Processing only personal data that is adequate, relevant and limited to what is necessary for the stated purpose.
- Storage limitation:
- Keeping personal data in identifiable form no longer than the purpose requires.
- Privacy by Design (PbD):
- Ann Cavoukian's framework of seven principles for building privacy into systems proactively and by default.
- Data protection by design and by default:
- GDPR Article 25's duty to build in measures that make the principles effective, and to process only what each purpose needs by default.
- Pseudonymization:
- Replacing identifiers with codes kept separately; the data stays personal data for anyone who can re-link it.
- Memorization:
- A model's tendency to retain and reproduce parts of its training data, which can expose personal data.
Exam tips: Any option that minimizes data 'regardless of the model's effectiveness' is a trap. The IAPP answer balances privacy and utility. · Encryption is security, not minimization. Don't let a security control answer a minimization question. · Privacy by design is proactive: if an option fixes privacy after launch or after a complaint, it's wrong. · Pseudonymized data is still personal data; only true anonymization takes data outside the GDPR.
- ✓The controller decides why and how, and stays accountable when vendors do the work.
- ✓DPIAs are required for likely high-risk processing, including profiling with significant effects; many AI projects qualify.
- ✓Article 28 contracts bind processors to instructions, security, sub-processor approval, assistance and deletion.
- ✓EU transfers need adequacy (such as the DPF), SCCs or binding corporate rules with a transfer impact assessment, or a narrow derogation.
- ✓Accountability means records: an Article 30 ROPA in the EU, documented assessments under US state laws.
- Controller:
- The party that decides the purposes and essential means of processing personal data, and carries primary accountability.
- Processor:
- A party that processes personal data on a controller's behalf and documented instructions.
- Data protection impact assessment (DPIA):
- A GDPR-mandated assessment, done before processing, of high-risk processing's necessity, risks and mitigations.
- Standard contractual clauses (SCCs):
- European Commission-approved contract terms used as a safeguard for transferring personal data out of the EU.
- Transfer impact assessment:
- An analysis of whether the destination country's laws undermine the protection a transfer mechanism promises.
- Record of processing activities (ROPA):
- The Article 30 register of an organization's processing purposes, data, recipients, transfers, retention and security.
Exam tips: Roles follow conduct, not contract labels: a 'processor' that trains its own model on your data is a controller for that processing. · A FIRST question involving a border usually wants the transfer mechanism settled before any data moves. · DPIA means before processing. An option that assesses risk after deployment is a trap. · The IAPP favors extending an existing PIA or DPIA into an AI impact assessment over building one from scratch.
- ✓Access, rectification, erasure, restriction, objection and portability all reach AI data, including inferences.
- ✓Article 22 restricts solely automated, significant decisions: allowed only for contract, by law or with explicit consent, plus safeguards.
- ✓Explanations must cover the procedure and principles actually applied, not a raw formula.
- ✓US states offer profiling opt-outs; California's ADMT rules add notice, opt-out and access rights.
- ✓Breaches: tell the regulator within 72 hours if there's risk, tell individuals if the risk is high, and document everything.
- Right of access:
- The right to obtain one's personal data and information about its processing, including meaningful information about the logic of automated decisions.
- Article 22:
- The GDPR right not to be subject to solely automated decisions with legal or similarly significant effects, subject to three exceptions and safeguards.
- Meaningful human involvement:
- Review by a person with the authority, competence and information to change an automated outcome, as opposed to rubber-stamping.
- Personal data breach:
- A security breach leading to accidental or unlawful destruction, loss, alteration, unauthorized disclosure of, or access to, personal data.
- Automated decisionmaking technology (ADMT):
- California's term for technology that replaces or substantially replaces human decision-making, regulated under its 2026 privacy rules.
Exam tips: Article 22 needs BOTH 'solely automated' AND a legal or similarly significant effect. A rubber-stamp human doesn't break the 'solely' chain. · The three Article 22 exceptions are contract, law and explicit consent. Legitimate interests and ordinary consent are distractors. · Article 22 safeguards are human intervention, expressing a view and contesting the decision: not access to source code, and not automatic reversal. · The 72 hours is for notifying the supervisory authority; individuals are told 'without undue delay' and only when the risk is high.
- ✓Article 9 prohibits processing special categories unless an exception, such as explicit consent, applies, on top of an Article 6 basis.
- ✓Biometrics become special category when processed to uniquely identify someone; photos alone don't.
- ✓US protection comes from state biometric laws (BIPA, CUBI), sensitive-data opt-ins and FTC enforcement.
- ✓Inferences can create sensitive data, and bias testing needs a lawful, safeguarded route.
- Special categories of personal data:
- GDPR Article 9 data (ethnic origin, political opinions, beliefs, union membership, genetic, biometric for identification, health, sex life or orientation) whose processing is prohibited unless an exception applies.
- Biometric data:
- Personal data from technical processing of physical, physiological or behavioral traits that allows or confirms a person's unique identification.
- Explicit consent:
- A clear, express statement of consent (for example, a signed or separately confirmed statement), required for some Article 9 and Article 22 exceptions.
- BIPA:
- Illinois' Biometric Information Privacy Act, requiring written notice and release before collecting biometrics and allowing private lawsuits.
- Sensitive data (US state laws):
- Categories such as biometrics, health, precise location and children's data that state privacy laws usually allow only with opt-in consent.
