Copyright that builds Swedish AI

Who gets the fruits of people’s work?

Sweden must not meet AI with fear and bans. If we do, we lose both the technology and the money. But nor should we be so naïve that we give away an entire country’s language, culture, research, professional expertise and public knowledge – only to buy the results back from some of the world’s largest companies.

Training data did not fall from the sky. It consists of people’s work, experience and memory. Giving it away without conditions would not be modernity. It would be a bad deal.

Two Swedish traditions meet here. One says that people’s work has value and that economic power must be open to democratic scrutiny. The other says that those who understand the infrastructure and the market can also influence the terms. Together, they lead to a practical conclusion: we should organise copyright so that Sweden can build AI – and so that the people behind the knowledge share in its value.

A knowledge commons – one agreement instead of ten thousand lawsuits

We want to bring together Swedish literature, journalism, research, cultural heritage, public texts and other rights-cleared knowledge in a licensed Swedish knowledge commons. Not so that the state owns the works, but so that an individual author, photographer or researcher does not have to negotiate alone with the world’s largest technology companies.

There should be one door in, one agreement and a working payment route – but also a clear right to say no. Anyone who takes part should receive remuneration.

We are bringing the basic idea of the Nordic extended collective licensing system into the age of AI. Copyright that can be defended only through five years of litigation is not much of a right. In practice, it is a privilege for those who can afford it. We want to make it real for everyone.

Whoever sells the intelligence pays for the knowledge

The fee should not burden the creator, but the business. Those who sell AI services on the Swedish market and build value on Swedish knowledge should also pay back into the knowledge that made the business possible.

The terms should at the same time favour those who want to build together with Sweden. Companies that disclose their data sources, invest here and give Swedish research and the public sector access to models and computing capacity should be able to secure better terms. Those who merely extract the raw material, shut the model away and take the value elsewhere should pay more.

It should be cheaper to build with Sweden than to harvest Sweden.

The public sector must also use its economic strength. Without a valid licence and proper reporting of training data, there should be no major contracts with Swedish healthcare, schools, government agencies or other public services either. Companies can move their servers; they cannot move the Swedish market. That is where our bargaining power lies.

The money back into the next book, the next discovery, the next idea

The remuneration should not disappear into the national budget. It should go back into creation. When usage can be measured, the person who created the work should be paid. When it cannot be measured precisely, the remuneration should be distributed collectively, transparently and according to understandable rules.

When knowledge has been developed collectively within a profession – in healthcare, schools, research or public administration – the money should return there in the form of research time, continuing education, better working conditions and room for new creation.

This is not about paying for yesterday. It is about financing tomorrow. An author who can make a living from their work writes the next book. A journalist with resources carries out the next investigation. A researcher with time makes the next discovery. A nurse who is given the chance to develop their profession creates knowledge that does not yet exist in any model.

AI can process the archive. Only people can create what does not yet exist. If these people are not paid, future knowledge will also dry up. That is the real economics behind copyright.

That is why Sweden must act now

Some say that stronger copyright slows AI development. They have the problem backwards. What slows things down is uncertainty: tens of thousands of unclear rights, drawn-out proceedings, companies that do not know what they may use and rights holders who do not know whether they will be paid.

We should replace this disorder with a functioning market: a lawful route to Swedish training data, clear prices, understandable rights and real obligations. That is not less market, but a better market. It is not less technological development, but a precondition for Sweden to have its own.

We have something extremely valuable: an entire language, world-leading research, extensive archives, a highly qualified public sector and generations of documented professional knowledge. That knowledge should neither be locked away nor given away. It should be opened up on Swedish terms.

We are not going to ban the machines. We will build them and use them. Sweden should be a country where AI companies want to develop because the knowledge is good, the rules are understandable and the rights are clear.

But one condition stands firm: the person who creates the value should also share in the value. That is justice, but it is also good economics. Above all, this is how we make sure there is new Swedish knowledge for the next generation of people to create – and for the next generation of AI to learn from.

Copyright that builds Swedish AI – implementation and economics

Sweden should offer a lawful and practical route to Swedish training data. Rights holders should be paid, research should retain its exception and Swedish AI companies should be able to obtain clear licences without negotiating each work separately.

The knowledge commons should be a federated licensing system – not state ownership of books, images, music or research.

Implementation plan

The first 100 days

PRV, together with rights holders, collective management organisations, universities, cultural institutions and AI companies, will be tasked with submitting a robust legislative and implementation proposal. Facts, ideas, methods and artistic style must not be granted new exclusive rights.

2027

A machine-readable Swedish rights and licensing protocol will be built. A work or catalogue should be able to be marked as:

• permitted for training
• prohibited for commercial training
• available under licence
• permitted for research
• or permitted for RAG and search but not for foundation-model training.

Four voluntary licensing pilots will be launched: press, literature, professional photography and music. Small and newly established AI companies will receive a low basic fee or deferred, revenue-based payment.

