Bird's-AI view

“‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge.”

---

“‘deployer’ means a 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.”

---

“‘importer’ means a natural or legal person located or established in the Union that places on the market an AI system that bears the name or trademark of a natural or legal person established in a third country.”

---

“‘distributor’ means a natural or legal person in the supply chain, other than the provider or the importer, that makes an AI system available on the Union market.”

---

The definition of limited risk AI systems, as well as their obligations, are entirely encapsulated in Article 50. It centers around the idea that users need to be aware when interacting directly with AI systems or the content generated by these systems. Included are AI systems:

• Intended to interact directly with persons
• Synthetically manipulating or generating audio, image, video, or text content
• Generating or manipulating image, audio, or video content constituting deep fakes
• Generating or manipulating text for public information purposes
• Used for emotion recognition & biometric categorization purposes

High risk AI systems carry the brunt of legislation, mostly for providers. This includes the topics: documentation & record keeping, data governance, risk management system, post-market monitoring, human oversight, metrics and transparency - including the legislation set out under limited risk AI. As per chance, the most convoluted set of measures falls onto the most convoluted class definition. A system is classified as high risk via membership in one of two separate lists; it either falls under a set of applications or it is part of a set of industries and products. The set of applications can be found in Annex III and includes the following:

• Biometrics
  Exceptional applications of remote biometric identification, biometric categorisation and emotion recognition systems that are not prohibited by Article V.
• Critical infrastructure
  Essential services for the maintenance of vital societal functions or economic activities. These are defined in the Critical Entities Resilience Directive (CER).
• Education
  Admissions, assessments, evaluating & steering learning outcomes or the monitoring & detection of prohibited behavoir during tests.
• Employment
  For purposes of recruitment, promotion or termination, including targeted advertisement; deciding on work-related contractual relationshaps or the allocation of tasks; used to monitor & evaluate the performance & behavoir of a worker.
• Essential services
  Evaluating eligibility for essential public assistance benefits and services; risk assessment & pricing in case of life & health insurance; evaluation of creditworthiness or establishment of credit score; evaluation & classification of emergency calls as well as establishment of emergency priorities.
• Law enforcement
  Used in polygraphs or similar tools; assessment of risk to become a victim of criminal offences; evaluation of evidence reliability & prosecution of criminal offences; profiling during detection, investigation & prosecution of criminal offences; risk assessment of (re-) offending based on profiling and assessment of behavioral & criminal traits.
• Migration control
  Used in polygraphs or similar tools; assessment of security risks; also if used for examination of asylum, visa & residence permits applications or detection, recognition & identification of individuals.
• Democracy and judicature
  Research and interpretation of facts & law; application of law; influence of voting behavior & outcome, excluding those not directly interacting with voters.

The AI system may also be used within several sectors and products already affected by Union Harmonisation Legislation, either as a safety component or as the product itself. These are listed in the less accessible Annex I. Thankfully, Deloitte has presented this information in a more readable format, which we’ll use here. This is the final piece of legislation to come into effect in October of 2026

