Sweden's data protection authority is reviewing a *job centre's AI*

The Swedish Public Employment Service (Arbetsförmedlingen) uses an AI system to assess jobseekers and decide what support they should get: coaching, training, work experience or subsidised employment. Now Sweden's data protection authority, IMY, is reviewing it ^[1].

IMY hasn't said what prompted the review or whether it found anything wrong. What it has said is that the review is under GDPR Article 22, which gives people the right not to be subject to decisions that are solely automated and have legal or similarly significant effects ^[2]. There are exceptions, but when an exception applies the organisation must still provide a way for people to contest the decision and get human involvement.

The bigger picture is the AI Act. From 2 December 2027, systems that assess people for employment and vocational training will be high-risk AI under Annex III ^[3]. That brings logging requirements, impact assessments, human oversight and a right to an explanation.

This review is one to watch, because it asks questions every public body using AI will soon have to answer.

Think of it like a referral

Say a GP uses a screening tool to decide whether to refer you to a specialist. You want to know it worked properly, that a person checked the output, and that if you ask in six months someone can show you what the tool said and which version it was. The same logic applies when an AI system decides what a jobseeker needs.

The Employment Service says the AI supports the decision, not that it makes it on its own. The question IMY will look at is how much room the person actually has to go against what the system suggests, and whether jobseekers know that a decision affects them and can challenge it.

What Article 22 already requires

GDPR Article 22 has been in force since May 2018 ^[4]. It says people have the right not to be subject to a decision based solely on automated processing, including profiling, that produces legal effects or similarly significantly affects them. There are exceptions: when the decision is necessary for a contract, authorised by law, or based on explicit consent. Even then, the person must be able to get human intervention, express their view and contest the decision.

For a public employment service, deciding what support someone gets counts. If the decision is fully automated and it affects whether you get training, a work placement or financial support, Article 22 applies unless there's a legal basis that allows it. And even with a legal basis, you still need a human in the loop and a way for the person to challenge the outcome.

IMY's review will look at whether the Employment Service meets those requirements now.

What the AI Act adds from December 2027

The AI Act raises the bar. Article 6 and Annex III list employment and vocational training as high-risk ^[3]. Systems used to recruit, select or assess people for jobs or education fall under it. So does AI that assigns people to training programmes or decides what support they get from a public employment service.

From 2 December 2027, providers and deployers of such systems must:

• carry out risk management and data governance (Articles 9 and 10),
• keep technical documentation,
• log events automatically over the system's lifetime (Article 12) ^[5],
• have human oversight,
• go through conformity assessment before placing the system on the market.

For public bodies and for private actors offering public services, there's an extra step. Article 27 requires a fundamental rights impact assessment before the system is put into use ^[6]. That assessment must look at the purpose, whether less intrusive means exist, who is affected, the risks to fundamental rights, and what safeguards are in place.

Article 86 gives the person affected a right to an explanation. If a high-risk AI system under Annex III (except critical infrastructure) has been used to make a decision with legal or similarly significant effects, the person can ask for a clear and meaningful explanation of the AI's role and the main elements of the decision ^[7].

Those requirements apply to new systems placed on the market and to existing ones still in use after the date. For a system already running in 2027, that means compliance work starts now.

The logging question

Article 12 says high-risk AI systems must technically allow for the automatic recording of events over their lifetime ^[5]. What to log is spelled out: it's for traceability, for monitoring after the system is on the market, and for the deployer's oversight.

What Article 12 doesn't say is how to protect the logs. It asks for the logs to exist, not for any particular way of keeping them unchanged. But when a decision is challenged months later, the question shifts. Can you show that the log is exactly what was written on the day, and that nobody with access has quietly edited it since?

Access controls and monitoring are the foundation. An independent fingerprint per log goes one step further. A change shows up even when it's made by someone with the right access, because the fingerprint won't match.

Sealing the decision record

Here, the thing to seal is each assessment decision as it is written: the jobseeker reference, the model version, the recommendation and the name of the person who checked it. Sealing takes a fingerprint of that record, not the record itself, so the decision never leaves the Employment Service's systems.

The fingerprint is sealed with WitniumChain, which anchors it on a chain where any later change would show. To check, anyone with access to the record calculates the fingerprint again and compares it with the sealed one. Change a single character and it won't match.

The Employment Service, IMY, internal audit and, if needed, a court can all check it. Read the docs for how it works.

What it shows is that the decision record is exactly what was written at that time. What it doesn't show is that the decision was fair, lawful or in the jobseeker's best interest. Those are separate questions, which the impact assessment, the human oversight and the right to contest are there to address.

Your system sends a fingerprint of the record, not the record itself, through the REST API, the SDK or the MCP server, and gets a witness ID back. The first 2,000 witnesses a month are free.

What this means for you

If you run or build AI that assesses people for jobs, training, benefits or credit, ask:

• Does the system make decisions on its own, or does a person have real room to decide differently?
• If it's fully automated, is there a lawful basis, and can people contest the decision?
• Could you show, six months later, what the system recommended, which version ran, and who checked it?
• Have you assessed the impact on fundamental rights, as Article 27 will require from December 2027?
• Can people affected get a clear explanation of the AI's role, as Article 86 will require?

The logging part is often treated as a technical detail. It isn't. When a decision is questioned, the logs are what tell you what happened. If you can't show that they haven't been altered, they're easier to challenge.

What to watch

IMY's review is ongoing, and it hasn't published findings yet. Worth watching because it's a test case for how Swedish authorities will apply GDPR Article 22 to public employment AI, and because similar systems are used across the public sector.

The AI Act's high-risk requirements for employment and training systems apply from 2 December 2027. The clock is already running for impact assessments, technical documentation and logging.

For jobseekers, the right to an explanation under Article 86 comes into force on the same date. For public bodies, that means deciding now how you will provide it.