
Sign up to save your podcasts
Or


EU in Practice | REDefine
In 2023, Amnesty International researchers created TikTok accounts designed to resemble thirteen-year-old users. When those accounts repeatedly engaged with videos about mental health, the recommendation system responded quickly. Within three to twenty minutes, more than half the videos appearing in some of the simulated feeds concerned mental health struggles, including material that romanticised depression or self-harm.
The accounts belonged to researchers, not actual children. The experiment didn’t establish that a particular teenager had been harmed. What it demonstrated was how quickly a recommendation system could concentrate sensitive content once it detected a pattern of attention.
In a separate automated experiment, after five to six hours, almost half the videos delivered to accounts signalling an interest in mental health were potentially harmful mental-health-related recommendations—roughly ten times the proportion shown to comparison accounts.
The findings, published in Amnesty International’s Driven into Darkness investigation, raise a serious question.
What if the system is responding to a child’s worry as though it were an interest they want to pursue?
A teenager might watch a video about depression because they are concerned about a friend. Another might watch content about dieting because a remark at school has made them question their appearance. A third might replay a frightening video because they can’t understand whether what they have seen is true.
The platform can observe the viewing time, the replay and the next interaction. It can’t directly observe the reason.
This is where a system designed to predict attention can begin to misunderstand, and eventually harm, the person behind it.
This Substack is reader-supported. To receive new posts and support my work, consider becoming a free or paid subscriber.
When watching becomes a recommendation
On 6 November 2026, the European Commission published preliminary findings that TikTok’s For You recommender system may breach the Digital Services Act because of harmful rabbit-hole effects.
The Commission’s investigation concerns more than whether individual videos violate platform rules. It examines how a recommender system repeatedly selects material about subjects such as sadness, isolation, mental health and body image, potentially intensifying risks to users’ wellbeing.
One finding is particularly revealing. The Commission preliminarily concluded that TikTok’s algorithm gives viewing time greater weight than a user’s explicit preference to see less of certain content. In practice, pressing Not interested may not be enough to change the recommendations.
The Commission has suggested that TikTok should give greater weight to explicit preferences, reconsider viewing time as a recommendation signal, improve controls and regularly test the system’s behaviour.
These are preliminary findings, not a final decision. TikTok has the right to respond, and the investigation remains subject to due process.
Yet the underlying problem extends beyond one platform.
A recommendation system doesn’t merely observe interests. It also decides which opportunities for attention appear next. A person watches a video; the platform interprets that behaviour; another video appears; the response to that recommendation becomes the next piece of evidence.
The system is learning from a sequence it helped create.
Sometimes that works beautifully. A child discovers astronomy and is introduced to physics, engineering and the history of space exploration. Another begins watching a drawing tutorial and encounters new artistic techniques.
But the same process can produce a narrowing effect. A few videos about a particular worry become a continuous stream of similar material. The person sees more of it, reacts to more of it, and provides the system with additional evidence that the subject deserves their attention.
A child’s attention isn’t a standing instruction to keep recommending the same subject.
That distinction matters especially when the content concerns vulnerability, identity, health or belonging.
What children’s experiences tell us
The scale of children’s exposure to difficult online material isn’t speculative.
The EU Kids Online 2026 comparative study surveyed 28,465 children aged 10–16 across nineteen European countries.
Among the children surveyed, 39% reported encountering online content about ways to become very thin, while 37% had encountered material about physically harming oneself. Only 48% said they felt safe online.
The proportion reporting that something online had bothered or upset them during the previous month was 23%, compared with 7% in the 2020 survey.
These figures describe children’s reported experiences. They don’t establish that recommendation algorithms caused each encounter, nor that exposure automatically produces psychological harm. But they do tell us that children’s online experiences include forms of distress that can’t be treated as exceptional or irrelevant.
The same report reveals another tension. Only half the children surveyed said their parents talked with them about what they do online, compared with 66% in 2020. Meanwhile, parental reliance on technical controls had increased.
Those findings should make us question a familiar assumption: that providing parents with another safety setting is equivalent to protecting a child.
A parent may know that a platform has parental controls without understanding which recommendations a child receives, how those recommendations develop or whether the child feels comfortable discussing an upsetting experience.
And children don’t always interpret their feeds as adults expect.
In a 2025 Pew Research Center study, an American teenage girl described appearance pressure this way:
The people they see on social media, it makes them think they have to look and be like them or they won’t be liked.
