Who Controls the Context?

1. Information Has Never Been Neutral

We should not fool ourselves into believing that an innocent information world existed before AI.

A newspaper decides which story appears on the front page. A teacher decides which texts his students read. Google sorts search results. A television channel decides who gets three minutes of airtime. Parents explain the world to their children from their own perspective. Academic journals decide which papers they publish.

People have always lived in curated realities.

What is new is not that information is selected.

What is new is the speed, individualization, and invisibility of that selection.

A newspaper prints the same front page for hundreds of thousands of people.

An AI could theoretically create a different front page for every single person.

And no one except the system itself needs to know how different those front pages are.

That substantially changes the question of power.

2. Let Us Take a Simple Question

A person asks:

“Should I quit my job?”

The question consists of six words.

But the possible answer depends on what else the system knows.

Perhaps it knows that the user has been dissatisfied for months.

Perhaps it knows his financial situation.

Perhaps it remembers that two weeks ago he mentioned a conflict with his boss.

Perhaps it knows that he has three children.

Perhaps it has inferred from earlier conversations that security is particularly important to him.

Perhaps it knows about job openings.

Perhaps it knows that his industry is currently cutting jobs.

Perhaps it knows only some of these things.

And it may remember something incorrectly.

The visible question remains identical.

The context changes the answer.

This is precisely why memory is becoming a central component of modern AI systems. A recent survey of memory architectures for AI agents describes memory not simply as a repository of past conversations, but as a technical process of storing, organizing, and selectively retrieving information. The word “selectively” is crucial. A system can never consider everything at once. It has to choose. [2]

And selection means weighting.

3. The Actual Answer Begins Before the Answer

When we speak with a powerful AI system today, several layers may determine what ultimately appears on the screen.

First, there is the trained model with the statistical relationships it has learned.

The provider’s rules may sit above that.

Then there are the specifications of the application in which the model is used.

Then comes the specific question.

Information from documents or databases may be added.

There may be stored information about the user.

The system may use search engines, company data, calendars, emails, or other tools.

Then it must decide which of these things are relevant to the specific question.

In the end, we see twenty sentences.

But we do not automatically see the many decisions that led to the creation of those particular twenty sentences.

When assessing human-AI systems, the American National Institute of Standards and Technology explicitly notes that complex human phenomena can lose necessary context when translated into mathematical models. NIST therefore calls, among other things, for clearly defined human roles and responsibilities and for decision-making processes to be considered throughout an AI system’s entire life cycle. [3]

That is an important point.

But the problem now runs just as strongly in the other direction.

Context can not only be lost.

Context can also be added.

And the user may notice neither.

4. In the Past, We Had to Manipulate What Someone Saw

Anyone who wanted to influence people needed access to their information channels.

Newspapers.

Television.

Posters.

Advertising.

Search engines.

Social networks.

The platform economy has already turned this into a highly developed business. Content is selected, sorted, and personalized.

Generative AI adds a new layer.

It does not merely select.

It formulates.

That is a fundamental difference.

A recommendation algorithm may decide which article we see next.

A generative AI can formulate a single answer from a hundred articles, tailored precisely to our question.

The original information space disappears behind a synthesis.

We no longer necessarily receive ten sources.

We receive an answer.

That is convenient.

And that is precisely why it is powerful.

5. Who Actually Decides What Is Missing?

I consider this question more important than many debates about hallucinations.

We may be able to detect false information.

It is much harder to detect information we have never seen.

Suppose we ask an AI for arguments in favor of a particular energy policy.

It gives us ten factually correct arguments.

None of them is invented.

But perhaps three decisive counterarguments are missing.

Does that make the answer false?

Not necessarily.

Is it complete?

No.

Can it still influence our decision?

Of course.

Manipulation therefore does not begin only with a lie.

It can begin with selection.

We know this from human communication. A lawyer does not have to lie to emphasize the facts most favorable to a client. Advertising does not have to claim that a product has no drawbacks. Often, it is enough simply not to mention those drawbacks.