Exam tips: 'It was already public' is never a defense for biometrics or other sensitive data. · Biometrics are GDPR special category when used to uniquely identify someone; templates and hashes still count. · BIPA means a written release plus a private right of action. A US question about class actions over faceprints points to Illinois. · Inferred sensitive traits are still sensitive. An AI that guesses health or orientation is processing special category data.
Understand how other types of existing laws apply to AI.
- ✓Training copies works, so it needs a license or an exception.
- ✓US: fair use, decided case by case on four factors; early rulings turned on how data was acquired and on market harm.
- ✓EU: TDM exceptions allow commercial mining unless the rightsholder opted out in machine-readable form.
- ✓Purely AI-generated works lack US copyright and AI can't be a patent inventor, so use contracts.
- ✓Contracts, licenses, trade secrets and personality rights also limit training data.
- Fair use:
- A US defense that permits some unlicensed use of copyrighted works, decided by weighing purpose, nature, amount and market effect.
- Text and data mining (TDM):
- Automated analysis of text and data to find patterns; the EU allows it under two copyright exceptions.
- Reservation of rights (TDM opt-out):
- A rightsholder's machine-readable statement that its online works may not be mined under the EU's general TDM exception.
- Human authorship:
- The US requirement that a work be created by a human to qualify for copyright protection.
- Data provenance:
- A documented record of where training data came from and the rights attached to it.
Exam tips: Publicly available is not public domain. Any option built on 'it's online, so it's free' is wrong. · Copyright fixes live at the input stage (license, fair use analysis, honor opt-outs) and the output stage (similarity filters). Disclaimers and after-the-fact deletion don't cure copying. · Keep the lanes separate: Thaler v. Vidal is about patent inventorship; Thaler v. Perlmutter is about copyright authorship. · When default ownership of AI output is uncertain, the IAPP answer is to allocate rights in the contract.
- ✓Disparate treatment is intentional or explicit differential treatment; disparate impact is a neutral practice with unequal effects.
- ✓AI discriminates mostly through proxies, biased history and unrepresentative data.
- ✓Title VII, the ADEA and the ADA govern AI hiring, and employers stay liable for vendor tools.
- ✓Federal agencies retreated from disparate impact in 2025–26, but the statutes and private lawsuits remain.
- Disparate treatment:
- Intentionally treating people differently because of a protected characteristic, including deliberately using a proxy for it.
- Disparate impact:
- A facially neutral practice that disproportionately disadvantages a protected group, regardless of intent.
- Proxy variable:
- A seemingly neutral input, such as ZIP code, that correlates closely with a protected characteristic.
- Four-fifths rule:
- A rule of thumb from the 1978 Uniform Guidelines: a selection rate below 80% of the highest group's rate signals possible adverse impact.
- Business necessity:
- The US employer defense that a practice causing disparate impact is job-related and necessary, defeated if a less discriminatory alternative was refused.
- Reasonable accommodation:
- An ADA-required adjustment, such as an alternative assessment, that lets a qualified person with a disability compete fairly.
Exam tips: 'We don't collect race, sex or age' is never a defense to disparate impact. Look for the option that deals with proxies and outcomes. · The employer is liable for a vendor's tool. Contract clauses shift costs, not legal duties. · The four-fifths rule is a red flag, not a verdict: the best answer investigates and validates. · Withdrawn agency guidance isn't repealed law. Title VII, the ADEA and the ADA still apply to AI hiring.
- ✓ECOA and Regulation B require specific principal reasons for credit denials; complex models are no excuse.
- ✓FCRA adds notice and dispute rights whenever a consumer report drives an adverse decision.
- ✓The Fair Housing Act reaches tenant screening, lending and ad delivery, including disparate impact.
- ✓Insurers face state unfair discrimination laws, Colorado's algorithm law and the NAIC AI bulletin.
- ECOA and Regulation B:
- The US credit anti-discrimination law and its implementing rule, which also require specific reasons when credit is denied.
- Adverse action notice:
- A required notice telling a consumer about a negative decision and, depending on the law, the reasons or the consumer report behind it.
- Fair Housing Act:
- The 1968 US law banning discrimination in housing sales, rentals, lending and advertising, including through disparate impact.
- Fair Credit Reporting Act (FCRA):
- The US law governing consumer reports, with notice and dispute rights when a report drives an adverse decision.
- Unfair discrimination (insurance):
- The state-law standard barring insurers from distinctions not justified by risk, or based on protected traits.
Exam tips: 'Internal standards' or 'insufficient score' fails Regulation B. The best answer names the specific principal reasons. · FCRA vs ECOA: FCRA follows the consumer report; ECOA follows the credit decision. One scenario can trigger both. · Housing questions point to the Fair Housing Act, which covers screening, lending, advertising and disparate impact. · Don't over-read the 2025–26 rollbacks: agency rules changed, but statutes, courts and state laws still police AI bias.
- ✓Section 5 bans deceptive (misleading and material) and unfair (substantial, unavoidable, unjustified harm) practices.
- ✓AI capability claims need evidence before you make them; AI washing is deception.
- ✓The FTC can order deletion of models trained on ill-gotten data (Everalbum, Kurbo, Rite Aid).
- ✓Federal posture softened in 2025, but deception enforcement continues, alongside state attorneys general and the EU's Unfair Commercial Practices Directive.
- FTC Act Section 5:
- The US federal prohibition on unfair or deceptive acts or practices in commerce, applied to AI without any AI-specific statute.