That observation isn’t a technical account of an algorithm. It is a description of how the resulting environment can feel to a young person.
The same Pew research found that 74% of American teenagers credited social media with helping them feel more connected to their friends, while 63% valued it as a place for creativity. These are not European estimates, but they help correct an important distortion in the debate.
Children’s digital lives contain opportunities as well as risks. A teenager can find friendship, discover a talent, learn about mental health or encounter a supportive community online. Those benefits are reasons to improve digital environments, not reasons to dismiss the risks.
We should be wary of a debate that treats every hour online as evidence of harm. We should be equally wary of one that assumes a platform is safe simply because most of its individual videos are lawful.
What the research can, and can’t, establish
A 2026 European Commission research review, synthesising 55 longitudinal and experimental studies, found no uniform association between time spent on social media and poorer mental health across all young people.
The evidence is more consistent for problematic or dysregulated use and harmful experiences, including appearance-related social comparison and disrupted sleep. Even there, the researchers caution that findings require careful interpretation.
This changes how we should approach children’s online safety. Counting minutes may be useful in some circumstances, but it tells us relatively little about the quality of the experience.
A particularly instructive 2024 study published in Body Image examined 1.03 million TikTok recommendations delivered over one month to 42 people with eating disorders and 49 comparison participants.
Researchers found that participants with eating disorders received substantially more appearance-oriented, dieting and eating-disorder-related material. The differences in what their algorithms delivered were considerably larger than the differences in what participants actively liked.
The study was not a representa…
EU in Practice | Association REDefine
At 11:43 p.m., you find a rechargeable nightlight for €4.99. It looks like a perfectly good Christmas gift. The photographs are lovely, the reviews reassuring, and a countdown timer warns that the offer will expire in twelve minutes. You add it to your basket, together with another item that qualifies you for free delivery and something else you didn’t know you needed until the marketplace suggested it.
Three minutes later, the shopping is done. You know when the parcel should arrive, but probably not who manufactured the nightlight, whether its battery was properly tested, or which company would take responsibility if it stopped working. None of this means the product is dangerous or badly made. It might be excellent value. The curious thing is how much we can learn about a product’s popularity and how little we sometimes learn about the product itself.
And Europe has decided to intervene, starting with the price.
Brussels has entered the checkout
Since 1 July 2026, the EU has applied a temporary €3 customs duty to qualifying low-value imports from outside the Union. On 1 November, a separate €2 handling fee came into effect. Both are imposed on the relevant commercial operators, although some of the costs may eventually reach consumers through higher prices.
There is an important detail here: the charges don’t simply add €5 to every parcel. They are calculated according to distinct customs items, generally based on product classification rather than the number of individual objects. Five identical T-shirts can count as one customs item, while a parcel containing a T-shirt and a watch contains two. The €3 duty applies to qualifying consignments valued at no more than €150, while the €2 handling fee has a broader scope. VAT is separate and has already applied to low-value commercial imports since 2021.
The reason for these changes is partly historical. For years, imports worth no more than €150 generally benefited from an exemption from ordinary customs duties. That made administrative sense when small international purchases were relatively uncommon. Collecting a tiny duty could cost more in administrative work than the duty was worth. It became more consequential when online marketplaces learned to organise millions of individual deliveries directly to European customers.
In 2025, almost 5.9 billion low-value e-commerce items entered the EU, according to the European Commission. A customs exemption intended to simplify small transactions had become an advantage for an enormous retail model. European businesses importing goods in bulk could face duties that competing direct-import arrangements avoided, even when their products originated from similar factories.
The case for correcting that imbalance is reasonable. So is the concern about affordability. A €3 charge represents a substantial proportion of a €5 product’s value, and large marketplaces may be better equipped to absorb new costs than smaller sellers. They can renegotiate shipping arrangements, consolidate stock in European warehouses or adjust their pricing across thousands of products. A small independent overseas seller has fewer options.
This is where the politics become interesting. European retailers certainly have an interest in the reform, but protecting European retail businesses isn’t automatically the same as protecting European manufacturing or consumers. Buying from a European shop doesn’t always mean buying something manufactured in Europe, produced under better labour conditions or designed to last longer. Sometimes the same factory supplies products to several different brands and retailers, with the final selling price reflecting marketing, distribution and commercial margins rather than fundamental differences in production.