AI can industrialize this form of selection.

And it could individualize it.

6. The Provider Is Not the Only Problem

It would be too simple to turn this into a story about evil technology corporations.

Power over context is distributed.

The model developer decides on training and fundamental behavior.

The operator of an application may decide on additional rules and sources.

Companies determine which of their own data are connected.

Search systems influence which external information is found.

Creators and media outlets decide which information is publicly available in the first place.

Governments set legal limits.

And finally, users themselves influence the context through their questions, data, and previous decisions.

The interesting question is therefore not:

Who controls the AI?

The more interesting question is:

Who controls which part of the context?

Because in modern AI systems, that control will probably never lie entirely with a single party.

7. Memory Makes It More Personal

An AI without memory knows my current question.

An AI with memory may know my history.

That can be enormously useful.

A person with a rare disease does not have to explain their entire medical history again in every conversation.

An entrepreneur does not have to describe the structure of their company every time.

A student can work for months with a tutor who knows her learning progress.

A woman in Guatemala could speak in her native language with a system that takes her family and local context into account, instead of repeatedly giving her answers that seem written for people in California.

Context can make AI more human and more helpful.

But the same mechanism has another side.

The better a system knows me, the better it can estimate which information will work on me.

Not necessarily in order to manipulate me.

It is enough that some optimization objective exists.

Time spent.

Likelihood of purchase.

Consent.

Signing a contract.

Political support.

Health behavior.

Willingness to donate.

Personalization then very quickly becomes influence.

8. The System’s Objective May Be the Most Important Information

We often focus on data.

What data does an AI have about me?

But at least as important is:

What is it supposed to achieve with those data?

A navigation system knows my location. In itself, that is neither good nor bad.

If its objective is to get me home as quickly as possible, it uses that information differently than if its operator is paid to route me past certain stores.

The data are identical.

The optimization objective is different.

With generative AI, this difference is harder to see because the system speaks with us.

Language creates the impression of a counterpart.

A counterpart appears to have reasons.

A technical system, by contrast, has objectives built into it through development, training, product design, and the context of use.

Users usually do not see this objective structure.

That is precisely why transparency about the fact that we are speaking with an AI is not enough.

Increasingly, we would need to know whom this system actually works for.

9. The Personal Assistant Is a Political Idea

We like to treat personal AI assistants as the next evolutionary stage of the smartphone.

That underestimates their significance.

One day, a personal assistant could decide which news to summarize for us in the morning.

It could prioritize our emails.

It could suggest appointments.

It could compare offers.

It could assess which insurance policies make sense.

It could sort medical information.

It could decide which job applications appear interesting.

It could tell us which political statements are probably true.

It could support our children’s learning.

Each of these functions is useful in itself.

Together, they create something different.

An intermediary between people and reality.

And intermediaries have power.

The question of whom a personal AI assistant serves may therefore become one of the most important product questions of the coming years.

The provider?

The advertiser?

The employer?

The government?

The user?

Or some combination of them?

10. European Regulation Recognizes Part of the Problem

The European AI Act prohibits certain manipulative or deceptive AI practices when they materially impair people’s ability to make informed decisions and thereby cause, or are reasonably likely to cause, significant harm. The law also classifies certain systems intended to influence election or referendum outcomes or voting behavior as high-risk AI systems. [4]

That makes sense.

But legally prohibited manipulation is only the outermost boundary of a much larger field.

Between neutral information and illegal manipulation lies an enormous realm of legitimate influence.

An insurance adviser tries to persuade.

A politician tries to persuade.

A doctor may try to persuade a patient to undergo treatment.

Parents try to influence their children.

Media outlets select topics.

Companies present products in a favorable light.

Not every form of influence is manipulation.

The problem with AI is therefore not to prevent influence completely.

That would be neither possible nor sensible.

The challenge is to make influence visible and controllable.