- Deceptive practice:
- A material representation or omission likely to mislead a consumer acting reasonably, judged by its net impression.
- Unfair practice:
- A practice causing substantial injury that consumers can't reasonably avoid and that isn't outweighed by benefits.
- Algorithmic disgorgement:
- An FTC remedy requiring deletion of models and algorithms built from improperly obtained data.
- AI washing:
- Overstating or misrepresenting the use or capabilities of AI in a product.
- Substantiation:
- The reasonable basis, such as test evidence, a company must have before making objective product claims.
Exam tips: No AI-specific statute is needed: Section 5 reaches AI claims and practices directly. · 'We never sell the data' answers the wrong question. Deception is about misleading statements or omissions; unfairness is about unavoidable harm. · Disclaimers don't cure a misleading net impression. · Algorithmic disgorgement means deleting the model built on improperly obtained data, which is why data provenance matters.
- ✓Product liability can apply without fault: defect plus damage plus causation.
- ✓Know the three defects: manufacturing (one unit), design (every unit) and warning (an undisclosed risk).
- ✓US courts are still deciding whether AI software is a product; some now say yes for design claims.
- ✓The EU's new Directive (transposition by 9 December 2026) covers software and AI, reaches post-sale updates and eases the burden of proof.
- ✓Product safety regulators (CPSC, FDA, NHTSA; the EU's General Product Safety Regulation) work alongside liability, and federal approval can sometimes preempt state claims.
- Strict liability:
- Liability for harm caused by a defective product without proof that the maker was negligent.
- Manufacturing defect:
- A flaw where a particular unit departs from its intended design.
- Design defect:
- A flaw in the design itself that makes every unit unreasonably dangerous, typically where a reasonable safer alternative existed.
- Failure to warn:
- A defect arising from inadequate instructions or warnings about foreseeable risks.
- Product Liability Directive (EU) 2024/2853:
- The EU's revised no-fault product liability law, which treats software and AI as products and eases the burden of proof.
- Liability reform:
- Changing the legal rules that decide who is accountable for harm, as opposed to cost-shifting tools like insurance or recalls.
Exam tips: One faulty unit is a manufacturing defect; a flaw in every unit is a design defect; an undisclosed risk is a failure to warn. · Strict liability doesn't need negligence. An option saying the victim must prove carelessness under strict liability or the EU Directive is wrong. · Under the new EU Directive, software and AI are products, and updates within the maker's control keep it on the hook. · Liability reform changes who is accountable. Recalls, insurance and waivers are operational mechanisms, not reform.
Understand the main elements of AI-specific laws.
- ✓The EU AI Act protects health, safety and fundamental rights, using a product-safety model.
- ✓Four tiers: prohibited, high-risk, transparency (limited), minimal. GPAI models have their own track.
- ✓Classification follows intended purpose and context, not technique.
- ✓Scope is extraterritorial: non-EU providers are covered when they target the EU market or their output is used there.
- Risk-based approach:
- Regulating AI in proportion to the harm a use could cause, so obligations rise with the risk to health, safety and fundamental rights.
- Intended purpose:
- The use a provider intends for an AI system, including its context and conditions of use; it drives risk classification.
- Fundamental rights:
- Rights protected by the EU Charter, such as dignity, privacy, nondiscrimination and a fair trial, which the AI Act aims to safeguard alongside health and safety.
- Extraterritorial scope:
- The AI Act's reach to non-EU providers and deployers when their AI is placed on the EU market or its output is used in the EU.
Exam tips: If a stem describes the technology but not the use, pick the answer that asks about purpose and context. Risk attaches to the use. · Social scoring is banned for companies too, not just governments. · 'No obligations for minimal-risk AI' is almost right: AI literacy (Art. 4) still applies to every provider and deployer. · Military, defense and national-security AI is out of scope entirely. Don't classify it as high-risk.
- ✓Art. 5 bans eight practices: manipulation, exploiting vulnerabilities, social scoring, profiling-only crime prediction, facial scraping, workplace/education emotion recognition, sensitive biometric categorization, police real-time remote biometric ID.
- ✓The exceptions are narrow: medical or safety for emotion recognition; authorized, targeted cases for police biometric ID.
- ✓The Omnibus added bans on non-consensual intimate deepfakes and AI-generated child sexual abuse material from 2 Dec 2026.
- ✓Prohibited practices sit in the top fine tier: €35M or 7% of worldwide turnover.
- Prohibited AI practice:
- An AI use banned outright by Art. 5 because it's incompatible with EU values and fundamental rights.
- Social scoring:
- Evaluating people over time based on behavior or personal traits in a way that leads to unjustified or disproportionate detrimental treatment.
- Real-time remote biometric identification:
- Identifying people at a distance, live, by comparing their biometric data (usually faces) with a database, without their active involvement.
- Emotion recognition system:
- AI that identifies or infers people's emotions or intentions from biometric data such as facial expressions or voice.
- Untargeted scraping:
- Indiscriminately harvesting facial images from the internet or CCTV to build or expand a facial-recognition database.
Exam tips: Emotion recognition is the classic near-miss: banned at work and in education (except medical or safety reasons), high-risk elsewhere. · The real-time biometric ban targets law enforcement in publicly accessible spaces. After-the-fact ('post') identification is high-risk, not banned. · Neither consent nor a CE mark makes a prohibited practice lawful. · Dates: original bans since 2 Feb 2025; the two Omnibus bans from 2 Dec 2026.
