Summary of the key takeaways from the webinar presented by Edern TRANVOUEZ, Brandon IP – Brandon Group, alongside its partner PI Motion, represented by Frédéric DUQUET.
- The reality: AI is advancing rapidly, but at what cost?
You probably use an AI assistant on a daily basis as part of the evaluation and development of your innovations: for monitoring, summarising and comparing prior art, testing your ideas, researching possible developments… The time savings are real. The question is no longer ‘should we get started?’, but ‘how can we use it without unwittingly revealing your company’s most valuable assets?’.
The risk is no longer theoretical. Last September, an undisclosed ChatGPT model claimed to have solved one of the seven millennium problems (the Navier–Stokes equations). At the same time, mathematicians Buckmaster and Alpöge, using OpenAI’s Codex, announced a comparable breakthrough and were asked whether OpenAI’s model could have been influenced by their own work. OpenAI’s response speaks volumes:
“Although this is unlikely, we cannot rule out the possibility that anonymised data derived from the use of our products may have contributed to the improvement of our models.”
Applied to your own situation: submitting an idea, a technical hypothesis or a workaround without taking precautions means accepting that this contribution may one day become the raw material for your competitors’ innovations. For a patent that has not yet been filed, there is a risk of losing novelty and the competitive advantage you are currently building.
Data is the lifeblood of these models. Sources of high-quality human text are running dry. Hence the recent wave of acquisitions and partnerships: Microsoft’s takeover of GitHub ($7.5 billion), the OpenAI–Reddit deal ($60 million), the reported $60 billion acquisition of Cursor, and Anthropic’s ‘Panama’ project. The key point to bear in mind is that a free or very low-cost service is often based on a data-collection business model. The price you don’t pay in euros, you pay in data.
- The landscape: a structural risk, not just a contractual one
The LARA study (Aithos) concludes that in real-world situations, without human supervision, virtually all available AI models and agents regularly breach the GDPR and the AI Act. Recent incidents, such as the breach at Hugging Face (July 2026) and the cyberattack targeting Australian government departments and UN agencies via OpenAI agents (September 2026), illustrate this unequivocally.
For an SME, the challenge is twofold:
-
- Contractual: you are liable for the data you entrust to a technology processor.
- Reputational: your customer, payroll or strategic data does not always belong to you; using it in a service where you have no control over the retention policy leaves you personally exposed.
What’s more, the Cloud Act (2018) complicates matters. It allows US authorities (the FBI, DOJ) to obtain data from providers subject to US jurisdiction, even if that data is hosted in Europe. A Microsoft, AWS or Google data center located in Paris, Frankfurt or Dublin remains under the control of a US company. Data stored in France – even if it is encrypted and covered by the GDPR – can, by law, be disclosed upon a court order without you being informed. It is therefore important to know where your data is hosted and by whom.
When it comes to providers, vigilance is essential:
Anthropic (Claude): the company’s communications highlight a ‘Privacy by design’ approach, but the consumer version trains the models on users’ conversations by default. It was not until August 2025 that an opt-out option was offered, and data retention can last up to five years. Worse still, changes introduced in July 2026 allow data to be passed on to the authorities based solely on internal assessment, without any external oversight. In practice, only the Team, Enterprise or API plans guarantee that your data will not be used to train models.
OpenAI / Google: they are subject to US jurisdiction, which exposes them directly to the Cloud Act. Once again, OpenAI’s and Google’s free services expose your data directly to the training of their models. In Google’s case, there is also deep integration into its ecosystem (Search, Workspace, Cloud), which multiplies the channels through which data flows beyond the simple chat service.
Chinese models (DeepSeek, etc.): the cost is up to 80 times lower, but at the expense of extensive collection of personal data (email address, telephone number, typing speed, content of conversations) hosted in China and shareable with the authorities. There is no reference to the GDPR or the AI Act in the terms and conditions of use. As things stand, providing data is not recommended. There is one caveat, however: when the model’s weights are open (i.e., made freely available to enable use on third-party servers), it is possible to host them yourself on sovereign infrastructure, which eliminates the risk of data collection.
The CNIL guide serves as a practical reference: it details, platform by platform (Gemini, Meta AI, ChatGPT, Copilot, Grok, DeepSeek, Mistral, Claude, WhatsApp, LinkedIn), the practical steps to take to object to the reuse of one’s data for training purposes. Refusing this reuse is not merely a matter of convenience; it is a right. We therefore encourage you to follow this advice.
- A pragmatic strategy to protect your assets
Whilst there is no single solution that suits everyone, three credible and complementary approaches are available to you:
(i) Prudent use of major US models: this involves using general-purpose assistants (Copilot, Claude, Gemini, etc.) whilst applying rigorous privacy settings, in accordance with the recommendations of the CNIL guide. This approach is strictly reserved for non-sensitive data. Let us be clear: this is a solution for use, not a solution for protection. The structural risk linked to the provider’s jurisdiction and business model remains.
(ii) A solution developed and hosted in Europe: you entrust your data to a provider whose infrastructure is located in France (Mistral AI, Safebrain, Algo IA, Scaleway) or in Switzerland (Infomaniak/Euria, Proton, Lumo). The service complies with the GDPR and the AI Act, and a contractual commitment guarantees that no training is carried out on your data. Furthermore, you fall outside the scope of the Cloud Act without sacrificing performance.
(iii) Hosting of open-source pre-trained models: you install the model (Gemma, Muse, Qwen, DeepSeek, Nemotron) yourself on your own infrastructure, and you control its entire lifecycle. Scaling is now very accessible: a recent laptop can run a model with 4 to 9 B of parameters; a high-quality graphics card or certain MacBook Pro models can handle models with 17 to 35 B of parameters; and a dedicated server can host models with 70 to 512 B of parameters, or even more. The advantage is that your data never leaves your organisation.
For example, how we operate at Brandon IP: a three-tier hybrid approach:
- Non-confidential, non-identifying data (prior art search, previously published application, search report): Large language models (Copilot, Claude) with privacy settings in line with CNIL recommendations.
- Non-sensitive but identifiable data (internal files, communications with clients): AI hosted in Europe, compliant with the GDPR and the AI Act (Euria by Infomaniak).
- Sensitive data (invention disclosure, patent application, litigation file, etc.): A local server hosting open-weight models (Qwen 3.8 27B, Qwen 3.6 35B A3B, Gemma 26B A4B).
This approach allows us to harness the benefits of AI whilst ensuring that neither our data nor that of our clients is compromised.
- Key takeaways
AI has become an indispensable tool in your executive, R&D or management roles. Used wisely and with the right provider, it boosts productivity. Used blindly, it exposes your sensitive data, sometimes without you even realising it.
The good news is that control is within your grasp. As outlined above, a few measures are all it takes to separate value from risk: a clear internal policy on what can be entered into a wizard, rigorous privacy settings, and a hosting choice suited to the sensitivity of your data.