The new charges can make customs treatment more consistent, which is a legitimate goal. But they also redistribute costs and may alter which businesses can compete successfully. The question is whether European consumers will receive something meaningful in return, beyond the possibility of paying more for precisely the same goods.
The worrying thing isn’t always the price
Consider what happens when the nightlight arrives. You plug it in, switch it on and assume that a product legally available for purchase has met the relevant safety requirements. That is a reasonable expectation, and one that European regulation is supposed to protect.
Unfortunately, the evidence shows how difficult this becomes when billions of items cross borders.
During a coordinated European enforcement operation in 2025, authorities examined 20,040 toys and small electronic products sold through major online marketplaces. Some 64% failed compliance checks. These were targeted inspections, so the result can’t be applied to every product sold online. Nor does every compliance failure mean a product is physically dangerous. Some problems involve missing documentation, incorrect labels or inadequate warnings. Nevertheless, laboratory testing of selected items revealed real risks, including electric shock, choking and exposure to harmful chemicals.
Those administrative failures matter more than they might initially appear. If a manufacturer can’t be identified, recalling a dangerous product becomes difficult. If the necessary technical information is unavailable, authorities may struggle to determine whether it was properly assessed. And if the same product appears under dozens of different names and sellers, removing one listing may accomplish very little.
The scale of the problem is extraordinary. In 2025, European customs authorities conducted approximately 388,000 controls, while handling billions of imported items. Physical inspection of everything is impossible, which means that effective supervision increasingly depends on identifying risky products through reliable information, cooperation between authorities and better digital systems.
The problem extends beyond individual sellers. In May 2026, the European Commission fined Temu €200 million for failing to adequately assess systemic risks involving illegal products on its marketplace. In October, the Commission accepted a corrective action plan, whose implementation remains subject to supervision and independent audit.
One interesting part of that plan concerns recommendation systems. Temu must examine how its algorithms and promotional mechanisms may contribute to the spread of illegal products, rather than assessing risks only in general terms. This goes to the heart of a much larger issue in digital commerce.
Online marketplaces can predict which products we might want next, test which photographs generate purchases and optimise the precise moment to offer a discount. Yet identifying dangerous products and establishing reliable accountability can remain significantly more difficult. The technology exists to sell products at speed and scale; the question is whether the systems responsible for preventing harm are receiving similar attention.
This is where Europe’s customs reforms become more interesting than their price tags suggest. From November 2026, product identifiers became mandatory in relevant customs declarations. Better identification could help authorities connect imported goods with safety information, detect repeated violations and establish which businesses are responsible when something goes wrong. The wider customs reform also places greater responsibility on the commercial operators arranging imports.
These changes are potentially more important than either fee. If authorities can trace unsafe goods more reliably, prevent repeat offenders from simply reappearing under different listings and intervene before problematic products reach consumers, the reform could deliver substantial benefits.
But better information isn’t the same as better enforcement. A product identifier is useful only if it leads to accurate records, effective risk assessment and action when something suspicious is discov…
On 3 November 2026, as Americans vote in the midterm elections, we are examining a question that reaches far beyond the United States: How do political campaigns turn fear, anger and hope into certainty, and what happens when that certainty outruns the evidence?
In 1964, Lyndon Johnson’s presidential campaign showed a little girl picking petals from a daisy, followed by a nuclear explosion. Twenty years later, Ronald Reagan’s Morning in America offered weddings, new homes, sunlight and optimism.
Fear and hope. Two extraordinarily different emotions, deployed with extraordinary precision.
Neither emotion is inherently illegitimate. Politics concerns our security, rights, livelihoods and future. People have good reasons to feel frightened, angry, hopeful or disappointed. The problem begins when emotional certainty starts doing the work of evidence.
In The Feeling Behind the Vote, our new REDefine video essay, we explore the psychology of political persuasion, a revealing 2026 Italian experiment involving 12,000 participants, the distance between memorable political promises and workable solutions, misleading claims about electoral procedures, and the information systems that can reward antagonism more readily than understanding.
We also revisit the Brexit campaign’s famous £350 million NHS promise, because a political promise can travel considerably further than its implementation plan. The latter, admittedly, wouldn’t fit on the side of a bus.
But this is not another argument that voters should simply calm down and become more rational. Emotional intelligence isn’t a political ideology, and it doesn’t tell anybody which candidate or party to support. It helps us recognise what we feel, examine what we are being told, and remain open to evidence that complicates our preferred conclusions.
