AI Distillation: How Swiss SMEs Benefit from Big Tech's Protection Measures

Nearly half of Swiss SMEs now see Artificial Intelligence as a business advantage. This is according to a recent survey by kmu.admin.ch from 2025. However, as interest in AI grows, a new, critical challenge emerges: How do major AI developers protect their models from unauthorised copying, and what impact does this have on Swiss companies?

The answer is complex. Behind the scenes, giants like OpenAI, Google, and Anthropic are working hard to combat so-called AI model copying – a process where the knowledge of their advanced systems is siphoned off and transferred into their own, often inferior or unauthorised models. This affects not only Big Tech itself, which fears billions in losses, but also every Swiss SME that relies on the integrity and exclusivity of these technologies.

As Lukas Huber from schnellstart.ai, who has been working with AI implementations in the Swiss SME environment for years, I am closely observing this development. The protection measures of the major players are not an abstract debate; they directly shape the rules of engagement for your AI strategy. It's about how you, as an SME, can benefit from innovations without falling into legal grey areas or jeopardising your investments.

How can Swiss SMEs leverage the benefits of advanced AI models without incurring the risks of model copying?

Through strategic partnerships, a focus on data sovereignty, and the implementation of robust governance structures.

So-called AI distillation, or "siphoning off" models, is essentially an attempt to transfer the knowledge of a large, expensive, and elaborately trained AI model into a smaller, often cheaper one. This is typically done by generating vast amounts of queries to the source model and using the responses as training data for the target model. For Big Tech, this means massive economic damage and a loss of competitive advantage. For you as a Swiss SME, it means the integrity of the AI solutions you rely on is directly affected.

The key is not to blindly rely on the underlying technology but to carefully examine the interfaces and the origin of the data. We must understand that the strength of an AI model lies not only in its size but also in the quality and exclusivity of the data it was trained on, and the specific "hyperparameters" that guide the training. These parameters, such as learning rate or batch size, are not learned by the model itself but must be manually set and optimised. They are part of the intellectual property.

For SMEs, this means that direct access to the latest proprietary models via secure APIs is the preferred route. This eliminates the risk of using unauthorised or inferior copies. At the same time, you must maintain sovereignty over your own data. When implementing an AI solution that processes your internal data, it is crucial that this data remains protected and cannot inadvertently contribute to the distillation of your own specific use cases. The design of the data architecture, including cloud infrastructure and data integration pipelines, must consider security and compliance, especially with the Swiss Data Protection Act (DSG), from the outset. This is a point we at schnellstart.ai repeatedly emphasise: Swiss hosting is not a luxury, but a necessity for trust and legal certainty.

Another important aspect is internal governance. An AI Governance Council, an interdisciplinary body comprising IT, Legal, Business, and Ethics, can define guidelines for AI deployment and review high-risk use cases. This ensures that all AI applications within the company are not only efficient but also legally sound and ethically justifiable. Such councils help to identify risks early and ensure compliance with laws such as the upcoming EU AI Act, which is also relevant for Swiss companies with EU ties.

What measures are Big Tech companies taking to protect their AI models from unauthorised use, and how does this affect external users?

Big Tech relies on technical hurdles like watermarking and API monitoring, as well as legal action and industry initiatives, which present both opportunities and increased requirements for SMEs.

Leading AI developers are investing heavily in protection mechanisms. A common method is "watermarking." This involves embedding invisible patterns or signatures into the outputs of AI models. If these outputs are then used to distill another model, the watermarks can prove that the knowledge originated from the original. This is comparable to a digital fingerprint that uniquely identifies the source. Such techniques require deep technical expertise and are part of the MLOps strategy, particularly in the area of operational monitoring.

Another protection mechanism is comprehensive monitoring of API usage. Every request to a large model leaves traces. Patterns indicating systematic distillation – such as an unusually high number of requests in a short period or specific query sequences – can be detected and blocked. This is a complex field that requires advanced analytics tools and machine learning itself to identify abusive behaviour. Such monitoring goes beyond simple access statistics; it analyses the intent behind the requests.

