This week, three Dutch studies appeared that together paint one picture: AI is already an everyday work tool, but organisations often don't know which tool is being used and with what data. What does that mean for you?
Shadow AI is the norm
According to the behavioural study Alert Online 2026, reported on by Computable, 49% of employees sometimes, often or always use a private AI account or app for work. 35% use AI tools without the organisation knowing about or having approved them. Only 39% know which AI tools are permitted at work. And 35% sometimes or more often bring substantive work problems to an AI tool.
That is precisely where the GDPR risk lies. A private chatbot account usually falls under consumer terms. Different rules for retention and training apply there than with a business subscription. With some providers, consumer chats are used for training by default unless you switch that off yourself. On /veiligheid/ you can see, per model, how this is arranged.
What you can do: draw up a short list of permitted tools and put it where everyone can find it. A ban without an alternative works poorly when half the staff already use private AI.
Financial professionals consult AI more often than the Tax Authority
Tax software company Nextens asked over 800 financial decision-makers in Dutch businesses where they get their tax and financial information from. Accountant.nl reports that 51.5% use AI tools such as ChatGPT, Copilot and Perplexity. That is more than the 46.5% who consult the Tax Authority. 79.1% use AI at least weekly, and 39.8% even several times a day.
Trust isn't keeping pace. 37.7% don't trust AI enough, 30.6% are worried about privacy, and only 22.6% almost always check the information. Notably, 53.6% are unaware of the obligation under the EU AI Act to make staff AI-literate. Just 6.4% have trained all their staff.
Are you working as an accountant or financial adviser? A general chatbot doesn't guarantee that tax rules are up to date. Always check amounts and deadlines against the source, and consider whether a professional tool with current tax sources would suit you better.
No single model refuses human rights violations entirely
Bits of Freedom published research by Aithos into seven models: Claude, ChatGPT, Grok, Gemini, DeepSeek, Mistral and Kimi. In eight scenarios, they were asked to assist with violations of, among other things, the prohibition of discrimination and the right to education. Claude refused in 96% of cases, Mistral in 11%. No model refused entirely. Bits of Freedom therefore calls for stronger European rules to better protect human rights.
This matters for HR, recruitment and the public sector: there, a model can directly influence decisions about people. A European model is therefore not automatically the safest in every respect. Compare models at /modellen/ and look beyond just where the data is stored.
In brief
- Is your team using private AI? Provide a clear, approved alternative and explain why.
- Do you use AI for tax questions? Always check the answer against the source.
- Choosing a model for decisions about people? Weigh up both privacy and behaviour.
