Skip to the articles
LAW RIGHTNEWSLETTER / 013

THE DECISIONS THAT AFFECT YOU

17 November 2026

What is still yours

Read the articles ↓
Gemini film · 10 seconds · Film generated with Gemini from an illustration created with OpenAI. Fictional scene.AI generated

LAW RIGHT / 013

A walk with Milo and Anouk

Three subjects, three songs. Start once: the journey continues on its own, and you stay in control.

One edition. Three ways in.

Your pay. Your working future. Your memories.

You can work for years without knowing the criteria behind your pay. Be assessed without knowing what software has recorded about your work. Leave thousands of photographs without saying what your family may do with them. This edition makes these decisions easier to see, understand and, where necessary, challenge.

Lexi · AI assistant

LEXI · AI assistant

A question along the way?

Lexi, Milo and Anouk accompany your reading. Take the time to choose your path.

01 / Pay

Your colleague earns more than you. Can you find out why?

Compare the work before comparing the numbers.

The conversation starts almost by accident. A colleague mentions their pay. The amount surprises you: your jobs seem similar, with responsibilities you thought were comparable. You do the calculation in your head. Then comes the question you hesitate to ask: what explains the difference? This scene is fictional, but the legal problem is specific. How can someone establish discrimination when the information needed to demonstrate it remains with the employer?

AI — creation assisted or generated by artificial intelligenceText prepared with AI: drafting, analysis and translation. Uses and provenance

What Europe wants to make visible

Directive (EU) 2023/970 concerns equal pay for women and men doing the same work or work of equal value. It promises neither identical pay for everyone nor unrestricted access to individual payslips. Its purpose is to make it possible to examine differences that secrecy has kept out of view. Higher pay may be justified by objective, gender-neutral criteria. Those criteria, their relevance and their actual application still need to be identified. S1

Milo · CuriosityWhat can we really compare?Virtual character · AI-generated image

The real work therefore begins before the salary spreadsheet. Equal value does not necessarily mean the same job title. Skills, effort, responsibility and working conditions must be examined. Relevant interpersonal skills must not be undervalued. Conversely, an identical title does not by itself establish that actual responsibilities are comparable. The category chosen must enable a fair comparison; arbitrary groupings must not hide the gap being investigated. S1, S2

Pay also means more than the monthly transfer. The European definition covers basic pay and other employment-related consideration, in cash or in kind, including variable or complementary components. Comparing two take-home amounts alone can therefore mislead: working time, bonuses and benefits need to be considered on a consistent basis. The directive expresses pay levels as gross annual pay and the corresponding gross hourly pay. S1

A 4% gap is not authorised discrimination.
Ten seconds to reflect · English narration.AI-generated contentVoice and images generated with Gemini · fictional illustration.

The video starts silently when you reach it. You choose when to enable sound. Captions and a transcript are available.

Read the transcript

Comparing pay means comparing work. Transparency starts with asking the right questions.

Download the transcript

What a worker will be able to request

Article 7 establishes a right to request and receive written information about individual pay and average pay levels, broken down by sex, for categories doing the same work or work of equal value. Requests may be made through workers’ representatives or an equality body. Incomplete or inaccurate information can prompt requests for clarification. The response must arrive within a reasonable period and no later than two months. Employers must remind workers of this right each year. S1

Anouk · ReflectionLet’s begin with the right questions.Virtual character · AI-generated image

This mechanism concerns comparable averages. It does not automatically reveal the named colleague’s pay. The GDPR still applies. Where a small group could identify someone, Member States may restrict certain information to representatives, labour inspectorates or equality bodies. Privacy must be protected without making equal-pay scrutiny ineffective. S1

The criteria for setting pay and pay progression must also be accessible. A national option to exempt employers with fewer than fifty workers concerns progression, rather than a general exemption from equal pay. Before hiring, the text provides for information about initial pay or its range early enough to enable informed negotiation, and prohibits questions about pay history. These rules should not be turned into a uniform European requirement to state an amount in every advertisement. S1

The five per cent trap

A 4% gap is not permission to discriminate. The 5% threshold triggers a particular process for employers subject to reporting: a joint pay assessment with representatives where a category has an average gap of at least 5%, the employer has not justified it through objective, gender-neutral criteria, and it has not remedied the unjustified difference within six months of the date the pay report was submitted. All three conditions must be met. Discrimination may still be challenged below that threshold. S1

Milo · AttentionA threshold does not tell the whole story.Virtual character · AI-generated image

Reporting has a separate timetable: from 7 June 2027, annually for employers with at least 250 workers and every three years for those with 150–249; from 7 June 2031, every three years for those with 100–149. The figures concern the preceding calendar year. Member States may extend reporting below one hundred workers. The absence of mandatory reporting does not remove the text’s other requirements. S1

