In November 2023, the Swedish data protection authority issued a €12.3 million fine against Spotify for GDPR violations related to inadequate disclosure of how user data was being processed — including data fed into recommendation algorithms. The case illustrated a structural tension that now confronts every business deploying AI: the right of individuals to understand decisions made about them collides directly with the opacity of machine learning systems.
The EU's General Data Protection Regulation, operative since May 2018, was not written with transformers or gradient-boosted trees in mind. Yet its principles — lawfulness, transparency, data minimisation, purpose limitation — apply with full legal force to every AI system that processes personal data. Businesses that treat GDPR as a cookie-consent checkbox while deploying AI-driven hiring, pricing, or credit tools are accumulating significant unpriced legal risk.
Article 22 of the GDPR grants individuals the right not to be subject to a decision based solely on automated processing where that decision produces legal or similarly significant effects. This includes credit scoring, insurance underwriting, recruitment screening, and loan origination. When such processing does occur — with consent or contractual necessity — the data controller must provide meaningful information about the logic involved, the significance, and the envisaged consequences.
The word "meaningful" has generated enormous legal debate. In 2022, the Austrian Data Protection Authority ruled that a fitness app's transfer of user data to Google Analytics — even anonymised data — constituted a violation because it could not guarantee data wouldn't be accessed under US surveillance law. The principle extends to AI: if you cannot explain the variables your model uses, their weights, and how they combine to produce an output about a specific individual, you may be unable to demonstrate GDPR compliance.
The practical implication for business leaders is stark: deploying a black-box model against EU residents for consequential decisions is a compliance gamble. Interpretable models, post-hoc explanation tools (SHAP, LIME), and documented model cards are now elements of legal risk management — not merely good ML engineering practice.
Machine learning models generally improve with more data. GDPR's data minimisation principle says you should collect only what is adequate, relevant, and limited to the processing purpose. These two imperatives are in direct tension. Businesses building or fine-tuning AI systems on customer data must conduct Data Protection Impact Assessments (DPIAs) — mandatory under Article 35 for high-risk processing — before deployment, not after an incident.
In 2022, the Italian Data Protection Authority (Garante) ordered Clearview AI to stop processing Italian residents' facial recognition data and imposed a €20 million fine — the maximum permitted under GDPR. Clearview had scraped billions of public images to train a facial recognition system without any lawful basis for that processing. The case established that training data provenance is itself a compliance question, not merely a downstream application issue. If you fine-tune a model on customer data, you must have a lawful basis for each processing activity involved in that fine-tuning.
Between 2018 and 2024, GDPR fines exceeded €4.5 billion in aggregate. Meta has been fined over €1.3 billion for transatlantic data transfers alone. The Irish Data Protection Commission and French CNIL have both opened investigations explicitly targeting AI training data practices. Fines can reach 4% of global annual turnover — for a company with €10 billion revenue, that is €400 million at maximum exposure.
Six lawful bases exist under GDPR Article 6. For AI deployments the most commonly invoked are: legitimate interests (Article 6(1)(f)), contract performance (6(1)(b)), and consent (6(1)(a)). Each carries different obligations. Legitimate interests requires a balancing test — the controller's interest must not override the fundamental rights of the data subject. Consent must be freely given, specific, informed, and unambiguous; bundling AI profiling consent into terms of service has been repeatedly rejected by regulators.
For special category data — health, biometric, racial or ethnic origin — the bar is higher still: you need an Article 9 condition in addition to an Article 6 basis. AI systems that infer health status from behavioural data, or that use facial recognition, almost certainly process special category data. The 2023 enforcement action against TikTok by Ireland's DPC, resulting in a €345 million fine, highlighted how AI-driven recommendation systems that process children's data attract heightened scrutiny even when data is nominally behavioural rather than explicitly categorised.
DPIA (Data Protection Impact Assessment): Mandatory pre-deployment risk assessment for high-risk AI processing under GDPR Article 35. Must identify risks to data subject rights and describe mitigation measures.
Article 22: Right not to be subject to solely automated decisions with significant effects. Requires human review on request and meaningful explanation of logic.