The video concludes with REDefine’s 120-Second Political Judgment Check: four practical questions for examining an emotionally charged message before accepting, repeating or acting upon its claims.
We chose Election Day deliberately. The immediate context is American, but the questions belong to every democracy. And they will remain relevant long after the votes have been counted.
Go deeper: Read the full article
Our accompanying essay, The Feeling Behind the Vote, examines the research, historical cases and wider responsibilities of campaigns, media institutions and digital platforms in greater depth.
Historical footage — sources and acknowledgements
This video includes short excerpts from two historical US presidential campaign advertisements, presented for historical examination, criticism and educational commentary.
1. “Peace, Little Girl” (”Daisy”), 1964
Lyndon B. Johnson presidential campaign, Democratic National Committee. Produced by Doyle Dane Bernbach, with creative contributions from Tony Schwartz and the agency’s team.
Historical source: Museum of the Moving Image, The Living Room Candidate.
https://www.livingroomcandidate.org/commercials/1964/peace-little-girl-daisy
Archival video reference: Lyndon Baines Johnson Presidential Library.
2. “Prouder, Stronger, Better” (”Morning in America”), 1984
Ronald Reagan–George H. W. Bush re-election campaign. Created by the Tuesday Team, with Hal Riney.
Historical source: Museum of the Moving Image, The Living Room Candidate.
https://www.livingroomcandidate.org/commercials/1984/prouder-stronger-better
Editorial note: The excerpts are included as objects of historical analysis and criticism, not as political endorsements. Original footage and archival rights remain with their respective rights holders. Source attribution does not imply permission, sponsorship or endorsement.
Most of us are great at using digital systems. We log in, upload, approve, search, automate and increasingly ask AI to do parts of the work for us. But knowing how to operate a system is not the same as understanding the system we are operating.
This explainer from The Code Beneath the Floorboards looks at system literacy: the ability to see beyond the interface and understand the rules, data, dependencies, decisions and forms of power underneath it. That matters for citizens trying to access services, for educators working inside digital platforms, and for organisations that may depend on systems nobody inside the institution can fully explain anymore.
As AI makes technical execution easier, the gap between what we can do and what we can understand may become even wider. The question is no longer whether everyone should learn to code. It is whether enough of us understand the systems shaping our work and public life well enough to question them, change them and remain accountable for what they do.
The full REDefine longform goes much deeper into institutional intelligibility, epistemic dependency, decision lineage, knowledge concentration, shadow architecture and the organisational consequences of losing understanding while the technology itself keeps working.
This video is part of The Code Beneath the Floorboards, REDefine’s series exploring the invisible digital infrastructure underneath ordinary life, institutions and civic power.
If you opened a customer-service chat this morning and were told that you were speaking to AI, that notice sits inside a rule that already applies. If an employer buys an AI system to screen job applications today, the full set of AI Act requirements for that kind of high-risk system will not apply until December 2027. If the employer interprets that as permission to build first and think about governance sometime next year, however, the timetable has been badly understood.
Both situations belong to the same law, but they sit at very different points in its implementation timetable.
The EU AI Act has become harder to explain because it no longer has a useful single answer to the question “when does it apply?” The Regulation entered into force in 2024, some rules began applying in February 2025, obligations for providers of general-purpose AI models followed in August 2025, most of the Act became applicable in August 2026, and the 2026 Digital Omnibus then moved the detailed high-risk regime further into 2027 and 2028. Meanwhile, the Commission has published transparency guidance, enforcement has started, complaint mechanisms are operating, new prohibited practices are approaching their application date, and the first harmonised AI Act standard appeared in the Official Journal last week.
A four-colour risk pyramid can still tell us something about the architecture of the law. It can’t tell a citizen what should happen when an AI system is used around them today, whether a company developing a system in 2026 can safely ignore requirements that apply in 2027, or why a European standard with the deeply unromantic name EN 18286:2026 suddenly matters.
So we are going to leave the pyramid alone for a while.
The more useful questions in October 2026 are what the Act already requires, what remains in transition, what people should notice when AI is used around them, and what organisations should be building now even when the legal deadline is still more than a year away.
First, “provider”, “deployer” and “user” are not interchangeable
Before looking at the dates, one piece of terminology is worth clearing up because it causes a surprising amount of confusion.
Under the AI Act, a provider is broadly the organisation or person that develops an AI system or has it developed and places it on the market or puts it into service under its own name or trademark. A deployer is an organisation or person using an AI system under its authority. Personal, non-professional use is excluded from the deployer definition.
