Sweden's data authority tells healthcare: you own your AI risks

The guidance isn't new law

Sweden's data protection authority, IMY, is running sessions on safe AI use in healthcare ^[1]. The timing makes sense: healthcare providers are adopting AI faster than many of them are updating their risk frameworks, and IMY wants to get ahead of the compliance gap.

The guidance doesn't introduce new rules. GDPR has applied to healthcare AI since it came into force in 2018, and patient safety laws haven't changed either ^[2]. What IMY is doing is spelling out what those rules mean when a diagnostic tool runs on a model nobody at the clinic fully understands.

Think of it like food labelling. The rules about allergens were always there; the guidance just tells you where to put the sticker.

You can't outsource the risk

One point runs through the whole guidance: the healthcare provider is the data controller, and you can't hand that role to a software vendor ^[1]. If an AI system processes patient data incorrectly, or makes a clinical suggestion that turns out to be wrong, the provider carries the responsibility.

That sounds obvious until you look at how procurement actually works. A hospital buys a radiology AI, the vendor says it's compliant and the contract lands on a desk in IT. Six months later, someone in legal asks whether the system logs every decision it makes, and nobody is sure.

IMY's guidance says the provider has to know. That means understanding what data the system processes, how it reaches its outputs, who can override it and what happens when it gets something wrong ^[1]. The vendor can help with documentation, but the provider has to be able to stand behind it.

Two sets of rules at once

Healthcare AI sits at the crossing of data protection and medical device regulation. GDPR governs how patient data is handled ^[2]. If the AI is a medical device, the Medical Devices Regulation (MDR) governs safety and performance ^[3]. A diagnostic AI has to meet both.

IMY's guidance focuses on the data protection side, but it acknowledges the overlap ^[1]. An AI system that suggests a diagnosis is processing sensitive health data under GDPR, and it may also be making a clinical decision that patient safety law cares about. Getting one part right doesn't excuse getting the other part wrong.

For smaller providers, that dual framework is the hard bit. GDPR compliance alone takes time, and adding MDR on top means either hiring expertise or relying on vendors to tell you what matters. The guidance doesn't solve that, but it at least names the problem.

Human oversight is not a formality

The guidance emphasises that AI outputs must be reviewed by qualified people before they affect patient care ^[1]. That sounds like common sense, but it's worth stating plainly: an AI can suggest, but a person has to decide.

In practice, oversight gets tricky. If a radiologist reviews 200 AI-flagged scans a day, at what point does review become rubber-stamping? IMY doesn't give a number, but the principle is clear: the human has to be able to catch the AI when it's wrong, and that requires time, training and access to the reasoning behind each output ^[1].

For high-risk AI under the AI Act, human oversight will be a formal requirement from 2 December 2027 ^[4]. Healthcare AI that falls under Annex III of the Act will need documented oversight procedures, and the people doing the overseeing will need to understand what they're looking at. GDPR already implies much of this; the AI Act will make it explicit.

What this means for you

If you're a healthcare provider using or considering AI:

• Check whether you can explain what data each AI system processes, how it reaches its outputs and who oversees it. If the answer is "ask the vendor", that's a gap.
• Document your risk assessments and your decisions about where human oversight is needed. GDPR requires this already ^[2], and the AI Act will ask for more of it ^[4].
• Make sure contracts with AI vendors clearly assign responsibility for things like security, data access and incident response. You remain the controller; the contract should reflect that ^[1].
• If the AI is a medical device, confirm that it meets MDR requirements ^[3] and that your staff know how to use it safely.

Logs decide what happened

When an AI system in healthcare makes a decision or a suggestion, someone may question it months or years later: a patient, a supervisor or a court. At that point, the thing that matters is the record of what the system did, which version was running and who reviewed the output.

A log its own operator can quietly edit is easy to challenge. Here, the thing to seal is each AI decision or suggestion as it's written to the log: the patient reference, the model version, the output and the reviewer. Sealing takes a fingerprint of the record, not the record itself, so nothing leaves your systems. If the decision is questioned later, anyone you share the record with can check that it matches the fingerprint sealed at the time.

WitniumChain does that through a REST API, an SDK or an MCP server. The content stays with you, and the first 2,000 witnesses a month are free. A sealed log doesn't show that the AI was right, or that oversight was adequate, but it does show that the record you're relying on is the one written on the day.

How it would work

1. Each time the AI writes a decision to the log, your system sends a fingerprint to WitniumChain and gets a witness ID back.
2. When the decision is questioned, you calculate the fingerprint again and compare it with the sealed one.
3. If they match, the record hasn't changed; if they don't, something has.

What to watch

From 2 December 2027, AI systems in Annex III of the AI Act become high-risk and must meet the requirements in Articles 8 to 15, including logging, human oversight and transparency ^[4]. Many healthcare AI systems will fall into that category. IMY's guidance is a head start on what providers will need to show.

The deadline is 18 months away, which sounds like time but isn't. If you're running AI in healthcare now, use that time to close the gaps IMY is pointing at: who owns the risk, what the logs show and whether oversight is real or just on paper.