Information must be capable of changing the outcome

The directive provides for full compensation without a predetermined cap, covering matters including back pay, related payments, lost opportunities, non-material damage and interest. It also strengthens the evidential rules: where facts support a presumption of discrimination, the employer must rebut it. A breach of transparency duties may also shift the burden, subject to the exception for a manifestly unintentional and minor breach. National procedure and the facts of each case still matter. S1

Anouk · ListeningUseful information starts a conversation.Virtual character · AI-generated image

In Belgium, the European deadline of 7 June 2026 must be distinguished from the measures actually adopted and their scope. The European register consulted on 16 November 2026 listed six notified Belgian measures, including public-service and regional or community texts. That record does not certify complete federal transposition for private-sector employment. Regional measures do not establish that the same obligation already applies to every employer nationwide. S3

Belgium already has a collective mechanism: businesses habitually employing an average of at least fifty workers must analyse their pay structure every two years. The report goes to the works council or, failing that, the trade union delegation. It does not give individual workers access to colleagues’ payslips, but it provides a practical starting point for discussions with staff representatives. S10

THE DOSSIER IN VOICES & PICTURES

01 / AUDIO · VIDEO

Another way into the subject

Listen to the analysis

Automatic transcript available

AI generatedVoice generated with NotebookLM. Editing and visual presentation produced with Codex. Scenes are illustrative.

Read the transcript

Welcome to the deep dive. Just a quick note up front. Our voices today are generated with AI using NotebookLM but the scenario We're unpacking is all too human. Oh, definitely Imagine chatting with a colleague you're doing the math in your head and realizing they make more than you for the exact same job You're left wondering why but I mean, how do you actually prove discrimination when your employer holds all the cards, right? It's the ultimate information asymmetry. The employer has the master spreadsheet and Well, you only have your single paycheck But the new EU directive 2023-970 on equal pay transparency is designed to fundamentally flip that dynamic Okay But I want to be clear on how it flips it because the directive doesn't just like email you everyone's individual pay slips No, no, it doesn't do that and honestly just comparing base salaries is sort of like comparing the size of two houses by only measuring the living room You're completely ignoring the garage in the backyard in the corporate world That's you know the overtime the bonuses and the benefits that is a perfect way to look at it The European definition of pay actually includes that whole property. So the basic pay plus any variable or complementary components Whether they are in cash or in kind. Oh, wow. Yeah, okay Yeah, and the directive doesn't just look for matching job titles either It evaluates work of equal value, which means analyzing skills effort responsibility and working conditions So an employer can't just hide behind arbitrary job groupings to justify a gap Okay, but if I'm not getting my co-workers specific pay slip, what data can I actually ask for you? Have the right to request comparable averages broken down by sex for categories doing the same work or work of equal value So it's grouped data exactly So we know what data employees can ask for but when the companies actually have to you know open up their books, right? So it's a phased rollout based on headcount Employers with at least 250 workers have to start reporting annually on June 7 2027. Okay, companies with 150 to 249 workers also start in 2027 but they only report every three years and Finally companies with 100 to 149 workers join that three-year cycle starting in 2031 Wait, let me stop you there and push back of it. Mm-hmm because there's a specific threshold mentioned in the source material That feels like a massive loophole. Mm-hmm. It mentions a 5% pay gap threshold. Does that mean? Say a 4% gap is just permitted discrimination Absolutely not and I'm glad you brought that up a 4% gap is not permitted discrimination You can always challenge that the 5% figure Specifically triggers what's called a joint pay assessment process. Oh, I see and that assessment actually requires three cumulative conditions before it kicks in. Three conditions. Walk me through those. Sure. First, There must be an average category gap of at least 5% right second that gap is not objectively Gender-neutrally justified, okay, and third the gap has not been remedied within six months of the submission of the pay report All three of those must be met to trigger the joint assessment. Okay That makes a lot more sense now Let's talk about the legal reality because the impact here is fascinating Under this directive the civil burden of proof can shift right meaning the employer has to prove they aren't discriminating rather than you having to prove They are yes breach of transparency duties can shift the civil burden of proof But we need to be careful this only applies based on specific article conditions So it's not just automatic for everything, right? It is never a guilty until proven innocent situation And it never presumes every pay difference is automatically discriminatory It just means if the facts support a presumption of discrimination, it's on the employer to rebut it What about small businesses now? Do they get a free pass because the source mentions it under 50 exemption? Yeah, that's a common misunderstanding The national option to exempt employers with fewer than 50 workers strictly concerns pay progression criteria It is not a general exemption from equality reporting or equal pay principles. Got it Let's look at how this plays out in reality Like in Belgium the source mentioned six notified Belgian measures on a register. Does that mean they fully adopted the EU rules? Not quite Those six notified measures are just a bounded register observation Meaning what exactly? Essentially, they are officially logged on a register But that alone isn't proof of complete federal transposition for the private sector Belgium isn't exactly starting from scratch, right? Don't they already mandate a pay analysis every two years for businesses that Habitually employ an average of at least 50 workers. They do. Yeah, it doesn't hand over the individual pay slips But it brings the data out of the shadows giving a works councils and trade unions a huge head start So what does this all mean for you listening right now? If you want to start a useful discussion at work, you should be asking practical questions like which criteria apply to your role Exactly, which comparison category are you grouped into and which pay components are included in your evaluation? keep your requests and Definitely keep the answers and here's an important final thought to leave you with an explanation drafted by an employer After they are challenged is no substitute for consistent transparent criteria. Yeah, that's a great point So how will your employers hidden arbitrary groupings survive when secrecy is no longer an option?