- ✓Door 1 (Annex I): AI in regulated products that need third-party conformity assessment.
- ✓Door 2 (Annex III): eight sensitive areas, from biometrics to justice and democracy, with specific carve-outs.
- ✓The Art. 6(3) exit covers narrow, procedural or preparatory tasks, must be documented and registered, and never applies to profiling.
- ✓Art. 50 transparency: chatbots disclose, generative outputs are marked, deepfakes and emotion recognition are disclosed, from 2 Aug 2026.
- Annex I route:
- AI that is, or is a safety component of, a product covered by listed EU product-safety laws and requiring third-party conformity assessment (Art. 6(1)).
- Annex III route:
- AI used in one of eight listed sensitive areas, such as employment, credit, education or law enforcement (Art. 6(2)).
- Safety component:
- A part of a product or system that performs a safety function, or whose failure endangers health, safety or property.
- Art. 6(3) derogation:
- The exit from Annex III high-risk status for systems posing no significant risk, such as narrow procedural or preparatory tasks; never available if the system profiles people.
- Profiling:
- Automated processing of personal data to evaluate aspects of a person, such as work performance, finances, health, preferences or behavior.
- Deepfake:
- AI-generated or manipulated image, audio or video that resembles real people, places or events and would falsely appear authentic.
Exam tips: Profiling of natural persons means always high-risk under Annex III. The Art. 6(3) exit is closed. · Fraud detection is carved out of the credit-scoring entry; car insurance isn't listed (only life and health). · A human in the loop doesn't change the tier. Classification follows intended purpose. · Medical devices, toys, machinery, cars: think Annex I. Hiring, credit, education, benefits: think Annex III.
- ✓Seven requirements, Arts. 9–15: risk management, data governance, technical documentation, record-keeping, transparency to deployers, human oversight, accuracy/robustness/cybersecurity.
- ✓Risk management covers intended use and reasonably foreseeable misuse, for the system's whole life.
- ✓Data must be relevant, representative, as error-free as possible, and examined for bias.
- ✓Logs at least six months; technical documentation 10 years.
- Risk management system (Art. 9):
- A continuous, iterative process across the life cycle to identify, evaluate and mitigate risks to health, safety and fundamental rights.
- Reasonably foreseeable misuse:
- Use of an AI system in a way its provider didn't intend but which could result from predictable human behavior or interaction with other systems.
- Technical documentation (Annex IV):
- The dossier a provider draws up before market placement to show a high-risk system meets the requirements.
- Automation bias:
- The tendency to over-rely on an automated system's output, even when it's wrong.
- Instructions for use:
- Information a provider must give deployers about a high-risk system's purpose, performance, limitations, oversight and maintenance (Art. 13).
Exam tips: Risk management under Art. 9 is continuous and life-cycle long. Any option framing it as a one-time pre-launch check is wrong. · Disclosing a known risk in the instructions doesn't discharge the duty to mitigate it. · Logs = traceability (Art. 12). Technical documentation = proof of compliance (Art. 11). Instructions = what deployers need (Art. 13). · Human oversight is a design feature the provider builds in AND a duty the deployer staffs.
- ✓Path to market: QMS → conformity assessment → declaration → CE marking → registration → launch.
- ✓Most Annex III systems self-assess; notified bodies handle some biometrics and Annex I products.
- ✓After launch: post-market monitoring, corrective action, and serious-incident reporting within 15/10/2 days.
- ✓Providers register Annex III systems in the EU database; public-authority deployers register their use.
- Quality management system (QMS):
- A provider's documented policies and procedures ensuring a high-risk system is designed, built, tested and monitored in compliance with the Act (Art. 17).
- Conformity assessment:
- The process of showing a high-risk system meets the requirements, either by internal control or with a notified body (Art. 43).
- Notified body:
- An independent, accredited organization designated by a Member State to assess the conformity of certain high-risk systems.
- CE marking:
- The mark a provider affixes to show a high-risk AI system conforms to the AI Act's requirements (Art. 48).
- Post-market monitoring:
- A provider's ongoing, documented collection and analysis of a system's real-world performance throughout its life (Art. 72).
- Serious incident:
- An incident or malfunction leading to death or serious harm to health, serious disruption of critical infrastructure, breach of fundamental-rights law, or serious harm to property or the environment.
Exam tips: The default conformity route for Annex III is internal control. Notified bodies appear for biometrics (without harmonized standards) and Annex I products. · FRIA is a deployer duty. Conformity assessment, declaration, CE marking, registration and post-market monitoring are provider duties. · Incident clocks: 15 days generally, 10 days for a death, 2 days for widespread infringement or critical-infrastructure disruption. · CE marking isn't a lifetime pass: a substantial modification means a new conformity assessment.
- ✓Art. 26: follow instructions, staff human oversight, check input data, monitor, suspend if risky, report incidents, keep logs, inform workers and affected people.
- ✓Art. 27 FRIA before first use for public bodies, public-service providers, and credit or life/health insurance deployers; notify the authority.
- ✓AI literacy (Art. 4) applies to all providers and deployers; the Omnibus softened it to supporting literacy.
- ✓Art. 86 gives affected people a right to an explanation of the AI's role in significant decisions.
- Deployer:
- A person or organization using an AI system under its own authority, except for purely personal, non-professional use.
