A team inside the European Commission's research directorate has set out how applicants for EU research money should measure the risk their AI systems pose to fundamental rights, rating each right separately on a four-level scale and refusing to reduce the result to a number. The note, titled "Fundamental Rights Impact Assessment (FRIA) and AI - Ethics advice for researchers and ethics experts," is dated May 2026 and was prepared by expert Alessandro Mantelero for the Research Ethics & Integrity Team of DG Research & Innovation, with staff member Mihalis Kritikos coordinating. It states that it "does not constitute official EU guidance." Its reviewer checklist nonetheless tells ethics panels to ask for an assessment where a proposal lacks one, and its annex lists AI credit scoring and the pricing of life and health insurance among the uses that can affect consumer protection rights.

In Short

The European Commission's research ethics team wrote a guide for scientists who want EU money for projects that use AI, explaining how to check whether the AI could hurt people's basic rights, such as privacy, equal treatment or health. Each right gets its own score of low, medium, high or very high, based on how likely harm is, how many people it could reach, how bad it would be and how hard it would be to fix. The guide is advice rather than law, but it shows how Brussels thinks about the same kind of check that banks and insurers using AI for credit and pricing decisions will face under the AI Act.

What the note is, and what it is not

The document comes from RTD.02, the Science for Policy, Advice and Ethics unit of the Directorate-General for Research and Innovation, in Brussels. According to the Commission, its purpose is "to provide a methodology for Fundamental Rights Impact Assessment (FRIA) in the context of the Horizon Europe ethics review procedure for projects involving the use of AI." The disclaimer describes it as prepared by a panel of experts "to support applicants and beneficiaries of Horizon Europe and other EU-funded programmes."

Two caveats sit at the front. The first is that the advice "cannot anticipate the outcome of the ethics appraisal of any application for EU funding," and the note accepts that projects on very similar topics may be assessed differently. The second is a legal notice: the note "does not constitute official EU guidance" and "cannot create any new obligations on the European Commission, its Executive Agencies, or other EU Funding Bodies." Nobody acting on the Commission's behalf can be held responsible for how it is used.

The structure is simple. Part I frames the problem, tracing impact assessment from the United Nations Guiding Principles on Business and Human Rights to the AI Act. Part II is the operational section, containing the methodology, a template for researchers and a checklist for reviewers. Annex I gives examples of AI affecting rights drawn from 35 articles of the EU Charter of Fundamental Rights, and Annex II walks through a case study. According to the reader's guide, "the FRIA methodology presented in this document is aligned with the AI Act and its obligations regarding the assessment of the impact of AI on fundamental rights."

One detail dates the research behind it. A footnote states: "All the websites cited in this Document were accessed between January and May 2024." The only version of the AI Act the note cites is Regulation (EU) 2024/1689 of June 13, 2024. It does not mention the Digital Omnibus on AI or the revised application dates that followed in 2026.

Why a data protection assessment is not enough

Researchers already know one impact assessment well. Article 35 of the GDPR requires a Data Protection Impact Assessment (DPIA) where processing is likely to result in a high risk to people's rights and freedoms, and the note points out that a DPIA is already supposed to look beyond data protection. In practice, it argues, that is not what happens.

"Assessing impact on various fundamental rights through the lens of the DPIA leads to the use of data protection categories to justify the final decision, which largely obscures the rationale of the assessment in relation to these fundamental rights," the note states. It adds that, from analysis of DPIA practice, "attention paid to fundamental rights other than data protection is limited in scope and not fully developed."

The note also draws a line between legal and ethical reasoning. During what it calls the first phase of the debate on AI regulation, it says, there was a risk of an "ethicsisation" of AI, "turning fundamental rights into ethical values." The opposite extreme, where legal provisions become the only test of whether an AI use is acceptable, is rejected too. Ethics review, which the note says the AI Act treats only marginally, remains a complement.

Then there is the question of timing. Human Rights Impact Assessment (HRIA), the corporate due diligence tool associated with the 2011 UN principles, is described as "primarily an ex post facto response to critical situations," producing a list of options that a company may or may not adopt. Under the AI Act, as under the GDPR and the Digital Services Act, the assessment comes first. "Under the AI Act, FRIA is a mandatory assessment," the note says, and its results must be used to prevent risks.

The model the note builds on is what it calls value-oriented rather than a pure risk-benefit trade-off. In its words, "the protection of fundamental rights cannot be restricted for reasons of mere economic interests or more efficient solutions." The example given is a fully automated AI-based medical triage procedure with no human supervision.