That means an employee casually using an AI assistant, a company integrating an AI tool into recruitment, and the company that developed that tool do not automatically occupy the same legal position. Nor does a citizen asking a chatbot a question become a “deployer” simply by pressing Enter.
The distinction matters because the Act distributes responsibility across the AI value chain. Providers may have obligations concerning how systems are designed, documented, tested and supplied. Deployers can have their own responsibilities concerning how systems are used, monitored and overseen in practice. In other words, buying an AI system does not magically transfer every governance question back to the vendor.
Anyone who has ever bought enterprise software may recognise the general principle.
The timetable now has several different clocks running
The first substantial AI Act rules began applying on 2 February 2025. That included the original prohibited AI practices and the AI-literacy provision. Article 4 has since been amended by the 2026 Digital Omnibus, but it still requires providers and deployers to take measures supporting the development of AI literacy among staff and others operating AI systems on their behalf. What changed is that organisations are no longer required to guarantee some undefined “sufficient” level of literacy for every individual. The obligation is contextual: technical knowledge, experience, training, the way the system is being used and the people affected by that use all matter.
That distinction is worth making because two equally unhelpful interpretations have circulated around AI literacy. One turns it into a universal certification exercise in which everyone needs the same course. The other assumes that the Omnibus quietly removed the obligation. It did neither.
From 2 August 2025, obligations for providers of general-purpose AI models also began applying. Providers of those models have duties concerning technical documentation, information for downstream system providers, compliance with EU copyright rules and publication of information about training content. Providers of the most advanced models presenting systemic risk face additional risk-management, incident-reporting and cybersecurity requirements. Models already on the market before August 2025 have a longer transition period, until August 2027.
For an ordinary organisation using a general-purpose model, that doesn’t mean it suddenly becomes the provider of the underlying model. It does mean that the AI Act is already regulating parts of the infrastructure on which many familiar AI products depend. The statement that “the AI Act doesn’t really start until 2027” is therefore simply wrong.
Then came 2 August 2026, when the general application date arrived. Enforcement powers became operational for provisions already applicable, and Article 50’s transparency obligations began affecting the AI systems and synthetic content people are much more likely to encounter directly.
The next important date is 2 December 2026, when the new prohibition introduced by the Digital Omnibus on AI systems generating non-consensual sexually explicit or intimate material and child sexual abuse material begins applying. The same date is also the end of the limited transitional period for the Article 50 machine-readable marking requirement for generative AI systems that were already placed on the market before August 2026.
Then we reach the dates that have generated most of the confusion: 2 December 2027 for many Annex III high-risk AI systems, including systems in areas such as employment, education, essential services, migration and certain biometric uses, and 2 August 2028 for high-risk AI embedded in products covered by the legislation listed in Annex I.
Those are not dates on which the rest of the AI Act suddenly becomes real. They are application dates for specific parts of the high-risk regime.
That difference matters enormously.
What should a citizen actually notice today?
The most visible part of the Act in ordinary life is now transparency.
Article 50 requires providers of certain interactive AI systems to make people aware that they are interacting with AI, unless that is already obvious from the circumstances. Providers of systems generating or manipulating synthetic content also have obligations to make those outputs detectable through machine-readable marking. Deployers have separate disclosure obligations for uses including emotion recognition, biometric categorisation, deepfakes and certain AI-generated or manipulated texts dealing with matters of public interest.
These are related duties, but they are not the same thing.
The visible notice telling you that the customer-service agent is artificial is there for a different reason from the machine-readable information embedded in synthetic content. One is about making the nature of an interaction clear to a person. The other helps technical systems identify content as AI-generated or manipulated.
Deepfakes bring another layer. Where AI has generated or manipulated image, audio or video content in a way that could falsely appear authentic, the deployer may need to disclose that artificial origin. Artistic, fictional, satirical and similar works are treated more flexibly because transparency has to be provided without destroying the work itself. European law has, thankfully, not yet decided that every science-fiction film requires a warning across the actor’s…
We often talk about technology as if it simply “gets old.” A laptop stops receiving updates. A battery becomes impossible to replace. A sensor still works, but the software around it no longer does. A small failure suddenly decides the fate of an entire device.
This short explainer looks at the hidden systems that determine whether technology can actually remain useful: repairability, spare parts, software support, data access, maintenance knowledge and the increasingly blurred line between owning a device and controlling its future.
