Latvia AI Centre Law creates supervised AI sandbox powers
By Anthony Guerriero, Founder, The Leveraged Years. Published 19 March 2025. Last verified: 2026-07-25.
Most of the AI-law traffic runs through Brussels, so a statute from Riga rarely makes the headlines. It should here. On 6 March 2025 the Saeima adopted the Maksliga intelekta centra likums, the Artificial Intelligence Centre Law, and it does something few national AI statutes do in plain text: it lets a public body, within a special regulatory environment, permit an applicant not to apply specified requirements of laws in force for a set period and under safeguards, so an AI system can be built and tested. That is a live, adopted power, not a consultation paper. If you want a concrete example of a national AI regulatory sandbox written into primary law, Latvia is now one of the cleanest ones to cite.
What the law creates: a foundation and a sandbox
The law does two distinct things. First, it establishes an institution. Section 1 states that the purpose is to build an AI technology ecosystem and a legal framework for cooperation between the public sector, the private sector, and higher education, and to set up the foundation Maksliga intelekta centrs (the Artificial Intelligence Centre, referred to as the Centre). Under Section 3 the Centre is a nodibinajums, a foundation registered under the Associations and Foundations Law, and under Section 4 it sits, for delegated state tasks, under the supervision of the Ministry of Smart Administration and Regional Development.
Second, and more unusual, Section 4(1)(9) lists as a core task of the Centre "to ensure the special regulatory environment for the advancement of artificial intelligence systems." Section 8 then defines that environment. This is the sandbox, and the operative wording is worth reading exactly as adopted:
"The Centre, in cooperation with the competent authorities, shall organise a special regulatory environment in which provisions different from the existing laws and regulations may operate in order to ensure the development, testing, and advancement of artificial intelligence systems."
Latvian original: "Centrs sadarbiba ar kompetentajam institucijam organize specialo regulativo vidi, kura var darboties no speka esosiem normativajiem aktiem atskirigi noteikumi, lai nodrosinatu maksliga intelekta sistemu izstradi, parbaudi un attistibu." (Section 8(1))
Read that against the usual national sandbox, which is often a supervised advice channel with no power to set aside anything. Section 8 goes further: it contemplates rules that differ from the laws in force actually operating inside the test space.
How the sandbox works in practice
The mechanism is an administrative act. Section 8(2) says the Centre issues an administrative act to run the environment, and that this act sets out the conditions, grants permission not to apply particular requirements of the laws in force, and fixes the time period in which a person may develop and test an AI system inside the environment. The Centre must take into account the recommendations and objections of the competent authorities before it issues that act.
There is a hard safety floor. The same provision states that development and testing "may not endanger the security of the State, society, the environment, or economic activity, or the health and life of persons," and that departures from the laws in force are allowed only so far as objectively necessary to enable the development, testing, and advancement of the AI system. So the non-application power is real but bounded, and it is reviewable as an administrative act rather than exercised informally.
Who is covered, and who is shut out
Participation is open in principle to public-sector, private-sector, and academic developers, which matches the Centre's stated aim of pooling their resources. But Section 8(3) draws two hard exclusions. The special regulatory environment cannot include:
- a trader that is subject to international or national sanctions; and
- a capital company under the direct or indirect controlling influence of the Russian Federation or the Republic of Belarus, their nationals, or legal persons registered there (the law's shorthand: persons belonging to Russia or Belarus).
That security carve-out is a Baltic fingerprint. It is written into the sandbox itself, not left to a separate screening regime. Governance of the Centre reflects the same posture: under Section 7 the Centre's council has nine members, including one representative each from the Ministry of Smart Administration and Regional Development, the Ministry of Economics, and the Ministry of Defence, alongside private-sector and higher-education nominees.
How Latvia's sandbox compares with the EU AI Act model
The EU AI Act (Regulation (EU) 2024/1689) requires every Member State to set up at least one AI regulatory sandbox at national level, operational by 2 August 2026. Latvia has moved early and, notably, has legislated a stronger internal mechanism than the baseline the Act describes. The table below sets the Latvian instrument against the EU floor and an early mover.
