A teenager can understand perfectly well that TikTok, YouTube or another recommendation system learns from behaviour. She may know that pausing, replaying, searching and scrolling all contribute information about what the platform thinks she wants to see next. That knowledge matters. But it does not necessarily allow her to inspect the profile being constructed, correct a mistaken inference or prevent the system from repeatedly testing that inference against her attention.
An organisation can perform an equally sophisticated exercise. It can map its cloud services, identify the APIs connecting different tools, document suppliers and even understand which external systems are essential to everyday operations. Yet if one provider changes its terms, withdraws access or makes migration prohibitively difficult, that knowledge does not by itself restore control.
These examples reveal a distinction that is becoming increasingly important as Europe invests in AI literacy, digital resilience and technological sovereignty.
Understanding a technological system is not the same as having meaningful agency within it.
That sounds obvious when stated plainly. In practice, however, we regularly collapse the two.
We teach people how algorithms work and call the result empowerment. We produce transparency documentation and assume that informed users can protect themselves. Organisations map dependencies and feel more sovereign because the architecture has become visible. Public consultations invite participation and are treated as evidence that technological decisions have become democratic.
Sometimes those measures genuinely redistribute power. Sometimes they merely make an unequal relationship easier to describe.
The difference matters.
What should technological understanding actually allow someone to do?
AI literacy is often discussed as a problem of knowledge: people need to understand what artificial intelligence can do, where it can fail, how data influences systems, why outputs can be biased or inaccurate, and when human judgement remains essential.
All of that belongs in AI education.
But if literacy is supposed to produce something more than competent consumption, another question follows: what becomes possible once someone understands the system?
Consider a young person using a generative AI assistant. She may understand hallucinations, know not to share sensitive personal information and recognise that a fluent answer is not necessarily a correct one. That is considerably better than uncritical use.
Yet the surrounding relationship may remain largely unchanged. She may still have limited knowledge of what happens to her interactions, how the system is optimised, what behavioural signals are retained, how the model is designed to respond to emotional language or which commercial incentives shape the experience.
The problem here is not that literacy has failed. The problem is that we have expected literacy to solve something that partly belongs to governance, product design and institutional responsibility.
A person cannot educate themselves out of every asymmetry of power.
That observation becomes particularly important when responsibility starts migrating downward. If a system produces harmful effects, it is tempting to ask whether the user understood the risks, configured the settings correctly or exercised sufficient judgement. Those questions can be legitimate, but they can also obscure an earlier one: what degree of control did the user actually possess?
Three conditions make technological agency more meaningful
A useful way to think about this is to distinguish three qualities that are often bundled together under words such as transparency, control or empowerment: legibility, contestability and reversibility.
Legibility means that a person or organisation can understand enough of the system to recognise what is happening. Not every line of code needs to be visible, but the important relationships should be comprehensible. What information is being collected? What is being inferred? Which external providers are involved? What is the system optimising? Where are consequential decisions being made?
Without legibility, people cannot even locate the problem.
But visibility alone is weak if nothing can be challenged. Contestability means that when a system makes a consequential inference or decision, there is some meaningful way to question it, correct it or introduce human judgement.
A recommendation that misunderstands a teenager’s momentary curiosity as a persistent preference is a small example. An automated employment, education, financial or public-service decision raises much higher stakes. In both cases, a system can be perfectly transparent about the fact that automation occurred while still leaving the affected person with no practical way to alter the result.
Then there is reversibility, which becomes especially important for organisations.
Can we change supplier? Can we export our data in a usable form? Can we disable an AI feature without destroying the wider workflow? Can we operate temporarily if an external service disappears? Can we undo an integration whose risks only became visible after implementation?
This is where digital sovereignty becomes considerably more concrete than political rhetoric about European technological independence.
An organisation does not need to build every component itself to retain agency. It does need to understand which dependencies are difficult to reverse and which actors have the practical ability to define the conditions under which it operates.
A dependency becomes strategically important when the other party can change the terms and you cannot realistically leave.
The most dangerous dependencies often look convenient
Digital dependence rarely arrives labelled as dependence.
It often begins as convenience.
A service saves time. An integration eliminates manual work. An AI tool makes drafting faster. A cloud platform solves infrastructure problems an organisation has no desire to manage internally. None of these decisions is inherently irresponsible; in many cases they are entirely rational.
The architecture becomes more interesting later.
One service relies on another provider for authentication. A supposedly portable workflow depends on a proprietary data format. An internal process begins using an external AI model through an API. Staff gradually organise their work around features that did not exist when the original procurement decision was made.
Each individual decision can still make sense while the combined system becomes increasingly difficult to leave.
This is why dependency mapping matters, but also why mapping is only the beginning. The useful question is not simply “What do we depend on?” It is “What happens if the conditions of that dependency change?”
That shift turns a technical inventory into a governance exercise.
It also changes the meaning of digital sovereignty. Sovereignty is not the absence of interdependence. Modern digital systems make that fantasy almost impossible. It is the preservation of enough knowledge, bargaining power and alternatives that dependence does not quietly become submission.
AI makes these questions harder because the system also interprets us
Artificial intelligence adds another layer to the problem because many contemporary systems do not simply wait for instructions. They interpret behaviour.
Recommendation systems infer interests. Fraud systems infer risk. Hiring systems infer suitability. Generative systems infer what kind of answer a person is seeking. Conversational systems increasingly respond not only to the literal content of a message but to tone, context and patterns in interaction.
This creates a peculiar situation.
We are investing heavily in helping people understand technological systems at precisely the moment those systems are becoming better at constructing models of the people using them.
That does not mean the machine “understands” a person in the human sense. But it can still produce classifications, predictions and responses that materially shape what happens next.
And this is where technological literacy becomes psychologically and politically interesting.
The systems we build contain assumptions about what human behaviour means. Attention becomes evidence of interest. Repetition becomes evidence of preference. Engagement becomes evidence of value. Speed becomes evidence of good service. Agreement can begin to resemble empathy.
Those assumptions do not originate in the machine. They originate in human choices about what should be measured, rewarded and optimised.
Once encoded into infrastructure, however, they can begin to look natural.
The real test of responsible technology is not whether people understand it
Transparency is important. AI literacy is important. Digital skills are important. None of them should be dismissed merely because they cannot solve every problem.
But we should be more precise about what they can accomplish.
If a user can understand a recommendation system but cannot correct what it infers, knowledge has not produced contestability.
If an organisation can identify a critical digital dependency but cannot realistically leave it, transparency has not produced reversibility.
If citizens can participate in technological governance but cannot tell whether their participation influenced the decision, consultation has not necessarily produced democratic agency.
A stronger approach to responsible technology therefore asks not only whether people can understand the system, but whether the relationship gives them meaningful room to act.
Can they see enough to understand what is happening?
Can they challenge consequential decisions?
Can they change direction when the relationship stops serving them?
Those questions are useful well beyond artificial intelligence. They apply to platforms, cloud infrastructure, data governance, educational technology, public digital services and almost every organisational technology stack.
They also expose something that the language of digital literacy can sometimes hide: the problem is not always that people know too little. Sometimes they understand perfectly well what is happening and still do not control the terms.
That is where the technical question becomes a political one.
Our latest REDefine weekly essay, Who Sets the Terms?, takes this argument further. It connects AI literacy with digital sovereignty and democratic AI governance, explores what happens when human choices become embedded in infrastructure, and asks an even more uncomfortable question: what happens when the systems we are trying to understand are simultaneously becoming better at interpreting us?