Three functions, and who carries them out

The note splits the work into three functions: awareness raising, risk analysis and risk management. Awareness raising belongs to the researchers, who must understand whether AI is appropriate at all and which areas it may touch. The other two require fundamental rights expertise, which, if the research team lacks it, must come from experts. "Researchers may be less familiar with the legal framework of fundamental rights than with ethical principles," the Commission's note observes.

Two warnings follow. "FRIA itself cannot be reduced to a mere questionnaire-based exercise," the note states, because its contextual nature conflicts with a one-size-fits-all form. Templates can help, but "the assessment of the impact on fundamental rights cannot be fully automated."

General-purpose models get particular attention. Systems built on them raise specific concerns, according to the note, "due to their high versatility of use in many contexts" and because they are embedded in systems with different characteristics. A footnote illustrates the point with public administration chatbots: giving general information on procedures is, in risk terms, quite different from advising on a real divorce case.

Data choices are flagged as well. The note describes a hypothetical tool predicting student learning performance that relies only on data schools hold. Nationality and family size might be used as risk factors, it says, but language matters more than nationality, and a large family can help a child as easily as it can hinder.

The scoring model

How does the method turn an abstract concept such as harm to rights into something that can be compared? The note starts from the premise, "in line with general risk theory, that there is no zero risk," and places each impact on a scale from a minimum to a maximum. Because social phenomena are hard to measure with exact figures, it uses ordinal values: low, medium, high and very high. The note gives its own example: "we cannot quantify in the same way the impact of a credit scoring system on some groups of people in terms of discrimination."

Likelihood: probability and exposure

Likelihood combines two variables. Probability runs from low ("The risk of prejudice is improbable or highly improbable") to very high ("The risk is highly likely to occur"). Exposure runs from "Few or very few of the identified population of rights-holders are potentially affected" to "Almost the entire identified population is potentially affected." The note stresses that for exposure, "the focus is on the group of potentially affected people and not on the whole population." Duration over time, used in some other models, is left out because it is hard to estimate for AI projects.

Severity: gravity and effort

Severity also combines two variables. Gravity describes the prejudice to the exercise of rights, taking into account group-specific impact, vulnerability and dependency; it ranges from "only minor prejudices" to "serious or even irreversible prejudices." Effort measures how hard it is to overcome the harm and reverse its effects. At the low end, the examples are time spent amending information and annoyances. Medium covers extra costs, fear and stress; high covers economic loss, property damage and worsening health. At the very high end, the prejudice "may not be overcome," with long-term psychological or physical ailments and death as examples.

Each pair is combined in a four-by-four matrix, and likelihood and severity are then combined in a third matrix to produce an overall impact for each right. A radial graph plots the results so that the most affected rights can be prioritised.

Why four levels, and why no numbers

The note surveys common matrix formats, from 3x3 and 4x4 to 5x5, 5x4 and 6x4. A 4x4 grid, it says in a footnote, "may be the most appropriate" because it "reduces the risk of average positioning, gives more attention to the high and very high levels in a way that is consistent with the focus on high risk in the current regulatory approach to AI, and does not excessively fragment the lower part of the scale."

On numbers, the guidance is firm. Since the components of impact on rights "cannot be quantified in a precise and granular manner," the note states that "variables, indicators and indices should not use numerical values." It warns against purely mathematical approaches to scaling, asks for transparent scaling criteria, and calls for a record of "all choices made in the matrix design" for accountability.

No averaging across rights

A composite index combining all impacts into one score is possible in social research, the note acknowledges. It rejects that approach for fundamental rights, where each right must be considered on its own. "The fact that one right is less affected than another cannot lead to any form of compensation," it states. Splitting one right into several components, which would count the same harm twice, is also ruled out in the note's list of best practices.

Exclusion factors and the balancing test

Competing interests enter only after each right has been scored. The note calls them exclusion factors: "all those elements that justify the expected impact." The first group is legal provisions. Its example is a national law that authorises real-time remote biometric identification in publicly accessible spaces to search for a suspect, citing Article 5(1)(h) of the AI Act. If a research project develops that technology within the scope of such a law, the impact on rights such as freedom of movement and the presumption of innocence is treated as excluded, "because the prevalence of competing interests has been established by law." Features that the law makes mandatory fall into the same group. A second group covers case-by-case balancing where no law has settled the question.

Even then, mitigation is required for any risk not justified by the prevailing interest. And measures need not touch the model at all. "Acting on the context of use can exclude risks without changing the design of the AI system," the note states, citing industrial robots whose collision risk cannot be designed out but can be avoided with mandatory paths, no-go zones for people and physical barriers.