Download

Watch the story

Automatic captions available

AI generatedVideo overview generated with NotebookLM; local editing and editorial corrections with Codex. AI-generated voices and visuals. Fictional illustrations.

Read the transcript

The conversation usually starts by accident. A colleague mentions their pay, and the number catches you completely off guard. You immediately run the calculation in your head. Your jobs look the same, and you carry the exact same responsibilities. Then comes the question you hesitate to ask. Why is there a difference? Workplace salaries have long been protected by a culture of strict secrecy. Directive 2023/970 is the European Union's mechanism to dismantle this exact information vacuum. To be clear, this text does not guarantee identical pay for everyone across the board, nor does it open up individual payslips for public viewing. Its purpose is to force visibility. It allows workers and equality bodies to examine differences that secrecy has historically kept out of view. Employers can still pay someone a higher salary, but they now have to justify that difference using objective, gender-neutral criteria. By mandating this access to information, the directive shifts the power dynamic of the workplace from blind trust to verifiable transparency. The directive introduces a rigorous new framework for examining equal work. Notice how equal value is not determined by a single label. Value is measured through specific criteria—skill, effort, responsibility, and working conditions. Each pillar carries structural weight independent of a job title. Relying on identical job titles alone is no longer sufficient. Two people might hold the same title, but if actual responsibilities and working conditions differ, the title can actively hide the pay gap. The directive also redefines compensation. A standard take-home pay block is contrasted against a much larger EU pay definition stack. Comparing only net monthly transfers is misleading. By establishing strict definitions for both the value of the work and the total compensation, the directive prevents employers from hiding discrimination behind arbitrary job categories or incomplete salary figures. Under Article 7, workers have the right to request written information about their individual pay level, as well as the average pay levels for their category, broken down by sex. This mechanism protects privacy under the GDPR. A new compliance mechanism starts with a trigger node labeled "5% Gap," which funnels into a mandatory joint pay assessment. Three prerequisite gates pulse simultaneously. However, a 4% gap is not a safe harbor. Pay differences below the 5% threshold can still be legally challenged by workers. If a dispute goes to court, the directive transfers the legal burden of proof from the employee to the employer. The first trigger for this shift occurs when a worker establishes facts that support a presumption of discrimination. Once that happens, the employer must actively rebut the claim. The second trigger is a breach of transparency duties. If an employer fails to provide the required pay information, they assume the burden of proof, unless the breach was manifestly unintentional and minor. Employers can no longer default to silence as a defense strategy. They are legally forced to actively justify their pay structures. The rollout of these rules follows a strict European timetable. Mandatory reporting begins on June 7, 2027 for employers with at least 150 workers. The subsequent phase starts on June 7, 2031, expanding the reporting requirements to employers with between 100 and 149 workers. But for member states like Belgium, the immediate hurdle is the transposition deadline of June 7, 2026, which requires embedding these rules into national law. This typographic layout contrasts that 2026 European deadline with a segmented map of Belgium, highlighting a highly fragmented implementation process. While some regional and public service measures have been notified, this does not certify that complete federal transposition is in place for private sector employment. The Belgium already has a collective mechanism in place. Businesses that employ an average of at least 50 workers are required to analyze their pay structure every two years. This biennial report is sent to the Works Council or trade union delegation. It provides a practical starting point for discussions with staff representatives today. To initiate a useful discussion about your own pay, there are three precise questions you should ask your employer or representative. First, which criteria apply to me? Second, which comparison category is used? Third, which pay components are included? Always document this exchange. Keep your request and keep the employer's answer. The practical advance of these legal mechanisms is the ability to examine a pay difference objectively. It replaces workplace rumors with structured, documented demands for fairness.

Download
Explore the sources

02 / Work & AI

A score to question.

Open the dossier ↗

AI-generated video backdrop · silent sequence

Continue reading →

02 / Work and AI

Will a machine decide your next performance review?

Behind the score, a decision you should be able to question.

Your manager opens a dashboard. A score summarises your targets, exchanges and activity. You ask why it is so low. “That is what the tool says.” This fictional scene raises a practical question: who can reconsider the conclusion, and on what information?