Lawful Basis: One of six GDPR-recognised justifications for processing personal data. Without one, processing is unlawful regardless of business purpose.
You are a senior legal or compliance officer advising on AI deployments. Use this lab to stress-test real scenarios against GDPR requirements: lawful basis, Article 22 obligations, DPIA triggers, special category data risks, and enforcement exposure. Describe your AI use case and get structured compliance analysis.
The EU AI Act entered into force on 1 August 2024, becoming the world's first comprehensive binding legal framework for artificial intelligence. Its prohibitions on unacceptable-risk AI began applying six months later; obligations for general-purpose AI model providers began in August 2025; and the full high-risk AI system requirements apply from August 2026. For businesses already deploying AI or planning to deploy, the compliance clock is running.
Unlike GDPR — which is fundamentally about data — the AI Act regulates the AI system itself: how it is built, deployed, monitored, and documented. It imposes different obligations on providers (those who develop or place AI on the market) and deployers (those who use AI in a professional context). Crucially, a company that purchases an AI system from a vendor and deploys it in a consequential context may still bear deployer obligations — including maintaining human oversight and conducting fundamental rights impact assessments for certain high-risk systems.
Unacceptable Risk (Prohibited): A small category of AI practices banned outright from the EU market. These include social scoring by public authorities, real-time remote biometric identification in public spaces (with narrow law enforcement exceptions), subliminal manipulation that impairs free will, and AI that exploits vulnerabilities of specific groups. Companies building or selling these systems into EU markets face penalties of up to €35 million or 7% of global annual turnover.
High Risk: The largest and most consequential category for business. High-risk AI systems are those used in: critical infrastructure, education and vocational training, employment and HR (hiring, promotion, task management), essential private services (credit scoring, insurance risk assessment), migration and border control, administration of justice, and systems that are safety components of products regulated under EU product law. All high-risk systems must comply with extensive requirements before market placement.
Limited Risk: Systems with specific transparency obligations — primarily chatbots and deepfake-generating systems. Users must be informed they are interacting with AI or viewing AI-generated content. No pre-market compliance requirements, but post-deployment transparency is mandatory.
Minimal Risk: The vast majority of AI applications — spam filters, AI-powered spreadsheet features, most recommendation engines not falling into high-risk categories. No mandatory compliance obligations beyond existing law.
If your organisation is a provider of a high-risk AI system — or a deployer of one used in an HR, credit, or essential services context — the AI Act imposes a suite of obligations that represent a significant operational and legal commitment:
Risk Management System: A documented, continuous process identifying and mitigating risks throughout the AI system's lifecycle. Not a one-time assessment — a living programme.
Data Governance: Training, validation, and test datasets must be relevant, representative, free of errors, and complete. This requirement alone has significant implications for organisations using historical datasets that may embed past discriminatory practices.
Technical Documentation: Providers must maintain detailed technical documentation demonstrating compliance. This documentation must be available to regulators on request.
Transparency and User Information: High-risk systems must provide deployers with instructions for use, including the system's purpose, accuracy levels, human oversight measures, and known limitations.
Human Oversight: High-risk systems must be designed to allow natural persons to understand, monitor, and override automated outputs. This is not optional human-in-the-loop theatre — it must be effective and documented.
Accuracy, Robustness, and Cybersecurity: Systems must achieve appropriate levels of performance and be resilient to adversarial inputs, errors, and inconsistencies.
Conformity Assessment: Before market placement, high-risk systems must undergo conformity assessment — either self-assessment or third-party assessment depending on the category. A CE mark equivalent is required.
Under the AI Act, providers (developers/vendors) bear primary compliance obligations. But deployers — companies that put the AI to work in their business — carry obligations too: conducting fundamental rights impact assessments for high-risk AI, ensuring effective human oversight, logging system operation, and reporting serious incidents. A company that buys an off-the-shelf AI hiring tool and uses it without modification is a deployer — and is legally responsible for how it deploys that tool.