The researcher template

The template opens with a gate. Question I.1 asks whether the proposal uses "any type of AI system or general-purpose AI models with systemic risk" that may directly or indirectly have an impact on people and society "other than in terms of advancing scientific knowledge." If the answer is no, a brief justification closes the assessment.

The planning table that follows asks researchers to describe the purposes and characteristics of the AI, and whether the system will be developed and put into service "for the sole purpose of scientific research and development." That question matters because the AI Act does not apply to AI systems and models built and put into service only for scientific research and development. Researchers must also say whether personal data will be used for training or fine-tuning, and in which countries the system will be used.

Several questions deal with third-party models. Has the AI provider already carried out an impact assessment, and have its results been shared? Has the provider given the researchers instructions on protecting fundamental rights? Will subcontractors be involved, and will they take part in the assessment? Each question is tagged by who should answer it: the researchers alone, the researchers with a fundamental rights expert, or with input from affected people and stakeholders.

Researchers then identify rights holders and duty bearers, list the rights that may be affected both by the research and by the later use of its results, and justify why AI is needed. If AI is not necessary and alternatives exist, the assessment stops. A footnote exempts programmes that require AI as a condition of participation. Existing assessments, such as a DPIA, are to be summarised and attached.

The risk section adds a table of dos and don'ts. Among the don'ts: cumulative assessment of impacts on different rights, double counting, unjustified overweighting of variables, and "ranges based on absolute numbers to estimate the magnitude of the impact." The note also explains, in a short FAQ, why it offers no taxonomy of AI components, no list of rights per application and no catalogue of mitigation measures. "Taxonomies are useful, but also pose a risk in the case of technologies and applications that can have a wide variety of different implementations," it says. On matrices, it argues that because AI providers and deployers must be accountable, "risk assessment cannot be based on statements that are not supported by evidence."

The FAQ is also explicit about scope. Asked whether the template can be used for the overall assessment of an AI system, it answers: "No, the template only covers the fundamental rights impact assessment. It does not address other relevant issues such as technology robustness, system security and conformity assessment."

Participation, and its limits

Involving rights holders and stakeholders is described as "often crucial," both to find impacts that researchers miss and to build trust. The note asks that participants receive accessible, neutral information about the technology and, where possible, feedback on how they shaped the design. Disclosure may clash with intellectual property, so limited disclosure or confidentiality agreements are suggested. It also names the risks of "participation washing," biased group composition and arbitrary limits on the voluntary nature of taking part.

What reviewers are told to check

The reviewer checklist has six questions that mirror the template. Reviewers check whether the AI may affect people beyond advancing knowledge, who the rights holders are, which rights are at stake, whether AI is essential, whether a proper assessment exists, and whether partners and subcontractors should be involved.

The fifth question sets out two cases. In Case A, "if the proposal does not include a FRIA, the ethics reviewers will ask the researchers to provide one." In Case B, reviewers check methodological consistency: no cumulative scoring across rights, separate likelihood and severity ratings, specific measures per right with expected outcomes, and periodic checks through the project lifecycle where needed. "In the event of shortcomings in the performance of FRIA, ethics reviewers may ask for additional information and revisions," the note states. Where researchers have not identified rights holders, reviewers will do it themselves. Where necessary partners are missing, reviewers are to identify them and ask researchers to involve them.

Annex I: where the note sees AI touching rights

The first annex table pairs Charter rights with examples of AI uses. The note calls it "not exhaustive" and "an initial benchmark." Several entries concern consumer-facing technology.

Under consumer protection (Article 38), it lists "AI system used to evaluate the creditworthiness of natural persons or to establish their credit score," AI "used for risk assessment and pricing in relation to natural persons in the case of life and health insurance," and AI products "at risk of causing harm, such as harm caused by safety or data processing issues (e.g., hallucinations)."

Under the rights of the child (Article 24), it lists AI used in social media or gaming that can affect children's wellbeing "by creating addictions and dependencies," and "AI-based recommender systems and conversational agents designed for a general audience with a level of explainability and transparency not appropriate for child users."

Under privacy and data protection (Articles 7 and 8), it cites emotion detection, biometrics, surveillance and the use of "human-like synthesised voices to deceptively gather personal information over the phone."

Under the right to good administration (Article 41), the note places this example: "AI technologies can be used to dynamically orient and personalise the choice environment, using nudging and manipulation techniques to direct individuals' behaviour towards specific economic, social or cultural orientations." The Charter article governs how EU institutions, bodies, offices and agencies handle people's affairs, so the example appears under public administration rather than under consumer protection.