AI — creation assisted or generated by artificial intelligenceText prepared with AI: drafting, analysis and translation. Uses and provenance

A score does not explain a person

Software can help organise information. The difficulty arises when its recommendation becomes decisive without meaningful scrutiny. Article 22 GDPR addresses decisions based solely on automated processing that have legal or similarly significant effects. Not every use of a dashboard automatically falls within that prohibition. But a manager who mechanically approves the result does not create genuine human involvement merely by being present. S4, S5

Anouk · ReflectionThere is a story behind the score.Virtual character · AI-generated image

Exceptions must be assessed, not assumed: contractual necessity, legal authorisation or explicit consent, with appropriate safeguards. In employment, it is particularly important to establish whether consent can genuinely be freely given. A worker should not be reduced to accepting a tool to avoid disadvantage. Other GDPR requirements, including lawfulness, transparency and accuracy, remain relevant even where a decision is not solely automated. S4, S5

Suppose a system interprets a late reply as lack of commitment. Was the worker on leave, carrying out fieldwork or performing a task that left no trace in the software? These questions do not prove that all tools assess people poorly. They show why data must be checked against actual work before it affects a bonus, promotion or continued employment.

The timetable has changed; responsibilities remain

The European AI Act lists certain recruitment, promotion, dismissal and performance-assessment uses among high-risk uses. The intended purpose and classification conditions of each system must nevertheless be checked. A writing function is not the same as a system intended to evaluate staff. S6

Milo · ListeningWho can still review the result?Virtual character · AI-generated image

The original timetable must not simply be repeated. The consolidated version of 27 July 2026, amended by Regulation (EU) 2026/1744, sets 2 December 2027 for the application of Chapter III, Sections 1–3, to systems covered by Article 6(2) and Annex III, except for Article 6(5). This postpones neither the GDPR nor the entire AI Act. The prohibition on certain systems inferring emotions at work is separate, with the medical or safety exception set out in the text. S6

High-risk classification does not mean prohibition. Since 2 August 2026, Article 86 has also provided, subject to its own conditions, a clear explanation of AI’s role and the main elements of certain decisions based on a high-risk system listed in Annex III. It concerns decisions with legal or similarly significant effects and an alleged adverse impact on health, safety or fundamental rights. It is not a general right to source code and applies insofar as a comparable right is not already guaranteed under Union law. S12

The issue goes beyond fear of machines. A good tool can reveal an inconsistency and support a better-informed conversation. It becomes dangerous when apparent precision closes that conversation. An assessment must remain an opportunity for a person to contribute what the data missed.

Your turn

Where does the person still have a say?

In your example, can a person genuinely reconsider the decision?

02 / AUDIO · VIDEO

Another way into the subject

Listen to the analysis

Automatic transcript available

AI generatedVoice generated with NotebookLM. Editing and visual presentation produced with Codex. Scenes are illustrative.

Read the transcript

Voices in the deep dive are generated with AI using NotebookLM. Okay, so picture this for a second. You walk into your performance review, right? And your manager just slides this dashboard across the desk. - Oh boy, yeah, I know where this is going. - Right, and your score is, well, it's shockingly low. And when you ask why, they just kind of shrug and say, you know, that's what the tool says. - Yeah, it's such a common scenario now, isn't it? But legally speaking, your boss can't just, you know, pass the buck to the software like that. I mean, under Article 22 of the GDPR, a decision has to be strictly solely automated. - Solely automated, right. - Exactly, solely automated. And it has to have a legal or similarly significant effect on you to actually trigger those strict protections. - Okay, but let me stop you right there because, I mean, if my manager is sitting right across the desk for me, physically handing me the review, is it actually solely automated? Like, what if they just glance at the AI score? Maybe they hesitate for a second and then they just mechanically approve it. That doesn't feel like a human made the call. - Well, that right there, you've hit on the exact friction point. I mean, real human scrutiny really matters here. Like, a manager merely hesitating before signing off that is definitely not proof of genuine oversight. - Wow, okay, so they can't just rubber stamp it? - No, absolutely not. The law requires them to actually assess the decision. And if the decision is deemed solely automated, then the employer has to rigorously prove an exception applies. - Like, what kind of exception? - Well, they need to show it's necessary for a contract or authorized by statutory law or that they have your explicit consent along with appropriate safeguards. - Right, consent, but in a workplace, who is actually going to say no to their boss? - Precisely, if we look at the reality of employment, consent is heavily scrutinized. You really cannot assume a worker is freely giving consent if they fear they'll face some sort of disadvantage for saying no. Consent is not, you know, a magic standalone remedy here. - Okay, so if consent isn't this get out of jail free card under the GDPR, what happens when we mix in the shifting timelines of the European AI Act? Because I'm looking at the July 27, 2026 consolidation, and it says certain rules apply by December 2nd, 2027. Does that mean companies can just, like, hit snooze on AI compliance for now? - No, no, no, that is a massive misconception. You absolutely cannot postpone the whole of Chapter III. - Oh, wait, really? - Yeah, that December 2, 2027 date is highly specific. It applies only to Chapter III, Sections 1–3 for systems covered by Article 6(2), and Annex III, except for Article 6(5). in the AI Act apply right now. Like, if a company uses a system intended to infer your emotions at work, that is prohibited today, unless it falls under the very specific medical or safety exceptions detailed in the text. - Okay, so with these active laws in play, what actual recourse do you have if the machine gets it wrong? Like, let's say the software flags you for a, quote unquote, "lack of commitment" because of a late email reply, completely missing that you were offline doing authorized field research. Do you just push a button and demand to see the AI source code? - No, it's actually much stricter than that. That brings us to Article 86, which is applied since August 2, 2026. This isn't just, you know, a general complaint box. or similarly significant effects. Plus, you have to allege an adverse effect on your health, safety, or fundamental rights. - So it's a high bar. - Very high. It's conditional and subsidiary. It provides a right to an explanation for certain high-risk systems, but it offers absolutely no broad entitlement to the actual source code. Think of it as a targeted safety valve, not, you know, an open door for everyone. - Man, it seems like while good tools can reveal inconsistencies and, you know, start a conversation, they cross a really dangerous line when their apparent precision is used to shut down human dialogue completely. - Yeah, that is the core issue here. An assessment has to remain an opportunity for you to contribute what the data missed. Just slapping a signature at the bottom of a page does not establish human oversight. - So here is your actionable takeaway from this. At your next review, use the practical questions from our source. Ask your manager which data informed this assessment. What does the system actually do and who has the authority to correct its output? - I love that. Those questions really force the conversation back into human hands. - Exactly. Which leaves us with this to ponder. If algorithms are only capable of measuring our trackable digital footprints, what deeply valuable human contributions are quietly becoming completely invisible in the modern workplace?