The AI Act includes a dedicated regime for general-purpose AI (GPAI) models — large language models and foundation models that can be adapted to a wide range of downstream tasks. Providers of all GPAI models must maintain technical documentation, comply with EU copyright law, and publish summaries of training data. Providers of models with systemic risk — currently defined as models trained using more than 10^25 FLOPs — face additional obligations: model evaluation, adversarial testing, incident reporting to the European AI Office, and cybersecurity measures.
As of 2024, the models in scope for systemic risk designation included GPT-4, Gemini Ultra, Claude 3, and Llama 3 400B. Businesses building products on top of these models are downstream deployers — they inherit obligations but also need to understand what their GPAI provider is and is not warranting about the underlying model's compliance.
Prohibited AI practices: Up to €35 million or 7% of global annual turnover, whichever is higher.
High-risk system non-compliance: Up to €15 million or 3% of global annual turnover.
Providing incorrect information to authorities: Up to €7.5 million or 1% of global annual turnover.
For SMEs, the lower of the monetary amount or percentage applies. Penalties are administered by national market surveillance authorities coordinated by the European AI Office.
Use this lab to classify your AI systems or proposed deployments under the EU AI Act's four-tier risk architecture. Describe the AI system — its purpose, who uses it, what decisions it informs, and whether it operates in the EU market. Get a risk-tier classification with rationale, the specific legal obligations that apply, and a prioritised action list for compliance.
Amazon built a machine learning hiring tool designed to automate CV screening. Trained on ten years of submitted applications — the vast majority from men, reflecting the male-dominated tech industry — the model learned to systematically downgrade CVs from women. It penalised CVs that included the word "women's" (as in "women's chess club") and downgraded graduates of all-women's colleges. Amazon scrapped the project in 2018, having never used the tool in actual hiring decisions. But the episode became one of the most cited examples in AI bias literature — a system learned to replicate historical discrimination with algorithmic precision.
Amazon's case was a self-discovered internal failure. Other organisations have faced external enforcement. In 2023, the EEOC — the US Equal Employment Opportunity Commission — published guidance explicitly warning that employers using AI for hiring decisions remain liable under Title VII, the ADA, and the ADEA even if the discriminatory output was generated by a third-party vendor's algorithm. The vendor does not absorb your liability.
Anti-discrimination law in both the US and EU distinguishes between disparate treatment (intentional discrimination) and disparate impact (facially neutral practices that disproportionately harm a protected class). The disparate impact doctrine — established by the US Supreme Court in Griggs v. Duke Power Co. (1971) and embedded in EU Equal Treatment Directives — does not require proof of discriminatory intent. If an AI selection tool produces statistically significant adverse effects on a protected group, and cannot be justified as job-related and consistent with business necessity, it may constitute illegal discrimination.
The four-fifths rule (or 80% rule), used by the EEOC as a rule of thumb, flags adverse impact when a protected group's selection rate is less than 80% of the rate for the group with the highest selection rate. An AI hiring tool that selects 40% of white male applicants but only 25% of Black female applicants has an adverse impact ratio of 62.5% — well below the 80% threshold. The burden then shifts to the employer to demonstrate job-relatedness and business necessity, and that no equally valid, less discriminatory alternative exists.
In 2023, the city of New York enacted Local Law 144, requiring employers using AI hiring tools to conduct and publish annual bias audits. The law was the first of its kind in the US and signalled a regulatory direction that has since been followed by Colorado (insurance algorithms), Illinois (AI video interview analysis), and Maryland (facial recognition in hiring) with their own AI-specific employment laws.
HireVue, a major AI-based video interview assessment platform, faced a 2019 complaint from the Electronic Privacy Information Center (EPIC) to the FTC, alleging that its system made employment assessments based on video and voice data using criteria that were not validated for job-relatedness. HireVue subsequently dropped facial recognition from its platform in 2021, citing algorithmic explainability concerns. The episode illustrates how regulatory and reputational pressure can force product pivots that significantly affect any business that has built on top of the vendor's capabilities.
For business leaders, the HireVue case generates two practical risk exposures: First, vendor dependency risk — a vendor product change can disrupt your hiring process midstream. Second, co-liability risk — using a third-party AI tool does not insulate you from discrimination claims. The EEOC's 2023 technical assistance document makes clear that employers are responsible for the discriminatory impact of any tool they use, even if they did not build it.