• Civil aviation
  All airports or parts of airports that are not exclusively used for military purposes as well as all operators, including air carriers, providing services at airports or parts of airports that are not exclusively used for military purpose.
• Agricultural and forestry vehicles
  Tractors as well as trailers and interchangeable towed equipment.
• Two-, three-wheel vehicles and quadricycles
  Get ready for a unicycle boom.
• Marine equipment
  On board of EU ships
• Personal protection
  Personal protective equipment designed and manufactured to be worn or held by a person for protection against one or more risks to that person’s health or safety, as well as some interchangeable components and connection systems for the equipment.
• Appliances burning gaseous fuels
  Appliances burning gaseous fuels used for, amongst others, cooking, refrigeration, air-conditioning, space heating, hot water production, lighting or washing; also, all fittings that are saftey devices or controlling devices incorporated into an applicance.
• Rail system
  Rail system including vehicles, infrastructure, energy and signaling system.
• Motor vehicles and trailers; systems, components and separate technical units intended for such vehicles
  Motor vehicles and their trailers; including, amongst others, autonomous driving.
• Civil aviation, european air space, aerodomes
  The design and production of products, parts and equipment to control aircraft remotely by a natural or legal person, as well as the design, production, maintenance and operation of aircrafts.
• Medical devices
  Medical devices for human use and accessories for such devices as well as clinical investigations concerning such medical devices and accessories.
• In vitro diagnostic medical devices
  In vitro diagnostic medical devices for human use and accessories for such devices.
• Machinery
  Machinery, interchangeable equipments and lifting accessories (e.g., robots).
• Toys
  Products designed or intended for use in play by children under 14 years of age (e.g., connected toys and IoT devices).
• Recreational craft and personal watercraft
  Recreational craft as well as propulsion engines installed on watercraft.
• Lifts
  Lifts permanently serving buildings and constructions for mainly the transport of persons with or without goods.
• Equipment and protective systems intended for use in potentially explosive atmospheres
  Equipment and protective systems intended for use in potentially explosive atmo- spheres as well as components incorporated within the equipment or protective system.
• Pressure equipment
  The design, manufacture and conformity assessment of pressure equipment and assemblies.
• Radio equipment
  Any kind of radio equipment that is anything connected via radio waves (e.g., WiFi, Bluetooth, 5G in laptops, phones, IoT devices)
• Cableway installations
  New cableway installations designed to transport persons, to modifications of cableway installations requiring a new authorization, and to subsystems and safety components for cableway installations.

Prohibited practices are detailed in Article 5. The formulation of this text can get rather precise regarding specifics and exceptions. We will give a summary largely based off of the same Deloitte article. When in doubt, please consult the official AI Act document. The following practices shall be prohibited starting the 2nd of february, 2025.

• AI-enabled manipulative techniques
  Persuasion to engage in unwanted behaviours, nudging for subversion and impairment of autonomy, decision-making and free choices causing or reasonably likely to cause harm.
  Excluding: common and legitimate commercial practices, e.g., advertising.
• Biometric categorisation
  Use of individual biometric data as face or fingerprint, to deduce or infer political opinion, trade union membership, religious or philosophical beliefs, race, sex life or sexual orientation.
  Excluding: labelling, filtering or categorisation of lawfully acquired biometric datasets. As well as law enforcement purposes.
• Real-time remote biometric indentification
  In publicly accessible spaces for the purpose of law enforcement. For targeted victim search or terrorist threat mitigation. Persons may be targeted when suspected of any of the crimes mentioned in Annex II. May only be used by secret services or with prior authorisation granted by a judicial authority and only to confirm the identity of persons.
• Social scoring
  Valuation or classification of persons or groups based on multiple data points related to social behaviour and leading to negative treatment.
  Excluding: lawful evaluation practices of persons done for a specific purpose in compliance with national and Union law.
• Risk assessments
  Assessing persons traits and characteristics to predict the risk of committing a criminal offence. Unless in support of a human assessment of the involvement of a person objectively and verifiably linked to a crime.
• Facial recognition
  Creation or expansion of facial recognition databases through the untargeted scraping of facial images from the internet or CCTV footage.
• Identifying emotions
  Identifying or inferring emotions or intentions on the basis of their biometric data elated to places of work and education institutions.
  Excluding: medical reasons such as therapy and safety reasons such as driver awareness assessment.

Your AI system is consideredgeneral purpose if it is capable of performing a broad range of tasks. Due to their broad scope, systems in this category may pose unforeseeable risks. Providers of these systems are subjected to an extra set of testing and transparency measures. These measures will supplement any measures already mandated by the AI system’s risk classification

“An AI model, including when trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks regardless of the way the model is placed on the market and that can be integrated into a variety of downstream systems or applications. This does not cover AI models that are used before release on the market for research, development, and prototyping activities.”

---

Exceptions are made for open source models, as well as models that are used in a research or development setting.

“The obligations set out in paragraph 1, points (a) and (b), shall not apply to providers of AI models that are released under a free and open-source licence that allows for the access, usage, modification, and distribution of the model, and whose parameters, including the weights, the information on the model architecture, and the information on model usage, are made publicly available. This exception shall not apply to general-purpose AI models with systemic risks.”