Download

Watch the story

Automatic captions available

AI generatedVideo overview generated with NotebookLM; local editing and editorial corrections with Codex. AI-generated voices and visuals. Fictional illustrations.

Read the transcript

Your manager opens a dashboard. A single score summarizes your targets, your email exchanges, and your entire activity over the last year. It's surprisingly low. You ask why. The manager points to the screen and says, "That is what the tool says." If a machine dictates your performance score, who has the actual authority to reconsider that conclusion, and based on what information? Software excels at organizing massive amounts of workplace data, but a critical danger emerges when an algorithm's recommendation becomes decisive without any meaningful human scrutiny. A manager sitting in the room mechanically signing the bottom of an automated report does not constitute genuine human oversight if they blindly accept the AI's output. Not every workplace dashboard triggers Article 22, but even if a human is in the loop, the employer still has to follow general GDPR rules. Bypassing the strict Article 22 prohibition requires specific exceptions. The automated decision must be a contractual necessity, it must be authorized by law, or it must be backed by explicit consent. In an employment context, consent is legally fragile. If refusing an AI evaluation leads to a career disadvantage, the worker's consent cannot genuinely be considered freely given. Look at this visualization, comparing tracked digital activity against untracked physical work. The algorithm sees a dense cluster of emails and logins, but interprets off-grid field work as a completely empty void. It flags a late reply as a lack of commitment, simply because physical work left no digital trace. Courts have already ruled on automated credit scoring, but those precedents cannot simply be copy-pasted into the highly nuanced relationship between an employer and an employee. Relying on algorithms to define a worker creates a false certainty. Data must always be cross-checked against actual, real-world work before it influences a promotion, a bonus, or continued employment. The Incoming European AI Act adds another layer of regulation by treating systems differently based on their purpose. A basic AI writing assistant faces entirely different rules than software built to evaluate staff. This timeline tracks the rollout. But notice this separate warning symbol. The AI Act includes an immediate prohibition against systems inferring employee emotions. Employers can only bypass this emotion inference ban if they have explicit medical or safety reasons set out directly in the legal text. The AI Act does not replace or postpone the GDPR. It adds a targeted, date-sensitive layer of legal conditions specifically designed to prevent machines from substituting human judgment on the job. Labeling an AI as high-risk does not mean it is banned from the workplace. It means the system is now subject to heavy regulatory conditions. One of the most powerful conditions is the right to an explanation under Article 86, applicable since August 2026. We also pass a final filter, the Subsidiarity Check. The right only applies if an explanation isn't already guaranteed by union law. When triggered, this right forces the employer to explain the AI's role and the main elements of the decision. However, it does not grant the employee general access to the underlying source code. A properly utilized AI tool should reveal inconsistencies and support a better-informed conversation. It becomes dangerous when its apparent precision is used to shut that conversation down. If you face an AI-assisted performance review, start by focusing on the inputs. Question the exact metrics that informed the algorithmic assessment. Next, interrogate how the software actually processes that information. Finally, identify who holds the actual authority to fix the output. Employers must give human reviewers the time, the specific information, and the actual authority to act. An assessment has to remain an opportunity for a person to contribute exactly what the data missed.

