What is the European AI Act?

How does the risk-based classification work?

The European AI Act divides artificial intelligence systems into four distinct risk categories. Rather than regulating the underlying technology in the abstract, the law evaluates the specific context, application, and potential harm of each deployment.

+-------------------------------------------------------------------------+
|                        EU AI ACT RISK TIERS                            |
+-------------------------------------------------------------------------+
|  1. Unacceptable Risk  -> Prohibited entirely (e.g., social scoring)    |
|  2. High Risk          -> Strict conformity, logging & human oversight  |
|  3. Specific Risk/GPAI -> Transparency, copyright & summary obligations |
|  4. Minimal Risk       -> Unregulated / Voluntary codes of conduct      |
+-------------------------------------------------------------------------+

1. Unacceptable Risk (Prohibited Practices)

Systems deemed a direct threat to safety, livelihoods, and fundamental rights are banned entirely. These bans entered into force in late 2024. Prohibited applications include:

• Cognitive behavioral manipulation that bypasses individual free will or distorts behavior to cause harm.
• Social scoring systems run by public authorities or private entities that evaluate individuals based on social behavior or personal characteristics.
• Real-time remote biometric identification in publicly accessible spaces by law enforcement, subject to narrow statutory exceptions.
• Biometric categorization systems that infer sensitive attributes such as political opinions, religious beliefs, or sexual orientation.
• Untargeted scraping of facial images from the internet or CCTV footage to build facial recognition databases.

2. High-Risk Systems

AI systems categorized as high-risk are permitted in the EU market only if they satisfy comprehensive conformity requirements. High-risk systems fall into two main groups: products already governed by EU safety legislation (such as medical devices, aviation software, and cars) and standalone AI systems used in sensitive areas like critical infrastructure, education admissions, employment screening, essential public benefits, law enforcement, and immigration management.

Deployers and providers of high-risk systems must establish risk management processes, conduct rigorous data governance, maintain automated event logging, provide clear technical documentation, and ensure adequate human oversight.

3. Specific Transparency Risk and General-Purpose AI (GPAI)

Systems that interact directly with individuals (such as conversational chatbots) must inform users that they are interacting with an artificial intelligence system unless this is obvious from the context. Systems generating synthetic audio, image, or video content (deepfakes) must clearly disclose that the material has been artificially generated or manipulated.

General-purpose AI models—including large language models—must maintain technical documentation, comply with EU copyright rules, and publish public summaries of the datasets used for training.

4. Minimal or No Risk

The vast majority of AI systems currently in everyday use, such as spam filters, recommendation engines, and video game mechanics, fall into the minimal risk category. These systems remain free from new legal obligations, though developers are encouraged to adhere to voluntary codes of practice.

What obligations apply to general-purpose AI models?

The European AI Act outlines a dedicated regime for general-purpose AI (GPAI) models. Because foundational models serve as downstream infrastructure for countless separate applications, regulators place specific disclosure duties directly on model providers.

All GPAI providers must establish policies to identify and respect copyright reservations expressed by rights holders under EU law. Providers must also publish a sufficiently detailed summary of the content used to train the model, using a template provided by the European AI Office.

Models classified as posing "systemic risk"—defined primarily by a cumulative computational capacity threshold exceeding $10^{25}$ floating-point operations (FLOPs) during training—face additional requirements. Providers of systemic GPAI models must:

1. Conduct continuous model evaluations, including standardized red-teaming tests to identify systemic vulnerabilities.
2. Assess and mitigate potential systemic risks at the European Union level, including cybersecurity hazards and algorithmic disruptions.
3. Track, document, and report serious incidents and corrective measures to the AI Office and relevant national competent authorities.
4. Ensure robust cybersecurity protection for the model weights and operational infrastructure.

These obligations apply regardless of whether the model is provided as a paid commercial API, integrated into consumer software, or released under certain open-source licenses that do not meet full exemption criteria.

How does the regulation treat data provenance and transparency?

Data integrity and verifiable provenance sit at the center of the AI Act. Regulators recognized that automated systems can amplify factual distortions, reinforce historical bias, and obscure intellectual property origins. Understanding what are the new global laws for AI systems helps contextualize why the EU prioritizes verifiable evidence trails.

Under the Act, high-risk systems must be trained and tested on data sets that satisfy strict quality criteria. Providers must document data collection methods, data cleaning workflows, labeling procedures, and assumptions regarding source material. This requirement directly impacts the mechanics of machine retrieval; information that lacks clear source attribution or shows structural evidence gaps introduces compliance vulnerabilities for downstream deployers.