| Feature | Latvia, AI Centre Law (Section 8) | EU AI Act baseline (Art. 57) | Spain, early pilot (RD 817/2023) |
|---|---|---|---|
| Legal basis | Primary statute adopted 6 Mar 2025 | EU Regulation, national sandboxes due by 2 Aug 2026 | Royal Decree, national pilot |
| Who runs it | State foundation (the AI Centre) with competent authorities | National competent authorities | Designated Spanish authority |
| Can it permit non-application of requirements? | Yes, by administrative act, for a set period, where objectively necessary and under safeguards | Supervised testing; no general power to set aside substantive law | Supervised testing environment |
| Hard safety floor written in | Yes: no risk to state, society, environment, economy, health, or life | Yes, via general Act safeguards | Yes, supervised |
| Security exclusions in the text | Yes: sanctioned firms and Russia or Belarus controlled entities barred | Not specified at Art. 57 | Not specified |
The point of the comparison is not that Latvia is stricter or looser. It is that Latvia wrote an explicit disapplication power and an explicit security exclusion into the statute, where the EU baseline leaves both to national implementation.
What the law does NOT do
- It is not a horizontal AI code. It does not classify AI systems by risk, ban practices, or impose conformity assessments. That job belongs to the EU AI Act. This law builds an institution and a test space.
- The sandbox does not grant blanket immunity. Section 8 permits non-application of specific requirements for a fixed time, subject to the safety floor. It does not switch off liability generally.
- It does not create a new enforcement regulator. The Centre advises, coordinates, funds, and runs the sandbox. It is not framed as the AI Act market-surveillance authority.
- One task is delayed. The Centre's role in giving opinions on AI or deepfake use during election campaigns (Section 4(1)(3)) applies only from 1 June 2026 under the transitional provisions.
- It says nothing about copyright or training-data licensing beyond the Centre building and organising datasets for AI training as one of its tasks.
Practical steps if you are watching this
- Confirm the current text. The law has already been amended (changes adopted 14 May 2026 took effect 1 June 2026). Always read the consolidated version at likumi.lv before citing a section number.
- Check the implementing rules. The transitional provisions required the Cabinet of Ministers to issue regulations under Section 8 by 31 May 2025. The sandbox conditions live partly in those regulations, not only in the statute.
- Map the security screen early. If any owner in your structure touches sanctions lists or Russian or Belarusian control, Section 8(3) closes the door before you apply.
- Track it alongside the EU deadline. Watch how Latvia's sandbox is positioned as the Article 57 obligation bites in August 2026.
Key Facts
- Instrument
- Maksliga intelekta centra likums (Artificial Intelligence Centre Law), Latvia.
- Issuer and adoption
- Adopted by the Saeima on 6 March 2025; promulgated by President Edgars Rinkevics in Riga on 19 March 2025.
- Publication and entry into force
- Published in Latvijas Vestnesis No. 55, 19 March 2025 (OP number 2025/55.1); entered into force the following day, 20 March 2025.
- What it creates
- A state foundation, Maksliga intelekta centrs (the AI Centre), under the supervision of the Ministry of Smart Administration and Regional Development, plus a Section 8 regulatory sandbox.
- Who is covered
- Public, private, and academic AI developers; sanctioned firms and entities controlled by Russia or Belarus are barred (Section 8(3)).
- Status
- In force (speka esoss), as amended by the law of 14 May 2026.
FAQ
What did Latvia's AI Centre Law create?
A state-backed foundation named Maksliga intelekta centrs (the Artificial Intelligence Centre) and, in Section 8, a special regulatory environment in which provisions different from existing laws may operate so AI systems can be developed and tested.
When did the AI Centre Law take effect?
The Saeima adopted it on 6 March 2025, President Rinkevics promulgated it on 19 March 2025, and it entered into force the next day, 20 March 2025.
Can Latvia's AI sandbox actually suspend other laws?
Within limits. Under Section 8 the Centre issues an administrative act that may permit an applicant not to apply certain requirements of laws in force for a defined period, provided the testing does not endanger public, environmental, or economic security or human health and life.
Who cannot use the Latvian AI sandbox?
Section 8(3) bars traders subject to international or national sanctions and companies under the controlling influence of Russia or Belarus or their nationals.
Primary sources
- Consolidated Latvian text: likumi.lv/ta/id/359339
- Official English translation: likumi.lv/ta/en/en/id/359339
- Official gazette publication (Latvijas Vestnesis, OP 2025/55.1): vestnesis.lv/op/2025/55.1
Related tracker entries: Portugal designates ANACOM as an AI Act market-surveillance authority, Romania names its AI Act competent authorities, and El Salvador's AI promotion law and the ANIA registry.