Download
Explore the sources

03 / Digital memory

Memories to pass on.

Open the dossier ↗

AI-generated video backdrop · silent sequence

Music enters the story

A voice we keep. Words we choose.

A musical pause · Missing You

Answer Me

Full film

Start listening whenever you wish.

Two full films, ordered by descending YouTube views.

Open the Missing You album

Absences · in French ↗

Milo Sarre and Anouk Revel are virtual characters. Music, voices and images were produced with AI. Artistic direction: Chris Beaurêve.

Continue reading →

03 / Memory

Who will inherit your photographs… and who may make you speak?

What you leave behind. What you allow.

A phone may hold a parent’s last recorded words. A mailbox, years of exchanges. An account, photographs nobody printed. After a death, a family may want to recover these traces. Then another possibility appears: using recordings to create a voice that answers back. These are different undertakings.

AI — creation assisted or generated by artificial intelligenceText prepared with AI: drafting, analysis and translation. Uses and provenance

Recovering a memory and creating a presence

In an interview conducted by Geneviève Simon and published by La Libre on 28 October 2026, I discussed whether a son may bring his mother back as a digital avatar. This newsletter takes the subject further with a practical question: what do we want to authorise, pass on or rule out during our lifetime? The interview is linked in the sources. S7

Anouk · DreamingA memory brings us closer.Virtual character · AI-generated image

Keeping a recording, sharing a copy and generating new speech with a reconstructed voice have different implications. The first preserves a trace. The last attributes words the person never said. That distinction should guide preparations even before every legal question has been resolved.

Under recital 27, the GDPR does not apply to deceased persons’ data. This does not make an account freely available for access and reuse. Exchanges may include living people’s data. Photographs or texts may be protected by copyright. Service terms, inheritance rules and the rights involved need separate consideration. Technical access alone establishes no entitlement to publish or create an avatar. S8

Image rights provide another Belgian reference point. The FPS Economy states that they continue for twenty years after death and may be exercised by the heirs. They are separate from copyright in a photograph or recording. This does not establish a blanket permission or ban on voice imitation: the relevant rights and intended use still need to be identified. S11

Technical access does not, by itself, establish a right to publish or create an avatar.

03 / AUDIO · VIDEO

Another way into the subject

Listen to the analysis

Automatic transcript available

AI generatedVoice generated with NotebookLM. Editing and visual presentation produced with Codex. Scenes are illustrative.

Read the transcript

Before we jump into things today, just a quick mandatory disclosure. The voices in this deep dive are actually generated with AI using NotebookLM. Anyway, imagine inheriting like your late grandmother's photo albums. You flip through them just to remember her, right? - Yeah, exactly. It's a very normal way to grieve. - Right, but now imagine taking her old voicemails, feeding them into an AI and literally making her speak to you today, like making her say words she never actually said in real life. Today we are exploring the massive legal gap between those two realities. - It is a huge gap. I mean, we are moving from just recovering a memory to actively creating a completely new presence. - So if keeping a voicemail is just a memory, but feeding it into an AI is creating a new presence, who actually gets to make that choice? Because, I mean, if I have my late grandmother's iPhone password, doesn't that inherently mean she gave me the keys to her data? - Well, it definitely feels like it should. But that is what we call the illusion of technical access. Like the GDPR was really designed to protect the data of living people. - Okay, but if the person passed away. - Right, so when someone passes away, their personal data actually loses that specific protection under Recital 27. But think about it, an old email thread or a message history isn't just the deceased person's data. - Oh, right, because other people are in the thread. - Exactly. It contains the replies of living people and their data is still fiercely protected. - That makes sense. But even if you sort out who owns the data, you still have to deal with what you are legally allowed to do with it, right? Because we tend to lump things together. - Oh, absolutely. They actually rely on entirely different legal mechanisms. Take a photo, for instance. Copyright protects the person who actually took the photo because it is their creative work. - Right, the photographer. - Yeah, but the image rights belong to the person in the photo. So in Belgium, for example, those portrait rights last for 20 years after death and they belong to the heirs. - 20 years? - Wow. - Okay, so just having access to the file does not give you permission to recreate the person. It sounds like we are treating digital data a lot like a vintage car. - How do you mean? - Well, getting the keys and the will means you own the car, right? But it does not give you a free pass to ignore the speed limit when you take it out for a drive. - That is a really great analogy. And the speed limit in this case is the new AI Act. Since August 2nd, 2026, Article 50 of the Act has imposed some really strict duties on how these technologies are used. - What kind of duties are we talking about? - You have very distinct obligations depending on the system. Like, you have to clearly disclose if a user is interacting with an AI. You must disclose if something is a deep fake. - Oh, wow. So you have to put a label on it. - Exactly. And providers have to ensure machine-generated content is clearly marked. Roles and exceptions do exist, but you cannot just call it an art project and assume there was some blanket exemption. - Right, you can't just claim artistic freedom to get out of it. - Well, the financial education site Wikifin suggests starting with a really clear inventory of your accounts. - Just literally writing down what you have. - Pretty much. You need to formally separate what should stay private correspondence from what can actually be shared or reused by your family. - But wait, you shouldn't just type your passwords into an instruction document and pass it around, right? That sounds super risky. - Oh, definitely do not do that. You need a secure, separate way to hand those passwords over. - It is just so wild to think about. I mean, we spend months discussing who gets the house, but we hesitate over what happens to a few minutes of recorded speech. - It is true. People do not really think about their digital legacy in the same way. - It really comes down to the practical questions you need to answer while you are still here. Like, what do you want to authorize, pass on, or just totally rule out? - Exactly. And it is incredibly valid to just leave an instruction that says, keep this recording, but do not make me say new things. - Right, because technology makes the imitation possible, but it literally cannot decide what you would have wanted. So listeners, if someone inherits the keys to your digital life tomorrow, what new words would you want them to put in your mouth, if any at all?