- Fundamental rights impact assessment (FRIA):
- An assessment certain deployers must complete before first using a high-risk system, describing affected groups, risks of harm and mitigation (Art. 27).
- AI literacy (Art. 4):
- Skills, knowledge and understanding that let people make informed use of AI and grasp its risks; providers and deployers must take measures to support it.
- Right to explanation (Art. 86):
- An affected person's right to a clear, meaningful explanation from the deployer of the AI's role in a significant decision based on an Annex III high-risk system.
Exam tips: FRIA = deployer duty, only for public bodies, private providers of public services, and credit-scoring or life/health-insurance deployers. · A deployer can't hide behind the vendor's CE mark. The mark covers the provider's obligations, not how you use the system. · Deployers keep the logs under their control for at least six months and must inform workers before workplace use. · AI literacy covers every AI system, applied from 2 Feb 2025, and since the Omnibus means 'take measures to support', not 'guarantee a level'.
- ✓Provider (builds and certifies), deployer (uses), importer (brings non-EU systems in), distributor (resells), authorized rep (acts for a non-EU provider).
- ✓Importers and distributors check CE marks, declarations and instructions, and stop nonconforming systems.
- ✓Art. 25: own name, substantial modification, or a new high-risk purpose makes you the provider; the original provider must cooperate.
- ✓Building for internal use still makes you a provider.
- Provider:
- Whoever develops an AI system or GPAI model, or has it developed, and places it on the market or puts it into service under its own name or trademark.
- Importer:
- An EU-established entity that places on the EU market an AI system bearing the name or trademark of a provider established outside the EU.
- Distributor:
- Any supply-chain actor, other than the provider or importer, that makes an AI system available on the EU market.
- Authorized representative:
- An EU-established entity with a written mandate from a non-EU provider to carry out specified obligations on its behalf.
- Substantial modification:
- A change after market placement, not foreseen in the original conformity assessment, that affects compliance or changes the intended purpose.
Exam tips: Roles follow conduct, not contracts or job titles. Rebrand, substantially modify, or repurpose into high-risk, and you're the provider (Art. 25). · Importer = EU entity placing a non-EU provider's branded system on the EU market. Distributor = anyone else making it available. · Distributors and importers verify; they don't re-certify. · One company can be provider and deployer at once, and owes both sets of duties.
- ✓All GPAI providers: technical documentation, downstream information, copyright policy, training-content summary.
- ✓Systemic risk (presumed above 10²⁵ FLOPs): add evaluation and adversarial testing, risk mitigation, incident reporting, cybersecurity.
- ✓The GPAI Code of Practice (July 2025) is the recognized way to demonstrate compliance; the AI Office enforces, with fines from 2 Aug 2026.
- ✓Applies since 2 Aug 2025; legacy models by 2 Aug 2027; not deferred by the Omnibus.
- GPAI model:
- An AI model with significant generality that can competently perform a wide range of distinct tasks and be integrated into many downstream systems.
- GPAI system:
- An AI system based on a general-purpose AI model, able to serve a variety of purposes directly or when integrated into other systems.
- Systemic risk:
- Risk from high-impact GPAI capabilities that could significantly affect the EU market, public health, safety, security or fundamental rights at scale; presumed above 10²⁵ FLOPs.
- Training-content summary:
- A public, sufficiently detailed summary of the data used to train a GPAI model, prepared using the Commission's template (Art. 53(1)(d)).
- GPAI Code of Practice:
- A voluntary code (July 2025) with transparency, copyright, and safety and security chapters that providers can rely on to demonstrate compliance.
Exam tips: 10²⁵ FLOPs = EU systemic-risk presumption. 10²⁶ = California SB 53 frontier models (and Korea's 'high-performance' AI). Don't swap them. · The open-source exemption is partial: the copyright policy and training summary are still required, and nothing is exempt for systemic-risk models. · GPAI models aren't CE-marked or entered in the high-risk database. The AI Office supervises them. · A GPAI-based system used for a high-risk purpose is high-risk; the model provider still has only GPAI duties for the model.
- ✓Two-level enforcement: the AI Office centrally for GPAI; national market surveillance authorities for AI systems.
- ✓The AI Board, scientific panel and advisory forum advise and support; notifying authorities oversee notified bodies.
- ✓Fines: €35M/7%, €15M/3%, €7.5M/1%; higher of the two for large firms, lower for SMEs (and small mid-caps for the lower two tiers).
- ✓Anyone can complain to a market surveillance authority; every Member State must offer a regulatory sandbox.
- AI Office:
- The European Commission's AI body that supervises GPAI models and, after the Omnibus, certain AI systems built on them; it can investigate and fine GPAI providers.
- European AI Board:
- A body of Member State representatives that advises the Commission and promotes consistent application of the AI Act; it has no fining power.
- Market surveillance authority:
- A national authority that monitors AI systems on its market and can investigate, order corrective action or withdrawal, and impose fines.
- AI regulatory sandbox:
- A controlled framework set up by a competent authority where providers can develop and test AI under regulatory supervision.
- SME penalty rule:
- For SMEs and start-ups, each AI Act fine is capped at the lower of the fixed amount and the turnover percentage (Art. 99(6)).
Exam tips: Ladder: 35/7 prohibited · 15/3 most obligations (including GPAI) · 7.5/1 misleading information. Expect swapped numbers as distractors. · SMEs and start-ups pay the LOWER of the two amounts; everyone else the HIGHER. · AI Office = GPAI models (and, post-Omnibus, AI systems built on a provider's own GPAI model). National authorities = most AI systems. The AI Board coordinates; it doesn't fine. · The GDPR's €20M / 4% is a separate regime. Don't mix the caps.