In 2022, the Austrian Labour Court ruled that an employer's use of an AI-based scheduling system that systematically assigned less desirable shifts to employees who had taken parental leave constituted indirect sex discrimination under EU law, since parental leave is disproportionately taken by women. The system was neutral on its face — it optimised operational coverage. The outcome was discriminatory. The court required the employer to demonstrate that the scheduling algorithm's business justification was proportionate and that no less discriminatory scheduling approach existed.
The legal and reputational risk of AI discrimination is asymmetric: the cost of an undetected discriminatory system that reaches regulatory attention or litigation is vastly higher than the cost of proactive bias testing. Business leaders deploying AI in HR, credit, insurance, or any decisioning role affecting protected groups should consider the following:
Pre-deployment disparate impact testing: Run your model's outputs against demographic breakdowns of your applicant or customer population before go-live. Document the analysis and the decisions made in response to any adverse impact findings.
Regular post-deployment audits: Populations shift, model drift occurs, and the interaction of your AI with real-world data may produce impacts not visible in pre-deployment testing. New York's Local Law 144 mandates annual independent bias audits with public disclosure — treat this as a template for best practice, regardless of jurisdiction.
Vendor contracts and representations: Require vendors of AI decisioning tools to provide bias audit results, demographic performance data across protected classes, and contractual representations about job-relatedness validation. Ensure your contract includes audit rights.
Human review processes: For high-stakes decisions — hiring, termination, credit denial — ensure documented human review exists. This does not mean a human rubber-stamps the AI; it means a human can explain the decision independently of the algorithmic output.
EEOC (Employment): Issued 2023 technical assistance on AI and Title VII/ADA/ADEA. Employers liable for discriminatory AI outputs regardless of vendor origin.
CFPB (Credit): 2022 circular clarified that "black box" credit models cannot be used to deny credit without providing specific, comprehensible adverse action reasons — explainability is a legal requirement under ECOA and FCRA.
HUD (Housing): Fair Housing Act applies to algorithmic property advertising targeting and tenant screening tools. Facebook settled a DOJ housing discrimination complaint in 2022 related to its ad targeting algorithms.
This lab helps you identify discrimination and bias risks in your AI deployment, understand your legal exposure under employment, credit, and fair housing laws, and develop practical mitigation strategies. Describe your AI use case — including what decisions it supports, what data it uses, and what population it affects — and get a structured risk and remediation analysis.
In December 2023, The New York Times filed suit against OpenAI and Microsoft in the Southern District of New York, alleging that millions of Times articles were used without authorisation to train ChatGPT and related models. The complaint included exhibits showing ChatGPT reproducing substantial portions of copyrighted Times articles verbatim — a phenomenon called "memorisation" in ML literature. The Times alleged both direct copyright infringement and contributory infringement, and sought statutory damages that, if awarded at maximum per-work rates, could theoretically reach billions of dollars.
The Times case is one of dozens of IP lawsuits now working through US courts. Artists' suits against Stability AI, Midjourney, and DeviantArt (Kelly v. Stability AI, filed 2023), musicians' suits regarding AI music generation trained on copyrighted recordings, and Getty Images' suit against Stability AI in both the US and UK courts collectively define the contours of the most consequential unresolved legal question in AI: does training a model on copyrighted material constitute infringement, and if so, what is the remedy?
US copyright law does not protect ideas, only expression. But training on copyrighted works involves copying that expression — temporarily or persistently — into model weights. The central legal question is whether this constitutes fair use under 17 U.S.C. § 107. The four-factor fair use analysis — purpose and character of use, nature of the copyrighted work, amount taken, and effect on the market — will be applied to AI training for the first time in binding court decisions expected between 2025 and 2027.
In the Google Books case (Authors Guild v. Google, 2d Cir. 2015), the Second Circuit found Google's scanning of millions of books for search indexing to be fair use — largely because the use was transformative and did not substitute for the original. AI training proponents argue a similar logic applies: a model doesn't reproduce a book, it learns patterns. Critics, including the Times, argue that memorisation of verbatim content, and competition with original creators in the AI output market, undermine the fair use defence.
