When we entrust information, analysis and decision-making processes to artificial intelligence, who are we really entrusting control to?
Imagine a Security Manager who has to prepare a risk assessment for a major event. There is little time and an artificial intelligence tool that can help. He uploads the floor plan, describes the vulnerabilities identified during the site survey, and adds a few details about the security systems in place. Within seconds, he gets a neat document, with scenarios, critical issues and mitigation measures.
An excellent result.
Or perhaps not.
Because before assessing the quality of the analysis, we should ask ourselves what happened to the information we have just shared. Where was it processed? Who can access it? How long will it be stored? Could it be used for other purposes? And who decided that transferring it to that system was acceptable?
These are not technology questions.
They are security questions, and they apply to any artificial intelligence, regardless of its nationality.
The problem is not China
The development of Chinese artificial intelligence has attracted increasing attention. The arrival of competitive models, some of them also released with open weights, has expanded the range of choices available to companies and professionals, and the discussion quickly moved from technology to geopolitics: are Chinese systems reliable? Are American ones safer? Should Europe build everything on its own?
These are legitimate questions, but they risk taking us off course.
A system developed in China is not automatically unsafe, and one developed in the United States or Europe is not automatically reliable.
Its origin may matter in a risk assessment, but it cannot replace the assessment itself.
For a Security Manager, the question is not whether an artificial intelligence system is Chinese, American or European.
The question is what risks it introduces into the system he or she is responsible for protecting.
And to answer that, we need to look beyond the manufacturer’s name.
One model, many dependencies
Behind a simple chat window there may be several different parties: the organisation that developed the model, the company distributing it, the operator managing the servers, the cloud provider, and the developers building the applications that integrate it into business processes.
Then there are data providers, software libraries, authentication systems and the tools through which AI can access other information.
It is a supply chain.
And like any supply chain, it introduces dependencies.
The ENISA Threat Landscape 2026, published on 22 September, confirms that attacks continue to target points of dependency, including supply chains and third-party providers, and notes that integrating artificial intelligence systems into business environments expands the attack surface.
The issue, therefore, is not only whether the model itself is reliable.
It is whether the entire system through which that model reaches us is reliable.
An excellent model may sit behind a vulnerable application.
A service with strong data-protection measures may rely on infrastructure we do not control.
Or it may work perfectly and suddenly become unavailable.
That too is a security risk, because operational continuity does not depend on the provider’s intentions. It depends on the system’s ability to keep operating.
Where does our data go?
Comparing Chinese and Western AI systems is interesting not because one must necessarily be preferred to the other, but because it shows how different the conditions of use can be.
Take one concrete example.
In its privacy policy updated on 17 April 2026, DeepSeek states that it collects, processes and stores personal data directly in the People’s Republic of China, and describes the possible use of inputs to improve its services and train its models, together with the rights and objection mechanisms available to users.
The policy applies to DeepSeek’s own services — its apps, websites and related services — and not to third-party applications built on its platform, for which the third-party developer is responsible.
That is documented information.
It does not demonstrate that the service conducts intelligence activities, nor that the data is being misused.
It does mean that a European organisation must assess the implications of that processing, including international transfers of personal data.
There is also an important distinction to make: a privacy policy concerns personal data, whereas floor plans, vulnerabilities and risk assessments are primarily confidential information.
For those, terms of use, contractual arrangements and confidentiality obligations matter.
The privacy policy is still informative because it states how the provider may use what it receives, including for model training.
Western AI services also have terms that need to be examined.
OpenAI, for example, distinguishes between business services — such as ChatGPT Business, ChatGPT Enterprise and the API — and individual services. For the former, it states that inputs and outputs are not used by default to train its models.
The same documentation makes clear, however, that training data can include information from consumer versions of ChatGPT and other individual services.
The comparison, it should be said, is not between equivalent products: we are comparing one provider’s general privacy policy with another provider’s business-focused documentation.
That is precisely the point.
Conditions vary from one provider to another, and even between services offered by the same provider.
Neither set of conditions automatically makes a business use safe or unsafe.
This is why evaluating a technology on the basis of the manufacturer’s nationality alone is misguided.
And we should not confuse where data is stored with where it may be processed, or with the jurisdictions to which the organisations handling it may be subject.
Digital sovereignty is not only about knowing where the servers are. It is knowing who can do what with our information.
Intelligence risk is not assumed. It is assessed.