Download

Watch the story

Automatic captions available

AI generatedVideo overview generated with NotebookLM; local editing and editorial corrections with Codex. AI-generated voices and visuals. Fictional illustrations.

Read the transcript

A son sits in a quiet room and swipes through his mother's phone. He holds years of text exchanges, a vault of unprinted photos, and an audio file containing her last recorded words. We rarely hesitate when passing down a physical property, like a home, but inheriting a few minutes of recorded speech presents a much more complicated reality. Holding the physical device doesn't automatically grant legal access to the accounts inside it, and possessing a digital souvenir certainly doesn't give someone the legal right to freely use it. This tension recently sparked a debate in a La Libre interview between Geneviève Simon and lawyer Christophe Boeraeve. The core question? Does a relative have the legal right to take those recordings and bring a deceased loved one back as a digital avatar? Navigating this space requires separating two completely different actions, preserving a trace to recover a past memory, and using that trace to fabricate a new synthetic presence. Logging into an inbox or a social account is just technical access. It establishes no legal entitlement to publish or reuse what you find inside. Under Recital 27 of the GDPR, data protection rules do not apply to deceased persons, but digital mailboxes and text chains are two-way streets. They contain the data of living people, whose privacy remains fully protected. Even a single audio file is bound by multiple layers of law. First, if the recording contains an original written work or composition, it is protected by copyright. Next, the actual sound recording of the performance is separately protected by what are known as related rights. Because these multiple legal protections overlap on a single digital asset, inheriting a recording does not give relatives blanket permission to reuse or alter it. France, for example, has a specific formal system allowing individuals to leave postmortem instructions for their digital privacy. Under Belgium law, image rights continue for exactly 20 years after death, and during that time they can be exercised by the deceased's heirs. However, this 20-year image right is highly specific. It establishes neither a blanket permission nor an absolute ban on voice imitation. When we cross the threshold from preserving a memory to creating a new digital presence, we enter the domain of the EU AI Act, specifically Article 50, which governs synthetic media and interactive AI. This article imposes two distinct duties. First, a duty to inform users when they are interacting directly with an AI system. Second, a separate duty to disclose when specific audio or video content is an artificially manipulated deepfake. There are exemptions built into the Act, clearly artistic works, have adapted disclosure rules, and purely personal, non-professional use falls entirely outside the legal definition of an AI deployer. But even if an avatar perfectly complies with AI Act disclosure rules, that compliance does not clear the underlying copyrights, image rights, or platform terms required to use the source material in the first place. While AI technology makes imitating a human voice highly accessible, a machine cannot legally or ethically determine what the deceased person would have actually wanted. Leaving explicit, practical instructions provides vital guidance for grieving relatives. The first action is to make a complete inventory of your digital accounts, files, and contacts, and discuss these assets with a notary. Second, explicitly state which photos or correspondence must remain private. You should directly authorize or refuse new uses of your voice, such as strictly forbidding the generation of new synthetic speech. Third, securely arrange the transmission of your access details. Passwords should never be placed in a standard document intended for general circulation. Your grieving son now understands the boundaries of the archive he holds. He possesses the data, but he respects the limits of its use as he sets the device down. The law provides a complex framework of overlapping rights.

