Free AI where it helps. Tough rules where it can harm.
Regulate the consequences – not the intelligence
An AI system that helps a nurse summarise a medical record is not the same as an AI system that decides on its own what treatment a patient should receive.
A tool that helps a small business write code is not the same as a system that identifies and monitors people in the street.
Yet the AI debate too often treats the technology as though it were one single problem with one single answer. It is an intellectually convenient position – and politically useless.
Sweden must not become a lawless laboratory. But neither should we build a digital permit society in which every idea has to pass through lawyers, consultants and authorities before anyone is even allowed to find out whether it works.
We want Europe’s simplest rules for safe AI – and Europe’s strongest protections where AI can harm people’s bodies, freedom or privacy.
The principle is simple:
Freedom for the tool. Accountability for the consequences.
Three risk levels – not three thousand rules
Swedish and European AI regulation must be based on what a system actually does, how much power it is given over people and what harm it can cause.
Not on the fact that the software happens to contain AI.
Green – let people build
Low-risk AI should be able to be developed and used without special permits.
This applies to tools for programming, translation, administration, search and knowledge support. It applies to research, product development, creative work, customer service, documentation, planning and logistics – not least tools that give people with disabilities greater freedom.
No special AI bureaucracy is needed here. Ordinary law already applies.
Fraud is fraud even when AI is involved. Discrimination is discrimination even when AI is involved. Breaches of confidentiality do not become less unlawful because a language model is used.
But the state must not add another layer of permits, fees and pre-approval simply because the technology is new.
No general AI licence. No AI tax. No pre-approval for ordinary AI tools.
When the risk is low, the threshold should be low too.
Amber – when AI is given power over people
Requirements must become stricter when AI moves from helping a person to assessing them.
This applies to systems that affect people’s rights, livelihoods, education or access to essential public services – for example in recruitment, lending, insurance, the exercise of public authority, social insurance, admission to education, the justice system and critical infrastructure.
In such cases, the individual must be told when AI has played a significant role. The decision must be logged and open to scrutiny. The decisive factors must be explainable, and anyone affected must have the right to a genuine human review.
Above all, a natural or legal person must carry the responsibility.
“The computer said no” is not an explanation. It is a capitulation.
A machine can calculate, recommend and identify patterns. But legal and moral responsibility can never be outsourced to mathematics.
Red – when bodies, freedom and privacy are at stake
Regulation must be tough here.
This applies to AI that makes medical decisions carrying a risk of serious patient harm or processes highly sensitive personal data. It applies to biometric identification and surveillance, the manipulation of children and other vulnerable people, and automated decisions that could deprive someone of their freedom or other fundamental rights.
The same strict approach must apply to sexual and fraudulent deepfakes, as well as systems that try to build permanent behavioural profiles of people’s private lives.
A user agreement on page 47 is not enough here.
What is needed are real rights, clear limits and someone who can be held accountable.
Healthcare should use AI – patients must not be treated as test material
Healthcare shows why both the technology romantics and the prohibitionists are wrong.
Swedish healthcare is weighed down by administration, waiting times, poor information transfer and staff shortages. Rejecting AI on principle would not be caution. It would be defending inefficiency.
AI should be used boldly for documentation, medical-record summaries, scheduling, medication checks, diagnostic support, image analysis, research and decision support.
But the dividing line must be crystal clear:
The greater the risk to a patient’s life and health, the stronger the requirements for human accountability, validation and follow-up.
When an AI system is materially important to diagnosis or treatment, there must be a named healthcare provider and clear professional responsibility. Clinical systems must be monitored on real Swedish patients and patient groups so that errors, biases and declining performance are detected. Serious incidents must be reported.
Patients must never end up in no man’s land between the doctor and the algorithm.
AI may help doctors do a better job. It must never become an excuse for no one to be responsible any more.
Patient data is not free raw material
A medical record contains some of the most private information a person can disclose: illnesses, mental health, genetics, sexuality, addiction, pregnancies and family history.
This must not be turned into commercial training material simply because the technology makes it possible.
Using patient data for AI must require a clear legal basis, a defined purpose and strong technical and organisational safeguards. Data must not be collected or shared merely because it is convenient.
Where the same result can be achieved through local processing, pseudonymisation, federated learning, secure research environments or other privacy-preserving methods, those solutions should be favoured.
Public healthcare must not sign contracts in which patients’ information effectively becomes the supplier’s payment.
The patient seeks care. The patient does not sell their privacy.
No biometric mass surveillance
A free society must not require people to identify themselves to a machine every time they cross a square.
We oppose general and continuous biometric mass surveillance in public spaces.
Targeted use in cases of particularly serious crime must be precisely that: targeted, subject to legal oversight, proportionate, time-limited and open to scrutiny.
There is a decisive difference between looking for a suspected murderer and making the entire population searchable around the clock.
Technology makes both possible.
Politics must understand the difference.
Personal AI should be allowed to be personal
Personal privacy must not become an argument against all personal AI.
People should be able to choose an AI that knows their calendar, language, documents and preferences. But the balance of power must be clear.
Users must be able to see what information is stored, delete it, transfer it and refuse secondary use. It must be clear whether the data is used for training or other commercial purposes.