Whenever AI and national security are discussed, the word intelligence inevitably appears.
But we need to distinguish facts from deductions and hypotheses.
It is a fact that digital service providers operate under different legal systems, with different obligations concerning data processing, cooperation with authorities and information security.
One documented example concerns transfers of personal data outside the European Economic Area.
The European Commission has adopted an adequacy decision for the United States, limited to commercial organisations participating in the EU-US Data Privacy Framework. There is no equivalent adequacy decision for China.
Appropriate safeguards are therefore required, such as Standard Contractual Clauses, together with an assessment of the conditions in the destination country.
The European Data Protection Board also points out that such safeguards need to take into account the actual level of protection provided by the legal system of the destination country.
Adequacy decisions themselves, moreover, need to be monitored by organisations exporting data, because they can be reviewed or invalidated.
It is a reasonable deduction that these differences affect the level of risk, especially when confidential or strategic information is involved.
It is, however, a hypothesis — requiring specific evidence — to claim that a particular system is being used to collect information on behalf of a government.
Confusing these three levels means conducting poor analysis.
And poor analysis produces poor decisions.
The issue is not to attribute intentions to a country.
It is to understand the operational and legal consequences of a technological choice.
Digital sovereignty does not mean isolation
At this point, the obvious solution might seem to be using only technologies developed and managed within our own borders.
That would be too quick a conclusion.
A model developed abroad can be deployed on infrastructure controlled by the organisation using it, without necessarily sending operational data back to the original service.
Conversely, a European model may be offered through platforms that introduce dependencies on non-European infrastructure.
And local execution does not eliminate every risk: software security, the provenance of updates, vulnerabilities and model reliability still matter.
The distinction that matters is between where the technology comes from and who controls how it is used.
An organisation should be able to choose which dependencies to accept, what information to share, and which activities to keep under its own control.
It should also know how to replace a provider, recover its data or continue operating if a service suddenly disappeared.
From a risk-management perspective, digital sovereignty does not necessarily mean technological self-sufficiency.
It means the ability not to lose control over our own decisions.
Who decides how much we can trust it?
Let us return to our Security Manager.
Before using an AI tool, he or she should know what information can be shared, where it will be processed, who may access it, what safeguards the provider offers, and whether the organisation has an established policy governing the use of artificial intelligence.
Above all:
Who has the authority to authorise its use in the most sensitive processes?
Because this brings us back to a problem we know well:
Technology arrives before the rules designed to govern it.
An employee discovers a useful tool.
A manager starts using it to speed up a task.
A department begins integrating it into its processes.
And gradually, the organisation becomes dependent on something it never formally assessed.
It is not necessarily a bad technological choice.
It is a governance decision that nobody realised they were making.
Trust is not a security measure
AI models will become more powerful and less expensive, and an increasing share of analysis and decision support will be entrusted to these systems.
That can be a great opportunity.
But the more important a technology becomes to an organisation’s operations, the more we need to understand its dependencies and know what we are paying for in terms of control, confidentiality, operational continuity and decision-making autonomy.
And we need to apply the same level of scrutiny to every provider.
Chinese.
American.
European.
Threats may differ. Regulations may differ. Levels of exposure may differ.
The risk-assessment method should not change with the flag.
Security is not about choosing whom to trust.
It is about building systems that do not depend solely on trust.
In the end, the question is not which artificial intelligence we are using.
It is another one.
If that system stopped working tomorrow, changed its terms, or produced a wrong assessment, would we still be capable of making an independent decision?
If the answer is yes, we are probably using the technology well.
If the answer is no, perhaps we did not simply adopt a tool.
We handed over part of the control.
And in security, once control is lost, it can be very difficult to regain.
Essential sources
- ENISA — Threat Landscape 2026, 22 September 2026. Digital infrastructure threats, technological dependencies and supply chains.
- DeepSeek — Privacy Policy, 17 April 2026. Processing, storage and use of personal data.
- OpenAI — Enterprise Privacy. Data handling in business services.
- European Data Protection Board — International Data Transfers. Adequacy decisions and safeguards for international transfers of personal data.
- European Commission — AI Act, enforcement framework. Governance, transparency and regulatory obligations.
The scenarios described are hypothetical. The observations on technological dependencies and organisational risks are security-management assessments and do not constitute allegations of unlawful conduct by any specific provider or country.
Leave a Reply