+-------------------------------------------------------------------------+
|                    DATA GOVERNANCE & PROVENANCE FLOW                    |
+-------------------------------------------------------------------------+
|  Source Content  ->  Ingestion & Extraction  ->  Evidence Verification  |
|  - Clear Entity      - Machine-readable text    - Source provenance     |
|  - Documented Date   - Metadata preservation    - Human audit logs      |
+-------------------------------------------------------------------------+

For generative systems, providers must ensure that synthetic outputs are detectable through technical watermarking or machine-readable metadata. When automated platforms retrieve corporate or institutional publications to generate synthesized answers, clear evidence tracking ensures deployers can verify how do authorities track where AI finds its facts.

Maintaining structured data and clear factual attribution is critical. As automated discovery platforms adapt to these transparency standards, adopting principles from why AI governance needs source and citation provenance helps organizations remain dependable sources of record.

Who is held legally responsible when an AI system produces non-compliant outcomes?

The European AI Act distinguishes between two primary market actors: providers and deployers.

• Provider: The natural or legal person, public authority, agency, or other body that develops an AI system or general-purpose AI model and places it on the market or puts it into service under its own name or trademark.
• Deployer: Any natural or legal person, public authority, agency, or other body using an AI system under its authority, except where the AI system is used in the course of a personal non-professional activity.

Providers bear the primary burden for technical documentation, risk management systems, conformity assessments, and registration in the EU database. However, deployers take on legal liability if they use a high-risk system contrary to its instructions, make substantial modifications to an existing system, or repurpose a general-purpose model into a high-risk domain.

This division of accountability is examined in detail in our analysis on who is responsible when an AI tool shares false claims. When an organization integrates an automated engine to make autonomous legal, financial, or employment determinations, the deployer remains accountable for ensuring the output relies on dependable, verifiable evidence.

What are the penalties for non-compliance?

Enforcement under the European AI Act is structured around significant administrative fines, designed to deter non-compliance across global technology providers and institutional users.

Penalties are calculated based on fixed monetary sums or a percentage of the offending company's total worldwide annual turnover for the preceding financial year, whichever is higher:

• Violations of Prohibited AI Practices: Fines reach up to €35,000,000 or 7% of total worldwide annual turnover.
• Non-compliance with High-Risk and GPAI Obligations: Fines reach up to €15,000,000 or 3% of total worldwide annual turnover.
• Supplying Incorrect or Misleading Information to Authorities: Fines reach up to €7,500,000 or 1.5% of total worldwide annual turnover.

For small and medium-sized enterprises (SMEs) and startups, the law stipulates that fines will be capped at the lower of the two amounts, preventing administrative penalties from disproportionately threatening early-stage business viability.

Enforcement is overseen at the EU level by the European AI Office (established within the European Commission) for GPAI models, and by designated national supervisory authorities in each EU member state for domestic deployers and systems.

How does the Act affect organizations outside the European Union?

The AI Act features strong extraterritorial reach, similar to the General Data Protection Regulation (GDPR). An organization does not need a physical presence within the EU to fall within its legal scope.

The rules apply to providers outside the EU if they place AI systems on the EU market or put general-purpose models into service for EU users. Crucially, the law also applies to providers and deployers located in third countries (such as the United States or the United Kingdom) if the output produced by their AI system is used within the European Union.

For example, if an American enterprise uses an algorithmic assessment tool based in California to process candidate applications for an office in Frankfurt, that deployment is subject to EU high-risk obligations. International enterprises must review their technology stacks to ensure their workflows align with broader AI regulation.

How can organizations prepare their information architecture?

Preparing for the enforcement of the AI Act requires operational clarity across an organization's digital assets, documentation, and citation structures. Regulators increasingly expect verifiable source material rather than unsubstantiated claims.

Organizations can take practical steps to align with emerging transparency standards:

1. Inventory AI Usage: Catalog all internal AI tools, API integrations, and public-facing retrieval interfaces. Categorize each use case against the Act's four risk tiers.
2. Audit Data Integrity: Review the factual accuracy, currency, and provenance of public-facing corporate information. Removing conflicting statements and broken claims ensures automated retrieval engines extract dependable facts.
3. Implement Structured Provenance: Ensure that primary research, case studies, and corporate filings contain structured metadata, explicit author credentials, and clear publication dates.
4. Strengthen Citation Integrity™: Review how internal and external assertions are supported. Adopting principles from what is citation integrity helps systems verify claims across automated discovery pipelines.
5. Establish Human Oversight: Create formal verification steps before AI-generated summaries or autonomous decisions are deployed in high-stakes settings.

For a broader operational roadmap, see our detailed guide on how can businesses prepare for future AI regulations, or review the comprehensive evaluation criteria in the CiteAbility™ framework.