Legally, companies are also acting more aggressively. Lawsuits for copyright infringement and enforcement of terms of service are commonplace. The Frontier Model Forum, a collaboration between OpenAI, Google, and Anthropic, is an example of an industry initiative that not only promotes the exchange of best practices but also forms a united front against model copying. They aim to establish standards that make the ecosystem safer while fostering innovation.

For external users, including Swiss SMEs, these measures have two sides. On the one hand, the trustworthiness of the models used increases. You can rely on the quality and exclusivity of the services you pay for being protected. This is a significant advantage, especially when dealing with sensitive data or business-critical applications. On the other hand, the protective measures can lead to stricter terms of use, potentially higher costs, and less flexibility. API access could become more regulated, and integration into your own systems might become more complex. It's a balancing act between security and accessibility.

What does the increasing collaboration among AI giants mean for the accessibility and cost of AI solutions for Swiss SMEs?

The consolidation of protection efforts can lead to a more stable, but potentially more expensive and highly regulated AI landscape, requiring a careful strategy from Swiss SMEs.

The collaboration of AI giants in the Frontier Model Forum signals a clear trend: the industry is attempting to self-regulate and set standards before external actors do. This has advantages: a united front against misuse can strengthen overall trust in AI technologies. It fosters a more stable development environment where companies can rely on the dependability of base models. Efforts to prevent harmful AI applications and establish ethical guidelines are also positive.

However, this consolidation also carries risks for SMEs. If leading providers collude or effectively form an oligopoly, it could lead to less competition. This, in turn, could drive up prices for access to advanced AI models. For small and medium-sized enterprises in Switzerland, often operating with tight budgets, the cost of cutting-edge AI solutions could become a significant hurdle. It is crucial for SMEs to plan their AI strategy meticulously and weigh the balance between the capabilities of Big Tech models and the potential need to develop their own leaner models or opt for open-source solutions.

The concepts of MLOps (Machine Learning Operations) play a central role in this balancing act. Training and tuning open-source models in-house can be an alternative. This includes selecting the right hyperparameters, performing cross-validation to ensure model quality, and continuous operational monitoring. However, this requires internal expertise or collaboration with experienced partners. We at schnellstart.ai support SMEs precisely in designing the technical architecture for such scenarios, including ensuring Swiss data protection and scalability.

The future of AI utilisation for Swiss SMEs will largely depend on how they position themselves in this complex landscape. It is an opportunity to strengthen their own competitiveness through smart technology decisions while keeping risks under control.

The protection measures taken by Big Tech against AI model copying are a double-edged sword. They offer increased security and integrity for the models used, but can also influence accessibility and costs for Swiss SMEs. It is clear that a passive stance is not effective here. Instead, a proactive and informed strategy is required that considers the specific needs and regulatory framework of the Swiss market.

As Lukas Huber, with my expertise in the IPSO certificate in AI Business and MLOps, I am convinced that Swiss SMEs have the capability to master these challenges. The key lies in the combination of technical understanding, strategic planning, and a clear focus on compliance and data sovereignty.

Three takeaways for your SME:

• ✅ Strategic Utilisation: Opt for transparent and licensed AI solutions. Evaluate whether direct API access to Big Tech models or your own, customised open-source implementation is the better choice for your specific use cases.
• ✅ Data Sovereignty & Governance: Prioritise Swiss hosting and ensure your data is protected according to the DSG. Establish an AI Governance Council to define guidelines for AI deployment and manage risks.
• ✅ Proactive Risk Management: Stay informed about developments in AI model protection. Consider potential cost changes and regulatory adjustments in your long-term AI strategy to remain competitive.

Would you like to learn more about how your company can leverage AI opportunities securely and effectively? Get in touch with us to discuss your specific challenges: schnellstart.ai/en/contact