Download
Explore the sources

Leaving something people can actually use

France has a particular system for post-mortem instructions described by the CNIL. It should not be presented as an identical Belgian procedure. For someone in Belgium, useful preparation starts with an inventory: which accounts, which files, which contacts and which wishes should be discussed with a notary? Wikifin highlights the importance of instructions for digital assets; practical arrangements need to reflect access and the nature of each asset. S8, S9

Milo · GratitudeLet’s leave room for choices too.Virtual character · AI-generated image

An avatar raises two distinct questions: does the person speaking to it know it is AI, and could its words be mistaken for a genuine recording of the deceased? Since 2 August 2026, Article 50 of the AI Act has imposed separate duties for informing people about AI interaction and disclosing certain audio or video constituting deepfakes. Clearly artistic or fictional works have an adapted disclosure rule, not a general exemption. Scope depends on the role and context; personal, non-professional use falls outside the definition of deployer. A disclosure does not resolve the other rights in the material used. S13

We readily discuss passing on a home. We hesitate over a few minutes of recorded speech. Yet “keep this recording, but do not make me say new things” can express an important wish. Technology makes imitation possible; it cannot decide what the person would have wanted.

Your turn

Prepare the conversation

0 / 3 points explored

Your choices stay on this page. No account or password is requested.

MISSING YOU / FILM

The Ones We Miss

Milo Sarre and Anouk Revel are virtual characters. Music, voices and images were produced with AI. Artistic direction: Chris Beaurêve.

Explore Missing You ↗
Return to the three articles →

Sources and reference points

Source titles are retained in their original language.

  1. S1 · Directive (UE) 2023/970, articles 3 à 12, 16, 18 et 34
  2. S2 · SPF Emploi, projet BE-MAGIC
  3. S3 · EUR-Lex, mesures nationales notifiées pour la directive 2023/970
  4. S4 · APD, décision individuelle automatisée ; RGPD, articles 5, 6 et 22
  5. S5 · Groupe Article 29, WP251rev.01, décision automatisée et profilage, version du 6 février 2018
  6. S6 · Règlement IA 2024/1689, version consolidée du 27 juillet 2026 après règlement 2026/1744
  7. S7 · Geneviève Simon, entretien avec Christophe Boeraeve, La Libre, 28 octobre 2026, sur les avatars numériques
  8. S8 · CNIL, Mort numérique, 31 octobre 2025 ; RGPD considérant27 ; droits distincts
  9. S9 · Wikifin FSMA, Pourquoi planifier votre succession, mise à jour 21 avril 2026
  10. S10 · SPF Emploi, égalité femmes-hommes : analyse bisannuelle des rémunérations
  11. S11 · SPF Économie, droit à l’image, mise à jour 3 décembre 2020
  12. S12 · Commission européenne, AI Act Service Desk, article 86
  13. S13 · Commission européenne, AI Act Service Desk, articles 50 et 113
Download the reading PDF
Lexi

A question stays with you?

LEXI

Information must enable understanding and challenge.

Lexi is an AI assistant. Its answers may contain errors; check the sources and the context of your situation.

The Lexi plugin in ChatGPT includes issues 001–013, their articles and source links. You can also contact Christophe Boeraeve.

Talk to Lexi ↗Contact Christophe Boeraeve

WhatsApp +32 467 11 17 77

How this edition was created

Drafting, analysis and translation: passages were generated with Codex; CUE/Manus and Claude contributed to research and critical reviews.

The nine short films across the three language versions and the new opening film were generated with Gemini. The opening’s reference image was created with OpenAI.

The audio analyses and video overviews were generated with NotebookLM. The video overviews shown are copies edited and corrected locally with Codex; the originals are preserved. Voices and visuals are AI-generated. Transcripts and captions are produced by automatic speech recognition and may contain errors.

About the European icons ↗

LAW RIGHT / 013

View the sources

The references for this edition, within reach as you read.

  1. S1 · Directive (UE) 2023/970, articles 3 à 12, 16, 18 et 34
  2. S2 · SPF Emploi, projet BE-MAGIC
  3. S3 · EUR-Lex, mesures nationales notifiées pour la directive 2023/970
  4. S4 · APD, décision individuelle automatisée ; RGPD, articles 5, 6 et 22
  5. S5 · Groupe Article 29, WP251rev.01, décision automatisée et profilage, version du 6 février 2018
  6. S6 · Règlement IA 2024/1689, version consolidée du 27 juillet 2026 après règlement 2026/1744
  7. S7 · Geneviève Simon, entretien avec Christophe Boeraeve, La Libre, 28 octobre 2026, sur les avatars numériques
  8. S8 · CNIL, Mort numérique, 31 octobre 2025 ; RGPD considérant27 ; droits distincts
  9. S9 · Wikifin FSMA, Pourquoi planifier votre succession, mise à jour 21 avril 2026
  10. S10 · SPF Emploi, égalité femmes-hommes : analyse bisannuelle des rémunérations
  11. S11 · SPF Économie, droit à l’image, mise à jour 3 décembre 2020
  12. S12 · Commission européenne, AI Act Service Desk, article 86
  13. S13 · Commission européenne, AI Act Service Desk, articles 50 et 113