A one-hour session with colleagues from 13 European countries, from Reykjavík to Athens. Most of it is practical, vendor-neutral advice you can use with any AI tool; the last part shows how we handle it in Validemic.
Seven takeaways
- Three questions decide almost everything: what data is it, which plan is the tool on, and who decided (and where is it written down)?
- No signed DPA, no personal data. The data processing agreement is the contract that says what the vendor may do with your data.
- Same logo, different contract. Free and enterprise versions of the same AI tool have different rules on training, retention and admin control.
- Classify the data, not the tool. A simple public / internal / sensitive traffic light helps staff decide in seconds.
- Review like airport security: a fast track for public data, a standard lane for internal data, extra screening for sensitive data.
- Most attackers don't break in, they log in, often through a vendor or an app someone connected. Every AI tool joins your attack surface.
- The renewal date isn't your deadline; the notice date is. Start looking at alternatives three months before it.
Copy-paste prompts
Use these with an AI tool your university has approved, and only with public documents such as privacy policies, terms and published DPAs. Treat the answer as a first pass and check every flagged clause yourself.
Check a vendor's terms for red lines
Here are this vendor's privacy policy, terms and DPA. Check them for each red line in the list. For every one, answer yes / no / not stated, and quote the exact clause. Red lines: 1. Trains AI on our data 2. Keeps data longer than needed 3. Sub-processors outside the EU 4. Vendor staff can read our content 5. Can change terms without notice 6. Breach notice slower than 72 hours Don't summarise. Only report what the documents say.
Compare vendors before you buy or renew
Compare these 3 vendors using only the attached DPAs, terms and privacy policies. For each, tell me: data location · sub-processors · AI training on our data · retention · renewal and price terms · exit and deletion. Quote the exact clause for every answer. Write "not stated" if a document doesn't say.
A data traffic light for your intranet
| Data | Examples | Which tools |
|---|---|---|
| 🟢 Public | Published papers, course descriptions | Any approved tool |
| 🟡 Internal | Emails, student essays, draft manuscripts | Signed DPA, no training, university login |
| 🔴 Sensitive | Interview transcripts, health data, grades | Only after a DPIA, or a locally hosted tool. Often: don't. |
What we covered
Part 1 · The basics, for everyone
Four GDPR questions to ask before personal data goes into an AI tool: why are we allowed to use it (usually because it's part of the university's job, rarely consent), does the data leave the EU (then the vendor needs the EU-US Data Privacy Framework or standard contractual clauses), is it high-risk (then do a DPIA first), and is it really anonymous (codes aren't anonymity if a key file links them to people). We then walked through the EU AI Act's four risk levels with university examples: AI literacy has applied to everyone since February 2025, and high-risk uses such as admissions, grading and proctoring need human oversight. Finally: the same AI can be used in very different ways, from a personal chatbot account with no contract to a tool with only EU sub-processors, and 25 of 35 European universities we checked offer Copilot Chat but with very different data rules.
Part 2 · Data protection
How to review AI tools without becoming the bottleneck: sort requests into lanes by the data involved, re-review only when something changes (new sub-processor, new feature, new terms, renewal), and only rely on a tool's findings if each finding shows where it came from.
Part 3 · IT and security
Recent incidents in Denmark, Germany and elsewhere show the same pattern: attackers came in through a vendor's or an app's legitimate access. We looked at shadow AI (the tools nobody approved), what to do about it (single sign-on, admin approval for apps that read mail and files, and one good approved tool before any ban), and a prompt for letting an approved AI tool check a vendor's terms for red lines.
Part 4 · Procurement
Track notice dates, not renewal dates; compare the quoted price with the published one; cap renewal increases; make sure exit includes export and deletion. We shared a prompt for a first AI-assisted comparison of vendors' public documents, and a four-step way to find alternatives before you renew.
For researchers and the library
Where AI helps researchers (reading, writing, analysis and study design), the EU's living guidelines on responsible use of generative AI in research (reliability, transparency, respect, accountability), and how to run tool trials that actually end in a decision.
Sources mentioned
- European Commission and ERA Forum, Living guidelines on the responsible use of generative AI in research (latest version, October 2026).
- Regulation (EU) 2024/1689 (AI Act) and Regulation (EU) 2016/679 (GDPR), EUR-Lex. Dates for high-risk obligations may change under the Digital Omnibus; check the current status.
- EU-US Data Privacy Framework participant list, dataprivacyframework.gov.
- Guidance quoted from Utrecht University, the University of Helsinki and the European University Association (2026).
- Validemic, AI tools approved by European universities: 35 checked (sources read 7 October 2026).
- Reporting on the Danish CPR register breach (Copenhagen Post, October 2026), the HSZ-BW attack (Staatsanzeiger, October 2026) and the Vercel breach via a third-party AI tool (Help Net Security, April 2026).