Other entries cover content moderation and AI-generated disinformation (freedom of expression), generative models "created using copyrighted material without the permission of the authors" (right to property), AI in recruitment, task allocation, performance monitoring and dismissal (occupation and working conditions), electoral disinformation (the right to vote) and the energy consumption of AI (environmental protection).

A second table works the other way, from technology to rights. It takes the domain and sub-domain descriptions from the AI Risk Repository maintained at MIT, compiled by Peter Slattery and colleagues in 2024, and adds a column, written by the author, mapping each risk to affected rights. Seven domains and 22 sub-domains are covered, from discrimination and toxicity to AI system safety. Overreliance and unsafe use maps to 14 rights. The repository's description of "Pollution of information ecosystem" refers to highly personalised AI-generated misinformation creating "filter bubbles"; the note maps it to freedom of thought, freedom of expression, freedom of the arts and sciences and the right to vote.

The worked example: a chemotherapy predictor

Annex II applies the method to a case that, according to the note, is "based on a real case analysed by the HRIA working group established by the Catalan Data Protection Authority in 2023" and led by Mantelero. The project would build an AI system using medical images to predict how patients respond to a chemotherapy treatment for one of the world's most common cancers, one associated with high levels of overtreatment and significant side effects. Training data come from 5,000 patients in ten European countries. Validation would follow in eight healthcare centres around the world.

The system is not a pure research tool. Asked whether it would be developed solely for scientific research and development, the case answers: "No, the system is being developed to be put on the market for commercialisation after the research and development phase." The training data include medical images, age and country, cancer stage and molecular subtype, pre-treatment information, treatment regimen and post-treatment pathology reports. These are special categories of personal data, as the case's own risk table acknowledges. Rights holders are people aged 18 to 85 with the cancer, and "as all the people involved are affected by cancer, they must be considered vulnerable."

Three rights are assessed. For the right to integrity of the person, faulty outputs could lead to ineffective or harmful treatment. Probability is rated medium for European patients and high for non-European patients, since the training data are European; exposure is very high. The overall impact is high for European patients where follow-up monitoring can catch errors, and very high in the remaining scenarios. For protection of personal data, probability is low because, as an EU-funded project, it is subject to ethics review and a data protection assessment; the overall rating is medium. For non-discrimination, the case states that "ethnicity may cause some differences in medical imaging, that may affect diagnostic accuracy," and every input is rated high or very high. The overall impact is very high. Correcting it, the case says, "it would be necessary to adapt or even retrain the algorithm with data that avoids discrimination."

The proposed measures are concrete. Healthcare professionals would be informed of the tool's limitations and of the error rate of the imaging equipment, and pathologies with and without rapid evolution would be distinguished to define follow-up. Data subjects would receive information on how training data were obtained. And the project would "add training data covering under-represented patient types" and "exclude/limit the use of the systems in areas with low representation in the training data." After those measures, the residual impact on non-discrimination falls from very high to medium, data protection stays at medium, and integrity of the person drops to medium or high depending on the scenario.

Gaps and inconsistencies in the text

The note contains several cross-reference errors. It defines the "AIA FRIA" as the assessment provided for in Article 27 of the AI Act and points readers to "section I.5 below," but Part I ends at section I.4. It refers to the Guidelines for Researchers as section II.3, while they appear in section II.4; II.3 covers participation. Exclusion factors are cross-referenced to "Part I, section II.2.2," a section that does not exist.

The final impact matrix relies on colour shading, described as "colours from lightest to darkest," with no written rule for how combinations of likelihood and severity translate into an overall level. The case study's results, such as medium probability and very high exposure producing high likelihood, therefore cannot be checked against stated rules in the text.

Two further tensions are visible. One of Part I's takeaways says that "fundamental rights cannot be overridden by any kind of competing interest," yet Part II allows a prevailing competing interest to make "acceptable a situation of high impact" after the balancing test. And in the case study, the answer on non-EU legal frameworks states that "potential prejudice due to limited human rights protection in some countries is not foreseen," while the same project receives a very high non-discrimination rating for use in non-EU facilities. The two answers address different issues, legal protection on the one hand and training data on the other, but they sit side by side without comment. The case also records that "no participatory risk assessment is planned," despite the note's emphasis on involving rights holders, and despite classing every affected patient as vulnerable.