- ✓Entry into force 1 Aug 2024; bans and AI literacy from 2 Feb 2025; GPAI from 2 Aug 2025; general application and Art. 50 from 2 Aug 2026.
- ✓The Digital Omnibus (Reg. (EU) 2026/1744, in force 27 Jul 2026) deferred Annex III to 2 Dec 2027 and Annex I to 2 Aug 2028.
- ✓New bans and the end of the legacy marking grace both fall on 2 Dec 2026.
- ✓Know both the original and amended dates; the exam outline was written before the Omnibus.
- Digital Omnibus on AI:
- Regulation (EU) 2026/1744, in force 27 July 2026, which deferred high-risk dates, added two prohibitions and simplified parts of the AI Act.
- Entry into force vs application:
- A law can be in force (legally existing) before its obligations apply; the AI Act entered into force on 1 Aug 2024 but applies in stages.
- Marking grace period:
- Extra time, until 2 Dec 2026, for generative AI systems placed on the market before 2 Aug 2026 to add machine-readable marking.
- Legacy system:
- An AI system or model already on the market before the relevant application date, which may benefit from transitional rules.
Exam tips: Never moved: prohibitions (Feb 2025), GPAI (Aug 2025), Art. 50 transparency (Aug 2026). · Moved: Annex III → 2 Dec 2027; Annex I → 2 Aug 2028. · The 2 Dec 2026 marking grace is only for generative systems already on the market before 2 Aug 2026. · If an option uses the original 2 Aug 2026 high-risk date, check whether the question asks about the law as first adopted. BoK v2.1 predates the Omnibus.
- ✓Art. 4 literacy, Art. 5 bans, Art. 6 classification, Arts. 9–15 requirements, Art. 16–17 provider duties and QMS.
- ✓Arts. 22–27: authorized reps, importers, distributors, role flips, deployers, FRIA.
- ✓Arts. 43–50: conformity, declaration, CE, registration, transparency.
- ✓Arts. 51–56 GPAI; 72–73 monitoring and incidents; 85–86 complaints and explanations; 99/101 penalties.
- Annex I:
- The list of EU product-safety laws (medical devices, toys, machinery, vehicles and more) whose regulated products can make embedded AI high-risk.
- Annex III:
- The list of eight sensitive areas of use that make an AI system high-risk, such as employment, credit and education.
- Annex IV:
- The required contents of a high-risk system's technical documentation.
- Chapter V:
- The part of the AI Act (Arts. 51–56) setting obligations for general-purpose AI models.
Exam tips: Common swaps: Art. 23 importers vs Art. 24 distributors; Art. 53 (all GPAI) vs Art. 55 (systemic risk); Art. 26 deployers vs Art. 27 FRIA. · Arts. 9–15 run in a fixed order: risk, data, docs, logs, instructions, oversight, robustness. · Fines live in Art. 99 (everyone) and Art. 101 (GPAI providers).
- ✓In force 22 Jan 2026: the second comprehensive AI law after the EU's.
- ✓Duties attach to high-impact AI, generative AI and high-performance (10²⁶ FLOPs) AI.
- ✓Self-assessment, meaningful explanations, user protection and human oversight; generative outputs labeled.
- ✓Enforced by MSIT; fines up to ₩30M; foreign operators above thresholds need a domestic representative.
- High-impact AI (Korea):
- AI that may significantly affect human life, physical safety or fundamental rights in listed areas such as health care, hiring, lending and public services.
- AI business operator:
- Under Korea's Act, an entity that develops AI or uses AI to provide products or services.
- Domestic representative:
- A local agent that foreign AI operators above set thresholds must appoint to deal with Korean authorities.
- MSIT:
- South Korea's Ministry of Science and ICT, which leads implementation and enforcement of the AI Basic Act.
- High-performance AI (Korea):
- AI trained with cumulative compute of 10²⁶ FLOPs or more, which carries life-cycle safety duties.
Exam tips: 'High-impact' = Korea; 'high-risk' = EU. The stem's vocabulary tells you which law. · Korea has no Article 5-style banned list and no mandatory third-party conformity assessment. · Fines top out at ₩30 million, and MSIT announced a grace period of at least a year from January 2026. · Light-touch isn't no-touch: generative AI labeling and high-impact duties still apply.
- ✓US federal AI governance = existing laws via existing agencies + targeted statutes + executive policy.
- ✓TAKE IT DOWN Act (signed 19 May 2025): NCII crimes including deepfakes; platforms remove within 48 hours; FTC enforcement from 19 May 2026.
- ✓Executive posture: EO 14179 and the AI Action Plan (deregulatory), EO 14365 (preemption push via litigation and funding).
- ✓No federal preemption statute yet, so state AI laws still apply.
- TAKE IT DOWN Act:
- A 2025 federal law criminalizing publication of non-consensual intimate imagery, including AI deepfakes, and requiring platforms to remove it within 48 hours of a valid request.
- Non-consensual intimate imagery (NCII):
- Intimate images of a person shared without their consent, whether authentic or AI-generated.
- Covered platform:
- Under TAKE IT DOWN, a public website, online service or app that primarily hosts user-generated content.