The full Code Beneath the Floorboards article goes much deeper into the idea of maintenance sovereignty — how organisations can identify support-horizon mismatch, lifecycle choke points, repair-pathway friction and supplier dependency before a small failure becomes an expensive replacement decision.
The paid edition also includes the REDefine Technology Lifecycle & Repairability Audit, our connected professional workbook for turning those hidden dependencies into practical evidence, actions and procurement requirements.
This video is part of The Code Beneath the Floorboards, REDefine’s series exploring the invisible systems underneath everyday technology, institutions and public life.
Someone asks whether you are over 18. You hand over an identity card.
They now know much more than whether you are over 18.
Your full name is there. Your exact date of birth. Your photograph. Your nationality. A document number. Depending on what you handed over, perhaps several other pieces of information that have nothing whatsoever to do with the question.
We are so used to this arrangement that it barely registers as odd. Physical identity documents come as bundles, and the bundle is what we show. You cannot peel your age off a passport, hand over that one fact and keep the rest in your pocket.
Digital identity doesn’t have to work that way.
That is what makes the European Digital Identity Wallet more interesting than the phrase “digital wallet” suggests. Europe is not only trying to move familiar documents onto phones. It is trying to make identity more precise: to let us prove the particular thing that matters in a particular interaction without automatically handing over everything else.
A system designed around that principle begins with a much better question than Who are you?
It asks: What do I actually need to know about you?
Your identity doesn’t have to travel as one big package
The European Digital Identity Wallet will allow people to hold and use identity information and official credentials digitally, including things such as driving licences, educational qualifications and other attestations. Member States are required to provide wallets capable of working across the European Union.
But putting a driving licence on a phone is the least interesting part of the idea.
The bigger change is selective disclosure.
The European framework is designed so that a person can disclose only the attributes required for a particular transaction. Where full identification is unnecessary, the system can also support pseudonyms and privacy-preserving methods that confirm a fact without exposing all the data underneath it.
In practical terms, that means an interaction could look like this: a service needs to know whether you are over 18, so it receives confirmation that you are over 18. It doesn’t automatically receive your date of birth. An institution needs to know that you hold a certain qualification, so it receives a verifiable credential confirming that qualification rather than a miscellaneous bundle of identity documents assembled around it.
That is a significant change in the logic of identification.
For years, European data-protection law has told organisations to collect only the personal information they actually need. The principle is called data minimisation. It sounds perfectly sensible written into legislation. Everyday digital life has been somewhat less disciplined.
The wallet attempts to turn that principle into something a person can actually use.
Europe is effectively asking whether privacy can be built into the transaction itself, rather than added afterwards through another notice explaining what happened to all the data you already surrendered.
The most interesting thing about the wallet may therefore be what it allows us not to say.
Europe is not creating one giant identity database in Brussels
The name can also give the wrong impression in another direction.
The European Digital Identity Wallet is not one enormous EU identification system into which every European will be deposited. National identity systems remain national. Member States provide the wallets, directly or through recognised providers, while common European rules and standards are meant to make those different systems interoperable.
Europe already has plenty of digital identity systems. The difficulty begins when people cross borders.
A digital identity that works beautifully inside one country may suddenly become much less useful when you need to enrol at a university, prove a qualification, access a service or complete an administrative process somewhere else. At that point, digital Europe can still become remarkably fond of scanned documents, attachments, translations and certificates proving the authenticity of other certificates.
The Commission says that only a relatively small share of key public services currently allow cross-border authentication through existing electronic-identification arrangements.
So the problem is not simply that Europe needs digital identity.
It needs portable digital trust.
If a university in Portugal issues a digital qualification, an institution in another Member State needs to know what it is looking at. Who issued it? Was that organisation entitled to issue it? Has the credential been altered? Is it still valid? Does it belong to the person presenting it?
None of those questions is solved merely by converting a diploma into a prettier file.
This is why the Wallet is really part of a much larger trust infrastructure. The visible credential is only the front end. Underneath it are standards for issuance, verification, revocation, security, authentication and legal recognition across different countries and institutions.
The QR code is easy.
Getting twenty-seven administrative systems to agree on what the QR code means is rather more European.
“Over 18: yes” is enough
Age verification gives us the clearest example of how different this could feel in practice.
The European Commission has already been developing a privacy-preserving age-verification solution based on technical specifications compatible with the future Wallet. The basic idea is deliberately simple: an online service that genuinely needs to know whether somebody is over 18 can receive confirmation of that fact without also receiving the person’s name, exact date of birth or broader identity.
