The previous post was about planning an AI security assessment. This one moves to what sits outside the attacks: privacy, governance and regulation.
In March 2023 Samsung opened ChatGPT up to engineers in its semiconductor division. Inside three weeks the South Korean business outlet The Economist reported that the company had already caught at least three leaks of confidential material (AI Incident Database). One person pasted in faulty source code to get help finding the problem, another pasted in the transcript of an entire internal meeting to have it written up as minutes. By May, Samsung had banned it outright (reported by Fortune).
Data privacy: once data is in the model, you cannot delete it#
AI Dark Arts (23) showed that training data memorized too well can be asked back out, and that vectors in a knowledge base can be turned back into text. Once data goes into a training set or a knowledge base it stops being a file you can delete. It is spread across the model’s weights and across a set of vectors, and pulling it back later means not knowing where to pull from.
So this line of defense goes in before the data does. Work out which data should never go in at all, and strip or mask sensitive fields like names and medical record numbers before anything goes to training or into a knowledge base. If something genuinely has to be used, then somebody owns that batch of data: who approved it and what it is used for get written down, rather than anyone sending in whatever they feel like.
AI governance: when something goes wrong, who is responsible#
Can your company answer these?
- Which data must not be pasted into an AI.
- Who approved connecting this AI tool.
- After an incident, whether you can find out who sent what out, and at which step.
There is a lot more to the policy side, but none of it has to start from nothing. When a company is genuinely building this out, NIST’s AI Risk Management Framework works as a starting draft: settle who is responsible, then work down through identifying, measuring and treating risk.
To get to the point of being auditable, follow ISO/IEC 42001. It is the AI version of ISO 27001.
What those two documents are for is turning “who, at which step, keeps what evidence” into a process people can actually run, and wiring security into how AI gets developed and operated.
You are not going to build the policy side single-handed. What you can do is ask whether the company has a written rule about which data must not be pasted into an AI, and if it does not, raise it with your manager.
If you run the company and you are reading this, one more thing worth keeping in mind. Announcing that AI is banned will not hold, because people will still use their own accounts to get their work done. You may as well pay for a commercial plan for everyone. OpenAI’s enterprise offering states publicly that it does not train on your data, and you can see who is using it and how much. This is not a sponsorship and I am not saying it has to be OpenAI. Apply the same logic to find a commercial option you trust, and make it good enough that nobody wants to go back to their personal account for company confidential work. Do that and your AI governance has come a long way.
AI regulation: how you plan to use AI is not entirely up to you#
The first two are internal matters, where how far you go is your own call. Regulation is the regional and global set of rules, and you are in the game whether you agree with it or not.
The EU’s AI Act came into force in August 2024, grading by risk from prohibited, to high risk, to limited risk, to minimal risk. What you plan to use AI for decides how heavy your obligations are. Using it to score the entire population is prohibited outright. Using it to screen CVs, do credit scoring or read medical images counts as high risk, which means risk assessments, record keeping, and leaving room for human intervention. Even outside the EU, an AI service that reaches EU users falls under it.
Taiwan’s Basic Act on Artificial Intelligence passed its third reading in December 2025, setting out seven principles the government must follow in promoting AI, among them privacy protection and data governance, security and safety, transparency and explainability, and accountability. It is a basic act, so it sets direction, and how it lands in each industry depends on the subordinate regulations that follow. But principles like “AI has to be explainable, and somebody has to answer when it goes wrong” are now written into law. On top of the Personal Data Protection Act, which already applies, using people’s data for training or feeding it to an AI needs a lawful basis in the first place. Having the data is not the same as being allowed to use it.
A short wrap-up#
The first step in governance is a list. Put down every AI in use anywhere in the company, and for each one write what it is used for, what data it touches, and who approved bringing it in. Then grade by impact: internal document drafting goes through a simplified process, and anything touching CV screening, credit scoring or medical imaging goes through full review. Each entry also gets the name of whoever signed off on it going live, and the name of whoever can call a halt when something goes wrong, and those two names should not both belong to the people who built the system.
The next post gathers what this series has covered into a list, so that when you finish it you know where to go next.