- Preemption:
- Federal law overriding conflicting state law; it requires Congress, a valid federal regulation or a court, not an executive order alone.
- EO 14365:
- The December 2025 executive order seeking a national AI policy framework and creating a DOJ task force to challenge state AI laws.
Exam tips: No comprehensive federal AI law. Expect answers built on existing laws (FTC Act, civil-rights, credit) plus the TAKE IT DOWN Act. · TAKE IT DOWN: 48 calendar hours, covers AI deepfakes, FTC enforces the platform duty since 19 May 2026. · Executive orders can't preempt state law on their own; state AI laws stay enforceable unless Congress or a court says otherwise. · The National AI Initiative Act coordinates federal AI R&D. It's the classic 'which law does NOT regulate companies' answer.
- ✓Colorado SB 26-189 (from 1 Jan 2027) replaced the never-effective SB 24-205 with ADMT notice, explanation, correction and human-review duties.
- ✓Texas TRAIGA (from 1 Jan 2026): intent-based prohibitions, government-only social-scoring ban, AG enforcement, sandbox.
- ✓California: SB 53 frontier safety (10²⁶ operations), AB 2013 training data, SB 942 provenance, CCPA ADMT rules.
- ✓Hiring: NYC LL 144 bias audits and notice; Illinois HB 3773 and the AI Video Interview Act. Utah requires AI disclosure.
- Automated decision-making technology (ADMT):
- Technology that processes personal data to produce outputs, such as scores or recommendations, used to make or assist decisions about people.
- Consequential decision:
- A decision with significant effects on access to things like employment, housing, credit, insurance, health care or education.
- Intent-based liability:
- Liability that requires proof the AI was built or used for a prohibited purpose; disparate impact alone isn't enough (the Texas model).
- Frontier developer (SB 53):
- A developer that trained a foundation model using more than 10²⁶ operations; 'large' frontier developers also exceed $500M in annual revenue.
- Bias audit (NYC LL 144):
- An annual, independent evaluation of an automated employment decision tool's impact by sex and race/ethnicity, with a public summary.
Exam tips: Colorado = impact (what happened); Texas = intent (what you meant). Both: Attorney General only, with a 60-day cure. · A 'duty of reasonable care against algorithmic discrimination' with impact assessments describes Colorado's repealed SB 24-205. BoK v2.1 was approved before the repeal, so know both. · California trio: SB 53 = frontier safety; AB 2013 = training data; SB 942 = provenance and detection. New York's RAISE Act (from 1 Jan 2027) mirrors SB 53 with a stricter 72-hour incident clock. · NYC LL 144 needs an independent auditor; an internal review doesn't count.
- ✓Many guides predate the final AI Act, the 2025 withdrawals and the 2026 Omnibus, so check their legal facts.
- ✓AI Act: in force 1 Aug 2024; fines €35M / 7%, €15M / 3%, €7.5M / 1%; SMEs pay the lower.
- ✓Gone: EO 14110 (revoked), Canada's AIDA (died), the AI Liability Directive (withdrawn), Colorado SB 24-205 (replaced).
- ✓The NIST AI RMF has a Core and Profiles, not implementation tiers.
- ✓Older guides still teach useful concepts; verify their law against current sources.
- Draft vs. final text:
- A legislative proposal or negotiating draft can change before adoption; only the final published text binds, so figures quoted from drafts go stale.
- Withdrawn proposal:
- A legislative proposal the European Commission formally abandons, such as the AI Liability Directive in 2025; it never becomes law.
- Revoked executive order:
- A presidential order cancelled by a later order, as EO 14110 was in January 2025; it no longer directs federal agencies.
- Implementation tiers:
- A NIST Cybersecurity Framework feature describing how rigorous an organization's practices are; the AI RMF 1.0 has no tiers.
- Artificial Intelligence and Data Act (AIDA):
- Canada's proposed federal AI law, part of Bill C-27, which died when Parliament was prorogued in January 2025.
Exam tips: Draft figures (€30M / 6%, €20M / 4%) are distractors. The final ladder is €35M / 7%, €15M / 3%, €7.5M / 1%. · If an option treats AIDA, the AI Liability Directive or EO 14110 as live, it's almost certainly wrong. · Art. 50 limited-risk duties are disclosure and marking. 'Must offer an opt-out' is a trap. · High-risk dates: the amended law says 2 Dec 2027 (Annex III) and 2 Aug 2028 (Annex I). BoK v2.1 predates the change, so read the stem's framing.
Understand the main industry standards and tools that apply to AI.
- ✓Five principles: inclusive growth and well-being; human rights and democratic values; transparency and explainability; robustness, security and safety; accountability.
- ✓Five recommendations for governments: R&D, ecosystem, interoperable governance, human capacity, international cooperation.
- ✓Adopted May 2019, updated May 2024; 47 adherents; non-binding but widely adopted.
- ✓Practical OECD tools: classification framework, tools and metrics catalog, AI Incidents Monitor, Policy Observatory, Hiroshima reporting framework.
- OECD AI Principles:
- The 2019 (updated 2024) OECD Council Recommendation setting five values-based principles for trustworthy AI and five recommendations for governments.
- AI actor:
- Anyone who plays an active role in the AI system life cycle, including organizations and individuals that deploy or operate AI.
- Framework for the Classification of AI Systems:
- An OECD tool that profiles AI systems across five dimensions to connect their characteristics with policy and risk considerations.