2028

The pilots will be evaluated with regard to prices, coverage, administrative costs, competition and actual distributions. If a representative organisation and a functioning market exist, a specific extended collective licence for AI training will be introduced within the defined area. Rights holders will retain the option to opt out.

2029

A federated register will be linked to European standards. Providers selling models on the Swedish market should be able to show a training summary, copyright policy and the licences relevant to the service being sold.

Requirements in public procurement should apply to the model and service actually being procured. They must be proportionate, open and non-discriminatory – not a general political ban on a supplier. The National Agency for Public Procurement’s fundamental principles.

2030

The first distribution should have been completed. If voluntary licences and extended collective licences leave a documented gap, a limited residual remuneration may be introduced for verified use that cannot be attributed to an identified rights holder. No use should be charged twice.

2031–2032

Sweden will promote a common European licensing and metadata architecture and international compatibility through the EU and WIPO. Systems without measurable net distributions or with unreasonable administrative costs will be wound up.

Costs

The local detailed analysis gives the following estimated cost ranges:

These are analytical estimates, not quotations. On their basis, the following state planning ceilings are set:

The initial appropriation framework will be SEK 80 million in 2027. The remaining SEK 220 million will be released following the reporting of costs and results. If the technical solution becomes more expensive, its scope must be reduced or coordinated at EU level – not automatically given a larger appropriation.

Revenue and funding

Licensing revenue is not state revenue. It should go to rights holders and affected professional groups, after openly reported administrative costs.

The supporting material contains the following examples:

The scenarios are alternatives and must not be added together. The same use should not pay both a full licence fee and a full residual fee.

The older example – that three per cent of an AI market worth SEK 10 billion would yield SEK 300 million and three per cent of SEK 50 billion would yield SEK 1.5 billion – is mathematically correct but is not a revenue forecast. We lack verified data on Swedish turnover and a secure legal basis for a general three-per-cent fee. The figures must therefore not be used to fund promises.

Once the system is mature, administration and auditing should aim for no more than 10–15 per cent of collected licensing revenue. Computing capacity or model access may be accepted as a supplement, but only after independent valuation. It must not replace a rights holder’s cash remuneration or create vendor lock-in.

The state’s maximum net cost during the parliamentary term is therefore SEK 300 million. Licensing income is kept outside the national budget.

Distribution

The money should be distributed according to three principles:

• Direct remuneration when the use of an identified work or catalogue can be verified.
• Representative and publicly reported distribution when usage can be statistically established but not identified at work level.
• A time-limited claims reserve for non-participants and orphan works.

Most of the net funds should reach authors and performing artists. A smaller share may fund metadata, minority languages, digitisation and shared professional knowledge. Every krona and every administrative deduction must be publicly reported.

What we know

• Swedish law already permits text and data mining of material to which the user has lawful access, but the rights holder can reserve the commercial right. Research has a stronger exception that cannot be overridden by contract. Copyright Act, Sections 15 a–c
• Sweden already has extended collective licences and a general possibility of using them within defined areas of use. Non-participants must be treated equally and may request individual remuneration. Copyright Act, Sections 42 a and 42 k
• PRV is already the supervisory authority for collective rights management. There is therefore an institution to build on rather than immediately creating a new authority. PRV on collective management
• The EU AI Act requires providers of general-purpose AI models to have a copyright policy and publish a summary of their training material. From 2 August 2026, the Commission can intervene against new models that do not meet the requirement. Fines may amount to three per cent of global annual turnover or EUR 15 million. Older models that were on the EU market before 2 August 2025 have a transition period until 2 August 2027. European Commission rules
• The European Parliament has called for clearer licensing rules and collective licensing, but the decision is not binding law in itself. European Parliament supporting material

What we do not know

We do not yet know:

• how many Swedish works and catalogues have actually been used
• how many AI companies will sign licences
• what prices the market will accept
• how much can be distributed directly to individual authors and creators
• how large the Swedish AI turnover subject to licensing is
• how national rules reach models trained outside Sweden
• or whether revenue will exceed administration costs in each sector.

A copyright licence does not solve GDPR, confidentiality, trade secrets, patient data or protection against unauthorised digital copies of people’s faces and voices either. Such material must not be brought into the knowledge commons merely because the copyright position is clear.

Decide now

The Riksdag allocates SEK 80 million for 2027 and instructs PRV to build the licensing protocol and launch four voluntary sector pilots. A robust proposal for a defined extended collective licence must be submitted by 2028 at the latest.

No general three-per-cent fee will be introduced without a verified market, legal review and impact assessment. No retrospective fee will be imposed on older models without legal authority. Research, quotation, criticism, parody and the lawful use of facts must be protected.

Before the end of the parliamentary term, the system must either have made its first payment or openly report why it did not work and be wound up. It is Sonja Kovalevsky’s and Bertrand Russell’s rigour joined with Jan Stenbeck’s determination to act: count honestly, build the infrastructure and start where the law actually allows us to start.