---

There’s a little bonus for any collector among you who wants to tick all the boxes, albeit a little expensive. This last box goes out to models commonly referred to as foundation models. Your general purpose AI System may get this extra ‘systemic risk’ mark when it meets any of the following conditions:

The criteria set out in Annex XIII mainly indicate training costs. The following features are mentioned: number of parameters, quality & size of the dataset, floating point operations and other training time indicators, number of input and output modalities, existing benchmarks, market impact and user count. New models may be classified as systemic risk when they compare with currently registered systemic risk models based on these criteria.

• Direct interaction
  AI systems shall be designed and developed in such a way to ensure persons concerned are informed that they are directly interacting with an AI system.
• Marked content
  AI generated audio, image, video or text shall be marked in a machine-readable format and detectable as artificially generated or manipulated. Excluded are AI systems that perform an assistive function for standard editing or do not substantially alter the input data or the semantics thereof.

• Emotion recognition & biometric categorization systems
  Insofar as these systems are legal, persons exposed to them need to be properly informed about how they are affected these systems.
• Full disclosure
  Is required for artificially generated or manipulated image, audio or video content constituting deep fakes. As well as generated or altered text with the purpose of informing public on matters of public interest, except were content has undergone human review and a natural or legal person holds editorial responsibility. Where content forms part of an evidently artistic, creative, satirical, fictional or analogous work or programme it’s sufficient to disclose the existence of such content in ‘an appropriate manner that does not hamper the display or enjoyment of the work’.

Measures for providers of high risk AI systems are by far the most expansive, We've ordered them into 3 sections; Information Management, Quality Management, and Human Oversight. Some of the AI Act measures may overlap with other Union legislation. The Union Harmonisation Legislation and the Union Financial Services Law are often mentioned throughout the AI Act. You are encouraged to integrate measures and procedures from different legislative pieces into a single comprehensive protocol.

Information Management

Keep technical documentation and logs. Records may be required for the proper implementation of the risk management & post-market management systems discussed below. Or, depending on the type of information, it is to made available downstream to deployers , who in turn will share it with distributors and importers. Alternatively, the information may need to be made available to a Union authorized representative upon request.

• General description of AI system
  This includes its intended purpose, requirements, a product overview and instructions for use.
• Detailed description of AI system
  How the AI systems was developed, what key design choices were made, what the system is designed to optimize, computational resources used, datasheets detailing which data was used in what way, validation & testing procedures and cybersecurity measures.
• Detailed information about the Quality Management System
  Detailed descriptions of the model’s inputs, the risk management system, the post-market monitoring system, data management protocols, human oversight measures, as well as assessments of potential risk and test & validation reports; dated and signed.
• Copy of declaration of conformity
  This document is part of the conformity assessment procedure which is the be executed prior to the product being placed on the market.
• Automated recording of events
  Keep records of events during the lifetime of the AI system, these records are to be used in risk management and post-market monitoring and have to be shared with authorized representatives upon reasonable request.
• Metrics
  Performance in accuracy, benchmarks, robustness & cybersecurity

Quality Management

Providers are required to take these measures in order to anticipate and contain potential risks involved with their AI system. In essence, these measures force providers to adopt a methodical approach to risk mitigation. This includes documenting common sense reasoning & analyses, setting up protocols, stringent testing & evaluation procedures and scrutinizing data collected through monitoring practices.

• Data governance
  Most parties already comply to many of these standard good practices. They include, among others: detect, prevent and mitigate data biases and data gaps and ensure datasets are sufficiently representative.
• Risk management system
  A continuous iterative process planned and run throughout the entire lifecycle of a high-risk AI system. This helps companies to identify, analyse and evaluate foreseeable risks, using information generated through the post-market monitoring system, as well as the performance, robustness & cybersecurity metrics. It shall include appropriate measures to mitigate or nullify foreseeable risks.
• Post-market monitoring system
  Actively and systematically collect, document and analyse relevant data on the performance of high-risk AI systems throughout their lifetime, this allows providers to evaluate the continuous compliance of AI systems. Include an analysis of the interaction with other AI systems.
• Performance, robustness & cybersecurity
  Ensure appropriate & consistent performance in accuracy, robustness & cybersecurity. Mitigate feedback loops. Ensure technical redundancy. Protect against data- and model poisoning, adversarial examples or model evasion, confidentiality attacks or model flaws.