- Catalogue of Tools & Metrics for Trustworthy AI:
- An OECD.AI library of tools and metrics that help put the principles into practice.
- AI Incidents Monitor (AIM):
- An OECD database tracking AI incidents and hazards reported in the news, to build evidence for policy.
Exam tips: Principles = AI actors; recommendations = governments. Questions often swap them. · The OECD Principles are voluntary soft law but foundational: the EU AI Act's AI-system definition comes from the OECD. · Frameworks and standards (NIST AI RMF, ISO/IEC 42001) operationalize principles. Principles alone don't tell you how. · 2024 update themes to recognize: safety (override, repair, decommission), mis- and disinformation, interoperability.
- ✓AI RMF 1.0 (Jan 2023): voluntary, flexible, for any sector and use case; a revision is underway but 1.0 is current.
- ✓Four functions: GOVERN (cross-cutting), MAP (context), MEASURE (assess and track), MANAGE (prioritize and act).
- ✓Harms to people, organizations and ecosystems; challenges in measurement, tolerance, prioritization and integration.
- ✓Seven trustworthy characteristics, with trade-offs balanced in context.
- NIST AI RMF:
- NIST's voluntary AI Risk Management Framework (AI 100-1, January 2023), organized around GOVERN, MAP, MEASURE and MANAGE.
- GOVERN:
- The cross-cutting function that builds the policies, accountability, culture and processes the other functions rely on.
- MAP:
- The function that establishes context and identifies risks and impacts of an AI system.
- MEASURE:
- The function that analyzes, assesses, benchmarks and monitors AI risks and trustworthiness.
- MANAGE:
- The function that prioritizes risks and allocates resources to treat, respond to, recover from and communicate about them.
- Risk tolerance:
- The level of risk an organization is willing to bear to achieve its objectives; the AI RMF leaves it to each organization.
Exam tips: GOVERN is cross-cutting, not 'step one then done'. Any option treating it as a one-time gate is a trap. · Seven characteristics; 'valid and reliable' is the base and 'accountable and transparent' cuts across the rest. 'Profitable', 'autonomous' or 'cost-effective' are never on the list. · What = transparency; how = explainability; why = interpretability. · Voluntary and no certification, and no implementation tiers (that's the Cybersecurity Framework).
- ✓Each function breaks into outcome-based categories and subcategories (19 and 72).
- ✓The Playbook gives suggested actions, documentation prompts and references per subcategory; it's not a checklist.
- ✓Use-case, temporal (current vs target) and cross-sectoral profiles tailor the framework.
- ✓The GenAI Profile (NIST AI 600-1) adds 12 GenAI risks; NIST's ARIA program supports measurement through model testing, red teaming and field testing.
- Category and subcategory:
- The outcome statements each AI RMF function breaks into: 19 categories and 72 subcategories in total.
- AI RMF Playbook:
- NIST's voluntary online companion offering suggested actions, documentation prompts and references for every subcategory.
- Profile (current vs target):
- A tailored application of the AI RMF; comparing a current profile with a target profile reveals gaps to address.
- NIST AI 600-1 (Generative AI Profile):
- A July 2024 cross-sectoral profile naming 12 risks unique to or worsened by generative AI, with actions mapped to the AI RMF.
- Confabulation:
- Generative AI confidently producing false or fabricated content; one of the 12 GenAI Profile risks.
Exam tips: Numbers: 4 functions, 19 categories, 72 subcategories. GOVERN has the most categories (6). · Framework = outcomes (what). Playbook = suggested actions (how). Profile = tailored to a use case, a sector, or current versus target state. · GenAI Profile = NIST AI 600-1, July 2024, 12 risks (confabulation, information integrity, CBRN information, value chain and more). · Texas TRAIGA's defense refers to substantially complying with the NIST GenAI Profile or another recognized framework, a sign of how laws lean on NIST.
- ✓ISO/IEC 22989 (2022): shared AI terms, stakeholder roles and life-cycle stages.
- ✓ISO/IEC 42001 (2023): certifiable AI management system; PDCA, risk and impact assessment, Statement of Applicability, Annex A controls.
- ✓ISO/IEC 42005 (2025): guidance for AI system impact assessments on individuals, groups and society.
- ✓They work together, and with 23894 for risk management.
- ISO/IEC 22989:
- The 2022 international standard defining AI concepts and terminology, including stakeholder roles and life-cycle stages.
- ISO/IEC 42001 (AIMS):
- The 2023 certifiable standard setting requirements for an AI management system, using Plan-Do-Check-Act.
- ISO/IEC 42005:
- The 2025 guidance standard on assessing an AI system's impacts on individuals, groups and society.
- Statement of Applicability:
- A 42001 document listing which Annex A controls an organization applies and why any are excluded.
- Plan-Do-Check-Act (PDCA):
- The continual-improvement cycle behind ISO management-system standards.
- ISO/IEC 23894:
- Guidance on AI risk management, adapting ISO 31000 to AI.
Exam tips: 22989 = words, 42001 = system (certifiable), 42005 = impact (guidance). · Only 42001 is certifiable. Any option offering 42005 or 22989 'certification' is a trap. · 42001 requires an AI system impact assessment; 42005 tells you how to do one well. · Standards support legal compliance but don't equal it: 42001 certification isn't EU AI Act conformity.
Independent study aid, not affiliated with the IAPP. Written with AI from primary sources; laws change, so check the sources. Last updated 7 Nov 2026.
AIGP Quest