Privacy means that people control their own information.
It does not mean that the state should ban them from using it.
Stop deepfake fraud – not satire
AI makes it possible to imitate people’s faces and voices at very low cost. Clear rules are needed here.
No one should be able to use the technology as a shield for fraud, blackmail, sexual abuse or portraying a real person as having said or done something that never happened.
Manipulated material that could reasonably be mistaken for reality must be clearly labelled.
But parody, art, satire and obvious fiction must not become a new area requiring official approval.
Regulate the fraud and the harm. Not the imagination or the tool.
The public sector must use its purchasing power
The state, municipalities and regions are among Sweden’s biggest customers. That gives them power – if they understand how to use it.
When the public sector buys high-risk AI, contracts must require documented security, clear accountability, the possibility of audits and control over sensitive data. There must be open export formats, incident reporting, reasonable transparency and a genuine possibility of changing supplier.
Sweden must not lock healthcare, public authorities and municipalities into a single American, Chinese or European company because no one thought about the exit clause when the contract was written.
Whoever controls the customer relationship and the data will eventually control the market.
The public sector must understand this before the contract is signed – not ten years later.
One door in – not five authorities and seven consultants
Companies that want to develop serious AI in Sweden should not have to receive five different answers from five different authorities.
PTS, IMY, the Medical Products Agency, the Swedish Financial Supervisory Authority, Swedac and other relevant authorities should therefore coordinate a single point of entry for AI issues.
One door in. One answer. One accountable process.
Small businesses, researchers, municipalities and healthcare providers should be able to get help determining which risk class a project falls into and which requirements actually apply.
It is better for an authority to help people do things properly before launch than to issue fines three years later.
Build sandboxes – not waiting rooms
EU rules provide for regulatory sandboxes. Sweden must use them properly.
Companies, universities and the public sector should be able to test new AI together with supervisory authorities under controlled conditions. Healthcare, public administration, industry, energy, education and cybersecurity should be prioritised.
Every trial must have an end date, a clear objective and measurable safety requirements. If the technology works and meets the requirements, it must also be able to move forward.
A sandbox in which the project is forced to remain is merely a waiting room with a more modern name.
Do not pile a Swedish mountain of rules on top of Europe’s
Europe has already adopted extensive AI legislation. Sweden must not add 27 national rulebooks on top of it.
As a general rule, we should not introduce our own general requirements for AI models and AI companies where the EU already regulates the same issue.
Swedish rules should focus on the areas where Sweden genuinely has its own responsibility: healthcare, the exercise of public authority, confidentiality, public procurement, working practices and the protection of fundamental rights and freedoms.
In Brussels, Sweden should pursue the same line. Every new rule must be able to answer three questions:
- What specific harm is the rule intended to prevent?
- Is there a simpler way to do that?
- Is the benefit greater than the cost for everyone who has to comply with it?
If no one can answer the first question, the rule should not exist.
Open models are not a crime
Open technology creates competition. It enables universities, smaller companies and European players to build on existing work without first asking a handful of global platforms for permission.
Open-source software and open models should therefore not be treated as more dangerous simply because more people can use them.
Regulation must be based on a system’s capabilities, use and actual risk – not its business model.
We will not fight a new AI monopoly with rules that only the monopolist can afford to comply with.
Every AI rule should have a best-before date
Technology develops fast. Legislation does not.
New Swedish AI-specific rules should therefore, as a general rule, be reviewed after three years.
If the problem has changed, the rule should change. If it has failed, it should go. If technology or the market has already solved the problem, the state should not continue administering yesterday’s risk.
The old system’s favourite word is permanent.
Technology’s favourite word is next.
Politics must be able to handle both.
Our first decisions
During the first 100 days of the parliamentary term, we want to:
- Create a single Swedish point of entry for AI regulation and regulatory sandboxes.
- Give healthcare a national fast track for the safe testing of AI that can reduce administration, waiting times and medical errors.
- Introduce common requirements for public procurement of high-risk AI: accountability, audits, data portability and security.
- Stop new general Swedish AI rules when the government cannot show what specific harm they are intended to prevent.
- Produce clear national guidelines for patient data, biometric data and other highly sensitive information in AI systems.
- Give people an effective right to have significant automated decisions explained and reviewed by a human.
- Press in the EU for future AI rules to be technology-neutral, risk-based and feasible for smaller companies to comply with.
- Require regular sunset reviews of Swedish AI-specific legislation.
Our principle
We are not afraid of artificial intelligence.
We are afraid of poor use of power, poor security and monopolies that no one can challenge any more. That is something entirely different.
AI can give doctors more time with patients, teachers more time with pupils and entrepreneurs the chance to compete with a company one hundred times their size. It can help researchers discover patterns that no human could have found alone.
It would be absurd to regulate all this away because the technology can also be misused.
But it would be equally absurd to pretend that medical records, faces, voices and decisions about people’s lives are just data like any other.
That is why we choose a different path:
Broad freedom where the risk is low.
Tough accountability where the consequences are serious.
Strong protection for the body, freedom and privacy.
Above all:
Regulate what AI does to people – not the fact that AI exists.
That is how Sweden protects people without closing the door to the future.
That is responsible intelligence in politics.