Instead of:
Jane Doe, born on X date, holder of document Y.
the service receives the information it actually needs:
Over 18: yes.
That separates two ideas that digital systems have spent years muddling together: verification and identification.
If the question is whether somebody meets an age threshold, identifying the person may be entirely unnecessary. A service needs an answer to one question, not an introduction.
Privacy-preserving cryptography can make that separation stronger. One of the technologies discussed in the European framework is the zero-knowledge proof, which allows a statement to be verified without exposing the underlying information used to establish it.
The name is considerably more dramatic than the concept needs to be. You can think of it as proving that you know something without revealing the thing itself, or proving that a condition is true without handing over the data behind it.
You don’t need to understand the mathematics to appreciate the change in philosophy.
For decades, digital systems have often treated more information as more certainty. If we know your name, birthday, address, document number and perhaps a little more for luck, surely we can trust the transaction.
The Wallet introduces a more disciplined possibility: sometimes a system can trust what you are entitled to do without knowing everything about who you are.
A digital diploma is useful only when somebody else can trust it
Age verification is neat because the proof is tiny. Cross-border credentials reveal the more complicated part of the project.
The Digital Credentials for Europe pilot, DC4EU, brought together organisations from 22 European countries to test wallet-based credentials in education and social security. Its use cases included educational diplomas, professional qualifications, European Health Insurance Card information and social-security documentation.
Consider the diploma.
You graduate from a university in Portugal and later need to use that qualification in Germany. The obvious digitalisation project would be to replace the paper diploma with a digital diploma.
Useful, certainly. Sufficient, no.
The institution receiving it still needs to know whether the university was authorised to issue the qualification, whether the credential has been tampered with, what the qualification actually represents, whether it remains va…
Schools teach far more than what appears on the timetable. Every platform, login, data trail and digital workflow also teaches students something about ownership, privacy, dependence, agency and who gets to shape the systems around them.
This short explainer looks at the hidden curriculum inside educational technology — from curriculum captivity and teacher attention to AI, data, interoperability and digital dependency.
The full article goes much deeper, with the professional section introducing the REDefine Educational Technology Dependency Model and our connected Open School & Educational Agency Audit for schools, universities, training providers and educational organisations.
Europe has 27 Member States, but one of the more intriguing ideas in its current competitiveness debate is a 28th regime: a European corporate framework that could sit alongside national systems rather than replacing them. In this episode, we unpack what EU Inc. could actually change for a founder trying to scale across Europe, where national law would still reappear, and what earlier European experiments can teach us about why some common regimes thrive while others quietly join the institutional graveyard. The deeper question is larger than company law: is Europe beginning to discover the institutional equivalent of an API — a way to make 27 different systems work together without first making them identical? Read the full article for the complete analysis.
The European Union is generally rather fond of categories.
There are Member States, candidate countries, potential candidates, European Economic Area countries, associated countries, strategic partners, third countries associated with individual programmes and third countries very specifically not associated with those programmes. Somewhere in Brussels, one suspects, there is a taxonomy for the taxonomy.
This week, however, Europe did something rather less characteristically tidy.
In her State of the Union address on 16 September, European Commission President Ursula von der Leyen proposed opening the door for Canada to become the EU’s first “associate member.” She placed the idea inside a much broader “Alliance for the Future”: a common prosperity and economic-security space covering advanced manufacturing, defence industries, the Arctic, energy, critical minerals, batteries, artificial intelligence, quantum technologies, cybersecurity and economic security.
There is just one small institutional detail.
The European Union does not currently have a general membership category called associate member.
The ordinary accession route under Article 49 of the Treaty on European Union is written for a “European State”. The EU does have powers under Article 217 of the Treaty on the Functioning of the European Union to conclude association agreements with third countries, involving reciprocal rights, obligations, common action and special procedures. But that is not an off-the-shelf status called “associate membership” into which Canada can simply be inserted after completing the appropriate form in triplicate.
A day after von der Leyen’s speech, Canadian Prime Minister Mark Carney welcomed the ambition and supplied considerably more detail about what Canada would like the relationship to contain. His proposals included closer cooperation in critical minerals, defence production, AI and computing capacity, energy, space and payments; smoother digital trade in non-agricultural goods and services; deeper financial integration; stronger youth mobility; Canadian membership of Erasmus+; and participation in the next generation of Horizon Europe.