Where Article 27 of the AI Act fits

The note's research context differs from the legal duty it references. Article 27 of the AI Act, in the text published in 2024, requires certain deployers of high-risk AI systems listed in Annex III to assess the impact on fundamental rights before first use. Those deployers are bodies governed by public law, private entities providing public services, and deployers of systems that evaluate creditworthiness or establish credit scores, or that price and assess risk for life and health insurance. Those last two categories closely track the wording of the consumer protection examples in the note's annex. Article 27 also states that where a DPIA already covers some of the obligations, the rights assessment complements it, and it tasks the AI Office with developing a questionnaire template. The note does not say whether its method is meant to inform that template.

The timetable for those obligations has moved. In May 2026, Council and Parliament negotiators agreed to push AI Act high-risk deadlines to December 2, 2027 for Annex III systems and August 2, 2028 for Annex I products, about a week after an earlier round of talks had collapsed. The Digital Omnibus on AI, Regulation (EU) 2026/1744, was adopted on July 8, 2026 and published in the Official Journal on July 24, and the amended Article 113 sets December 2, 2027 for the Annex III chapters of the AI Act, which include the deployer obligations. The note, whose sources were accessed in 2024, reflects none of this.

Why this matters for the marketing community

For most advertisers, the research ethics review of Horizon Europe is distant. The methodology is not. It sets out in detail, from inside the Commission's own research directorate, how a fundamental rights assessment of AI could be structured, and the uses it singles out overlap with commercial practice: credit scoring, insurance pricing, recommender systems, conversational agents, synthetic voices, and generative models trained on copyrighted work.

The contrast with other jurisdictions is sharp. Colorado's SB 26-189, which the state House passed 57 to 6 in May 2026, excludes advertising, marketing, differentiated product recommendations, search and content moderation from the consequential decisions it regulates. The Commission note treats recommender systems aimed at general audiences and content moderation as potential rights impacts. Kenya's data protection regulator took the opposite line in a draft dated July 2026, requiring a DPIA outright for content recommendation on consumer platforms, one of eight illustrative AI uses placed in that tier.

Assessments are also enforced, not just written. In March 2026, Spain's data protection authority fined FC Barcelona 500,000 euros over an Article 35 failure linked to a biometric membership census. And the four-level language of the Commission note has a familiar ring: in September 2025, SURF, the Dutch education and research cooperative, moved risks identified in its Microsoft 365 Copilot DPIA from high to medium while other privacy concerns persisted, the same residual-risk logic the case study applies.

The note's treatment of manipulation sits in a crowded field. The Commission's February 2025 guidance on prohibited practices prompted debate over where influence ends and manipulation begins under the AI Act, and the Council of Europe has issued guidelines for equality bodies on using the AI Act against algorithmic discrimination. On training data, the European Data Protection Board said in December 2024 that AI models trained on personal data cannot automatically be considered anonymous. Employment AI, another category in the annex, already has a test case in the United States, where a federal judge extended the age discrimination collective in Mobley v. Workday to applicants screened through HiredScore AI features.

What the note does not do is equally important. It creates no obligation for any company, sets no deadline, and is explicit that its template does not cover robustness, security or conformity assessment. Its significance lies in the method: per-right scoring, no numbers, no averaging, and a documented record of every choice in the matrix.

Timeline

Summary

Who: The Research Ethics & Integrity Team of the European Commission's Directorate-General for Research and Innovation, with a methodology prepared by expert Alessandro Mantelero and coordinated by Commission staff member Mihalis Kritikos. It is aimed at applicants, beneficiaries and ethics reviewers in Horizon Europe and other EU-funded programmes.

What: A non-binding ethics note setting out a Fundamental Rights Impact Assessment method for AI in research. Each potentially affected right is rated separately for likelihood (probability and exposure) and severity (gravity and effort) on four ordinal levels, combined through 4x4 matrices, with no numeric scores and no averaging across rights. It includes a researcher template, a six-question reviewer checklist, examples linked to 35 Charter articles, a map of 22 AI risk sub-domains to rights, and a case study on a chemotherapy prediction tool.

When: The note is dated May 2026. Its cited sources were accessed between January and May 2024. AI Act deployer obligations for Annex III high-risk systems, including the Article 27 assessment, apply from December 2, 2027.

Where: The European Union, within the ethics review procedure for EU research funding run from Brussels; the case study spans ten European countries and eight healthcare centres worldwide.

Why: According to the note, data protection impact assessments do not fully capture the effect of AI on rights other than privacy, and researchers often lack the legal expertise to assess those effects. The method is meant to make that assessment evidence-based and accountable, and the note says it is aligned with the AI Act's rights assessment obligations, which reach deployers of AI for credit scoring and life and health insurance pricing.