11. We Need a New Conception of Freedom of Information

Freedom of information was long understood as access.

Can I read a newspaper?

Can I open a website?

Can I buy a book?

Can I ask a public authority for documents?

In the age of generative AI, access may no longer be enough.

Because in theory we could have access to billions of documents and, in practice, still see only what a system selects for us.

A second freedom then becomes important:

The freedom to influence the selection mechanisms.

I want to be able to know which sources were used.

I want to be able to request a different perspective.

I want to be able to say: Show me the strongest counterarguments.

I want to be able to correct stored assumptions about me.

I want to be able to distinguish between fact, interpretation, and recommendation.

I want to know whether commercial interests play a role.

I may even want to be able to check different AI systems with different perspectives against one another.

Not because a machine is inherently less trustworthy than a person.

But because trust without the possibility of dissent eventually becomes nothing more than dependency.

12. Context Must Not Become the Machine’s Property

This leads to an important design question.

Who actually owns my digital memory?

If, over the years, an AI comes to understand how I work, think, decide, and communicate, something valuable emerges.

Not only for me.

Economically as well.

My context could one day be more valuable than access to a particular language model.

Models can be replaced.

My history cannot.

A provider that owns my context therefore gains enormous power to lock me in.

Switching to another system would mean starting from zero again.

That would be a new form of lock-in.

Not through files or software formats.

Through knowledge about me.

A serious discussion about AI sovereignty would therefore eventually also have to encompass the portability of personal context.

My digital memory should not automatically belong to the provider that technically stores it.

13. The Great Danger Is Not Necessarily a False World

Harari’s warnings are sometimes read as suggesting that machines will one day create a gigantic artificial reality and completely control people.

That is conceivable, but it is not even necessary for the current discussion.

The much more mundane version is enough.

Every day, we receive answers that are mostly reasonable.

Our AI knows us better and better.

It saves us time.

It increasingly makes a good initial selection.

So we check a little less.

Then a little less again.

Not because we are stupid.

But because the system works most of the time.

That is exactly how dependency on infrastructure emerges.

No one personally checks the water quality every morning before turning on the tap.

We trust the system.

That works because rules, measurements, responsibilities, and public oversight have developed around drinking water.

With AI context, we are still at a very early stage of this development.

We are already building the water pipes.

But we have not yet settled the question of who actually measures the quality.

14. Context Can Also Mean Liberation

For all the criticism, we should not forget the opposite direction.

A universal AI that gives every person the same answer would not automatically be fairer.

People live in different realities.

A woman farmer in Guatemala needs different information from an investment banker in Frankfurt.

A blind person needs different forms of presentation from a sighted person.

A beginner needs different explanations from an expert.

A person in crisis may need different communication from someone seeking a sober expert analysis.

Context is therefore not the problem.

Context is a prerequisite for good communication.

The question is who controls it.

Personalization can strengthen autonomy when the person controls it.

It can weaken autonomy when others control it.

This is the same technical mechanism with two entirely different political meanings.

15. Conclusion: So Who Controls the Context?

The honest answer is:

No one alone.

And that is exactly what makes the question difficult.

We will not meaningfully understand power over AI if we look only at models.

We have to look at the entire information chain.

At providers.

At data.

At memory.

At search systems.

At rules.

At economic interests.

At user interfaces.

At the decision about which information enters an answer at all.

And at ourselves.

Because we voluntarily give these systems more and more context, as better answers genuinely require better context.

The decisive conflict of the coming years may therefore not be between humans and machines.

It may be between two conceptions of AI.

In one, the machine knows us better and better and increasingly decides for itself which information is relevant to us.

In the other, it also knows us well, but we can see, correct, and influence how that knowledge is used.

Technically, the two systems may look remarkably similar.

Politically, they are fundamentally different.

It is therefore not enough to ask whether an AI tells the truth.

We need to know which reality it draws on to assemble its truth.

And who decided what that reality contains.