Intelligence Data for AI? What the Pentagon Plans with LLMs – What Does This Mean for Swiss SMEs?

The news sounds like something out of a spy thriller: The Pentagon is reportedly planning to train AI models with classified military data. A highly sensitive development that extends far beyond military applications. As the US realigns its defence strategy, Swiss SMEs face a crucial question: What do such global developments mean for our local economy, our data, and our competitiveness?

It's a paradoxical situation. Precisely at a time when the acceptance of Artificial Intelligence in Switzerland is clearly on the rise – 45% of Swiss SMEs now consider AI an advantage for their business operations, an increase of 10 percentage points year-on-year – new, complex questions about data sovereignty and security are emerging. The opportunity that 60% of Swiss SMEs see in AI must be linked to a clear strategy for data sovereignty and compliance. Only then can the potential risks arising from the use of global AI models be effectively minimised.

The Pentagon's plans are a wake-up call. They highlight that data used to train AI models is of immense importance – not only for military intelligence but also for the integrity and security of every single company relying on AI. For Swiss SMEs, often characterised by precision, confidentiality, and quality, it is crucial to fully understand the implications of these developments and act proactively.

What specific security risks do Swiss SMEs face when using AI models potentially trained on sensitive data?

The risks are manifold, ranging from unintentional data exposure to manipulated information. When global AI models, potentially trained on sensitive, perhaps even classified, data, find their way into commercial applications, significant uncertainties arise for Swiss SMEs. It cannot be ruled out that such models contain subtle biases or even "memories" of their training data, which could then resurface in generated content or analyses.

A primary security risk is so-called "data leakage" through inference. Even if an SME does not directly input its own sensitive data into a public AI model, the model could, through clever queries (prompt engineering), reveal information it learned during its training from similar, but originally classified, sources. This affects not only military secrets but potentially also trade secrets, customer lists, or financial data, if such information was included in the training datasets or processed in similar contexts. The model's ability to distinguish between fact and fiction, as reports about the Pentagon suggest, is a central vulnerability here.

Furthermore, there is the danger of "hallucinations" or the spread of misinformation. If a model was trained on biased or unreliable data, it can generate false or misleading answers that appear plausible at first glance. For a Swiss SME using AI for market analysis, customer communication, or decision-making, this could lead to costly errors or significant reputational damage. The origin and quality of the training data are crucial for the reliability of AI outputs. It is a fallacy to believe that a model that "knows everything" always tells the truth.

Another often underestimated risk is dependency on external providers. If an SME relies heavily on a global AI model operated by a foreign provider, it relinquishes a portion of its data sovereignty and control over the technology used. Changes in terms of service, data processing policies, or even geopolitical tensions could directly impact business operations. The AI supply chain, from training data to algorithm, must be transparent and traceable to manage such risks. Trust in technology requires trust in its origin and operation.

How can Swiss SMEs leverage the benefits of AI without compromising data sovereignty and confidentiality?

The key lies in a conscious strategy focused on local control, transparency, and robust data protection practices. Swiss SMEs do not have to forgo the advantages of AI simply because global models pose risks. There are viable ways to achieve efficiency gains and innovation through AI without relinquishing control over their own, often highly sensitive, business data.

One of the most effective strategies is the use of private or on-premise hosted AI solutions. Instead of relying on large, publicly accessible Large Language Models (LLMs), SMEs can use smaller, specialised models trained on their own data. These models then run either on their own infrastructure or in a dedicated, secure Swiss cloud environment. This ensures that data remains physically in Switzerland, is subject to the Swiss Federal Act on Data Protection (FADP), and is not subject to the access rights of foreign authorities. Control over the data and infrastructure remains entirely with the company.

Another important step is the consistent anonymisation and pseudonymisation of data before it even enters an AI model. This applies even to private models. Personally Identifiable Information (PII) should be removed or replaced with placeholders as early as possible. This minimises the risk that sensitive data, even in the unlikely event of a security incident, can be linked to individuals. This requires careful data architecture and processes designed for data protection from the outset.

Furthermore, SMEs should scrutinise AI service providers and products carefully. Inquire about the hosting location, security certifications, data processing policy, and the possibility of training or customising your own models. A Swiss hosting partner committed to local laws and whose data centres are located in Switzerland offers significant additional security. Transparency about the origin of training data and the functioning of the algorithm is also crucial. Lukas Huber, founder of schnellstart.ai, repeatedly emphasises that focusing on Swiss solutions and local expertise is the best way to maintain data sovereignty.

What regulatory or ethical considerations are relevant for Swiss SMEs when implementing AI solutions potentially based on global or military datasets?

Compliance with the Swiss Federal Act on Data Protection (FADP) and adherence to ethical principles are non-negotiable. Regardless of the origin of the initial training data, AI solutions used in Switzerland must meet local legal and ethical requirements. The FADP is the central anchor point here.

The revised FADP, which came into effect in September 2023, imposes high demands on the protection of personal data. For Swiss SMEs, this means they must be able to account at all times for what data they process, for what purpose, and on what legal basis. When AI models are used that are potentially based on global or military datasets, and whose origin and processing methods are unclear, compliance with the FADP becomes extremely difficult, if not impossible. It becomes particularly critical if these models process or generate personal data that cannot be clearly traced back to a clean and transparent data origin. A Data Protection Impact Assessment (DPIA) is essential for high-risk AI applications.

Ethical considerations are equally important. AI systems trained on opaque or potentially biased data can perpetuate discrimination or produce unfair outcomes. For an SME using AI in recruitment, credit scoring, or customer outreach, this would have not only legal but also significant ethical and reputational consequences. The responsibility for the decisions an AI makes ultimately lies with the company that uses it. A "black box" AI, whose functioning and data basis are not transparent, is therefore ethically problematic.

The question of AI governance is therefore becoming increasingly pressing. Swiss SMEs must develop internal policies for handling AI that ensure transparency, traceability, and human oversight. This includes training employees in the responsible use of AI, defining clear responsibilities, and implementing mechanisms for reviewing and correcting AI decisions. It's about understanding AI as a tool that must be controlled and monitored, rather than as an autonomous black box. Regulatory requirements are likely to increase rather than decrease in the future, and those who act proactively today will be better positioned tomorrow.

The Pentagon's plans are a clear signal that data in the digital age is a strategic resource of utmost importance. For Swiss SMEs, this means leveraging AI opportunities with a clear awareness of the risks and a robust strategy for data security and sovereignty. Those who protect their data, protect their business model.

Conclusion: Seize AI Opportunities, Secure Data Sovereignty

The Pentagon's ambitions to train AI models with classified data underscore the critical importance of data in the AI era. For Swiss SMEs, this serves as a reminder to proactively secure data sovereignty and compliance when implementing AI solutions. The benefits of AI are immense, but they must not come at the expense of security and confidentiality.

✅ Focus on Local and Private AI Solutions: Opt for on-premise or Swiss cloud solutions to maintain control over your data and comply with the Swiss Federal Act on Data Protection (FADP).

✅ Establish Strict Data Governance: Implement processes for anonymisation, pseudonymisation, and classification of data before it enters AI models. Transparency about data origin is crucial.

✅ Embrace Responsibility and Ethics: Train your employees in the use of AI and ensure that human oversight and ethical guidelines are guaranteed for all AI applications.

Would you like to learn more about how your SME can implement AI safely and in compliance with regulations? Contact us for a no-obligation consultation to discuss your specific requirements and find the right solutions for your business. Visit schnellstart.ai/en/contact.