Using AI Safely and Responsibly · 1 min read
Things you can't do right away
Some AI uses with personal data are restricted regardless of which assistant you use. Fully automated decisions and special category data both require extra steps.
Introduction
Section titled “Introduction”Learning goals
- Understand why fully automated decisions about people are restricted
- Know what profiling is and why it requires extra steps
- Understand the higher bar that applies to special category data
Some uses of AI with personal data are restricted regardless of which assistant you use. These restrictions exist because the risks to individuals are considered too high without additional safeguards. You’ll notice some overlap with the EU AI Act’s high-risk categories — see the AI Act course for that angle.
Fully automated decisions
Section titled “Fully automated decisions”You cannot use an AI assistant to make a final decision that significantly affects a person — for example, whether they receive a benefit, a permit, or access to a service — without a human reviewing and being accountable for that decision.
This isn’t just good practice; it’s a legal requirement. Individuals have the right not to be subject to a decision based solely on automated processing if that decision has significant consequences for them.
The key word is solely. AI-assisted decision support is completely fine — the assistant can summarize a case, flag relevant factors, or give a recommendation. What it can’t do is be the only decision-maker. A human must genuinely review the output and be in a position to override it, not just sign off on whatever the assistant produced.
Example: A colleague suggests having the assistant automatically approve or reject permit applications with no human check — that’s not allowed. A person must be responsible and in a position to disagree.
Profiling
Section titled “Profiling”A related restriction is profiling — using AI to systematically evaluate personal aspects of someone, like their performance, behavior, or reliability. Even an internal AI-generated profile requires a clear legal basis.
Example: Using an assistant to generate a “performance score” for employees falls here. A person must be responsible and able to override the assessment.
Special category data — higher bar
Section titled “Special category data — higher bar”As discussed in section 1, health data, religious beliefs, political opinions, ethnic origin, and similar categories require stricter handling. You should generally avoid feeding this data into an assistant unless:
- The specific assistant has been approved for it
- Your organization has confirmed there is a valid legal basis
- A DPIA has been completed if required
When in doubt, anonymize — if you can accomplish your task without including the sensitive details, do so.
Test your knowledge
Question 1 of 3
A citizen has applied for disability support. You use an AI assistant to summarize the case file, including their diagnosis, for your own internal review. Is this OK?