Human Oversight

The measures ensure AI systems are designed and developed so that they can be effectively overseen by employees of the deployer - for instance: an ‘AI Compliance Officer’, with aim to prevent or minimize foreseeable risks to health, safety or fundamental rights. To comply to measures concerning Human Oversight providers shall appropriately combine measures built into the AI system, as well as measures to be implemented by the deployer, identified a priori by the provider.

• Enable deployers to properly understand system capacities
  And monitor its operations: detect and address anomalies, dysfunctions and unexpected performances.
• Enable deployers to correctly interpret the AI system’s output
  Use explainable AI tools to provide explanations annotating model outputs. This can be done through Xaiva, or feature importance techniques, which we wrote in more detail about here.
• Mitigate automation bias
  Increase awareness and mitigate risks of automation bias in your AI system design choices.
• Enable disregard, override or reversal of AI system output
  Any AI system’s output shall be fully adaptable and reversable by the AI compliance officer, before being actuated.
• Stop button
  Deployers need an accessible method ‘that allows the system to come to a halt in a safe state’.

To use AI models responsibly, Human involvement is essential to check model decisions and verify that they behave as they should. The AI Act mandates deployers to adopt this human-in-the-loop approach. In this article we refer to these future governance roles as ‘AI Compliance Officers’. Providers of high-risk AI systems are responsible for facilitating human oversight. To learn what you, as a deployer, have to expect from your provider, read the ‘Human Oversight’ section under high-risk obligations for providers. These are the responsibilities set out for deployers.

• Implement human oversight measures
  Identified, facilitated and instructed by the provider. Assign AI compliance officers with necessary competence, training, authority and support.
• Control over inputs
  If able to exercise control over input data, make sure that data is relevant and sufficiently representatitve. Keep automatically generated logs for at least 6 months.
• Minimize risks to health, safety or fundamental rights
  in accordance with its intended purpose or under conditions of reasonably foreseeable misuse.
• Monitor operation
  Enable AI compliance officers to properly understand the relevant capacities and limitations in view of detecting and addressing anomalies, dysfunctions and unexpected performance.
• Inform parties
  When suspecting risks to health, safety or fundamental rights of persons, inform the provider followed by the distributor and relevant market surveillance authority and suspend the use of the system.
• Inform stakeholders
  That they are subject to the use of a high-risk AI system that makes decisions or assists in making decisions
• Amend outputs
  Decide not to use the high-risk AI system or to otherwise disregard, override or reverse the output of the high-risk AI system.
• Stop!
  Press the ‘stop’ button, allowing the system to come to a halt in safe state.
• Biometric identification
  When using biometrics to confirm the claimant’s identity, at least 2 persons need to verify and confirm system output. This measure does not apply for law enforcement, migration, border control and asylum purposes.

Obligations for importers and distributors largely overlap. In essence, they are responsible for verifying conformity of the entire supply chain. They are able to do this due to transparency requirements set out by the AI Act. Importers and distributors are able to verify all the documentation provided to them by the provider, deployer or importer they are dealing with. Some of these measures have to be performed before putting a product to market, others require contiuous attention after the product is marketed. I’ll start with the measures that apply to both roles before naming the exceptions.

• Ex-ante
  Instructions for use and other documentation are in proper order. The product bears a CE marking and there’s a declaration of conformity. Verify product name & trademark.
• Ex-post
  Do not make available, correct, withdraw or recall the product when suspecting noncompliance. Immediately inform provider or importer and competent authorities when system is presenting risk. Ensure storage or transport does not effect compliance.

Distributor

• Ex-ante
  Ensure quality management system is in place and compliant.

Importer

• Ex-ante
  Appoint authorized representative. Ensure provider completed conformity assessment procedure. Be watchful for falsification. State registered trade name or trademark and contact address on package or accompanying documentation.
• Ex-post
  Ensure all necessary information to demonstrate conformity can be made available to competent authorities in a language which can easily be understood by them. Keep documentation for 10 years after product has been placed on market or put into service.