Outside the US, the position is clearer and more restrictive. The EU's Directive on Copyright in the Digital Single Market (DSM Directive, 2019) permits text and data mining for research purposes but allows rightholders to opt out of TDM for commercial AI training. Several major publishers, including Axel Springer and Le Monde, have done exactly that. A business using an AI system trained on data scraped from EU sources without proper opt-out mechanisms may face copyright liability under EU law even if the training occurred outside the EU.
A separate but related question: who owns content generated by AI? The US Copyright Office has taken a clear position: copyright does not protect works generated by AI without sufficient human authorship. In 2023, the Office declined to register "A Recent Entrance to Paradise" — an image generated entirely by the AI system DABUS — and subsequent guidance confirmed that the degree of human creative control determines copyrightability. AI-generated images, text, or music with minimal human input cannot be copyright protected in the US.
This has significant commercial implications. If your organisation uses generative AI to produce marketing copy, product descriptions, software code, or creative assets, and those outputs lack sufficient human authorship, you may not be able to obtain copyright protection for them. Competitors could freely copy your AI-generated content without infringement. Your contracts with clients that warrant ownership of delivered work may be breached if the delivered work is unprotectable AI output.
For code specifically: GitHub Copilot has faced multiple lawsuits alleging that it reproduces copyrighted open-source code without attribution, in violation of licence terms. A 2022 class action (Doe v. GitHub/Microsoft/OpenAI) raised claims under the DMCA's Section 1202 (circumvention of copyright management information by stripping licence notices from reproduced code). The case signals that businesses using AI code generation tools face potential licence compliance exposure if the generated code is substantially similar to licensed open-source material.
Getty Images sued Stability AI in both Delaware (US) and the UK in 2023, alleging that Stable Diffusion was trained on over 12 million Getty images without licence, and that the model can generate outputs bearing distorted versions of Getty watermarks — evidence of training on watermarked images. The UK case proceeded under English copyright law, which does not have the same fair use doctrine as the US. As of 2024, both cases remain ongoing but have established that AI companies cannot assume public accessibility of images implies licence to train on them.
Beyond IP, a rapidly developing area of legal exposure is product liability for AI outputs that cause harm. In 2023, a US attorney was sanctioned by a federal court after submitting a brief citing non-existent cases hallucinated by ChatGPT (Mata v. Avianca). While the attorney bore direct professional responsibility, the episode foreshadowed a class of harm where AI-generated output is relied upon for consequential purposes and proves false or dangerous.
The EU AI Liability Directive, proposed in 2022 and still in legislative process as of 2024, would create a presumption of causation where an AI system operated in a high-risk context and caused damage — shifting the burden to the developer or deployer to disprove the causal link. Combined with the EU Product Liability Directive update (also 2022), which explicitly extends product liability to software including AI, European businesses face a legal environment where demonstrable harm from an AI output creates a credible liability claim without requiring the plaintiff to unpack the technical causation chain.
In the US, existing product liability doctrine applies. The key questions courts will address are: Is AI a product or a service? (Different liability frameworks apply.) Who is the "manufacturer" in the supply chain — the model developer, the API provider, or the business deploying the model? The answers will shape the AI industry's liability architecture for decades.
Training data: Audit and document the provenance of any training data used in proprietary models. Ensure licences permit AI training use. Where EU DSM opt-outs apply, respect them.
Generated content: Flag AI-generated content internally. For commercially important content — campaigns, code, product designs — add sufficient human creative input and document it to support copyright claims.
Code generation: Implement policies requiring review of AI-generated code against known open-source licence obligations. Consider tools that flag potential licence conflicts in generated code.
Output verification: For any AI output used in professional, legal, medical, or financial contexts, implement mandatory human verification protocols. Document the verification process. Do not deploy AI-generated professional advice without human expert review.
Use this lab to analyse your organisation's exposure to IP infringement claims (training data, generated content, code), product liability from AI outputs that cause harm, and related contractual risks. Describe your AI use case — what you generate, what data you used, and how you deploy outputs — and receive structured legal risk analysis and practical mitigation steps.