So we have reached an unusually interesting point in European integration: the political ambition is visible, several pieces of the practical integration already exist, and the constitutional label remains under construction.
This is therefore not a story about Canada joining the European Union.
It is about a more interesting question.
How many parts of Europe can a country join without joining the Union itself?
First, Canada and Europe did not suddenly discover one another in Strasbourg
The easiest way to misread this week’s announcement is to imagine Canada and the EU standing on opposite sides of the Atlantic until geopolitical anxiety suddenly caused them to exchange telephone numbers.
The relationship is already substantial.
CETA, the Comprehensive Economic and Trade Agreement, has been provisionally applied since September 2017. EU–Canada trade in goods reached €81.5 billion in 2025, 76% higher than in 2016, while bilateral services trade reached €49 billion, up 91% over the same period. The EU is Canada’s second-largest trading partner.
CETA also supplies an early warning about the difference between political announcements and European institutional reality. Almost nine years into provisional application, only 17 of the EU’s 27 Member States have completed their national ratification processes. Belgium, Bulgaria, Cyprus, France, Greece, Hungary, Ireland, Italy, Poland and Slovenia still have not.
Before Europe constructs an entirely new relationship with Canada, then, there remains the small matter of fully ratifying the old one.
That is not merely a Brussels joke. It tells us something important about what follows. European international agreements can operate, deepen and produce real economic effects while parts of their legal architecture remain unresolved for years. The announcement of an “associate membership” relationship is therefore the beginning of an institutional process, not evidence that the process has already happened.
Meanwhile, cooperation has continued growing around CETA rather than waiting politely for every national parliament to finish.
In 2025, Canada and the EU established a Security and Defence Partnership. In June 2026, Canada became the first non-European country formally able to participate in procurement under SAFE, the EU’s Security Action for Europe instrument. That means Canadian companies and Canadian-origin products can participate in parts of common European defence procurement under the agreement.
A few days before the Strasbourg speeches, Canadian and European leaders were already pointing to the first Canadian contract under SAFE. This is rather more consequential than the diplomatic language might suggest: Canadian firms are no longer merely cooperating with European defence companies from the outside. They can participate inside a specific EU procurement framework.
And defence is not the only example.
Horizon Europe has already shown what partial integration looks like
In July 2024, Canada became associated with Pillar II of Horizon Europe, the enormous part of the EU research programme devoted to collaborative work on global challenges and industrial competitiveness.
Canadian universities, researchers and organisations can participate there on broadly equal terms with their European counterparts. They can join and lead consortia and receive Horizon Europe funding directly.
This is easy to overlook because research programmes don’t receive the theatrical lighting of summits and parliamentary speeches. But institutionally, it is quite revealing.
Canada did not need to become an EU Member State to become integrated into one of the EU’s major research systems.
That existing arrangement also helps clarify Carney’s proposal for Canadian participation in the next generation of Horizon. He is not proposing that Canada suddenly discover European research cooperation. He is proposing continuity and potentially deeper participation beyond the current framework.
This is where the emerging model becomes visible.
Instead of imagining one enormous gate marked EU MEMBERSHIP, imagine a building with many doors.
Research is one door. Defence procurement is another. Digital trade may become another. Education and youth mobility could become another. Critical-mineral supply chains, financial services and technological cooperation may produce still more.
Canada is already through some of them.
Erasmus+ may matter more than the strategic sectors in the spotlight
The headline sectors in the new alliance are exactly the ones one would expect in 2026: defence, AI, critical minerals, energy, quantum technologies and economic security.
But one of Carney’s proposals may ultimately tell us more about the durability of the relationship than any of those.
He explicitly proposed Canadian membership of Erasmus+, arguing that it could expand opportunities for Canadian and European students and strengthen long-term ties between the two societies.
Canada is not currently fully associated with Erasmus+. Under the current programme rules, EU Member States and a limited group of associated third countries participate fully across the programme. Countries such as Canada belong to the much larger category of third countries not associated with the programme, which can participate only in particular actions under specific conditions.
The distinction matters because Erasmus stopped being merely a university-exchange programme a long time ago.
It now connects universities, schools, vocational education, civil society organisations, adult learning, staff mobility, sport and large networks of institutional cooperation. Since 1987, Erasmus and its predecessor programmes have provided international learning opportunities to more than 16 million people. The Commission’s latest evaluation found that more than 6.2 million participated in learning mobilit…