Part of the AI Regulation News hub.
A private member's bill registered at Romania's Senate on 28 October 2026 would create an AI Safety Authority inside ANCOM and bar frontier AI systems from operating without prior authorisation
This is a proposal at the first chamber, filed by one non-affiliated deputy, and the Senate record so far shows only requests for opinions. The text is demanding: authorisation before operation, liability cover floors starting at EUR 100 million, and prison terms. It also says it complements the EU AI Act, which is where it would be tested.
Bottom line: A bill, not law, and a private member's bill at that. It was registered at the Senate on 28-10-2026 as B597/2026 by one non-affiliated deputy. The Senate record lists opinion requests sent on 03-11-2026 and a Legislative Council opinion deadline of 03.12.2026. It shows no committee report and no vote, and nothing in the text binds anyone today.
Who this affects: Developers and operators of frontier AI systems who offer service in Romania, providers of high-capacity compute, insurers who would be asked to cover frontier risk, and counsel and public-affairs teams tracking how national proposals sit beside the EU AI Act.
Issue date: Registered 28-10-2026. Opinion requests sent 03-11-2026, including to the Legislative Council with a deadline of 03.12.2026. If a text like this were adopted, its art. 50 would bring it into force 30 days after publication in the Official Gazette, and 270 days after publication for the criminal-law articles 40-42, 44 and 45.
What changed: Nothing is in force. As filed, the bill would set up an AI Safety Authority within ANCOM; require 60 days' notice before training a system above a compute threshold of 10^25 operations (arts. 6 and 7); bar operation without prior authorisation (art. 9, decision within 90 days, licences of at most 24 months under art. 10); and add 24-hour serious-incident reporting (art. 20), shutdown capability tested at least quarterly (art. 14), provenance marking of AI-generated audio, image and video (art. 17), liability cover (art. 36) and fines of up to 10% of worldwide turnover (art. 37).
Analysis: Our reading, marked as inference: the drafters know the AI Act is the obstacle. The text says it complements the Regulation and leaves the Union's rules untouched, and it adds a simplified recognition route for systems lawfully operating in another member state. Whether a prior-authorisation gate on operation in Romania can coexist with that single-market logic is the open question, and nothing on this record tells us how far a one-deputy initiative will travel. The useful reading today is as a marker of what a frontier-safety licensing regime could contain: compute thresholds, mandatory shutdown tests, strict liability and insurance floors.
Primary sources: Senatul României, record of B597/2026 · Senatul României, text of the bill as filed by the initiator (PDF)
- Instrument (EN)
- Legislative proposal on the safe development of frontier AI systems and prevention of catastrophic risks for Romanian citizens (B597/2026). Native title: Propunere legislativă privind dezvoltarea sigură a sistemelor de inteligență artificială de vârf și prevenirea riscurilor catastrofale pentru cetățeanul român. English rendering is ours
- Authority
- Senatul României (first chamber). Initiator on the Senate record: Coarnă Dumitru, deputy, non-affiliated
- Jurisdiction
- Romania
- Status
- Registered at the Senate for debate (the record says înregistrat la Senat pt. dezbatere). The Legislative Council opinion field shows x. No committee report or vote appears on the record
- Bindingness
- Not law and binds no one. A legislative proposal that has not been adopted by either chamber on the record we read
- Issue date / next deadline
- Registered 28-10-2026. Legislative Council opinion deadline 03.12.2026
- Proposed regulator
- Autoritatea pentru Siguranța Inteligenței Artificiale, a structure without legal personality within ANCOM, led by an ANCOM vice-president appointed by Parliament for six years (art. 31)
- Proposed sanctions
- Fines by offence class from 1% to 10% of prior-year worldwide turnover (art. 37(1)); service cessation and blocking (arts. 38-39); offences and prison terms in arts. 40-45 (arts. 42 to 45 not read in full), for example 2 to 7 years for operating without authorisation (art. 40(1))
- Primary source
- https://www.senat.ro/legis/lista.aspx?nr_cls=B597&an_cls=2026
- Editorial Note
- Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.
What was filed and where it stands
The Senate's legislation page lists B597/2026 under the title above, with Coarnă Dumitru, a non-affiliated deputy, as the only initiator. The stage reads "înregistrat la Senat pt. dezbatere" and the first chamber is the Senate. The Chamber of Deputies registration field is empty.
The procedure log has two kinds of entry. On 28-10-2026 the bill was registered, with three PDFs attached: the covering letter, the initiator's text and the statement of reasons. On 03-11-2026 it was sent for opinions to the Romanian Academy, the Financial Supervisory Authority, ANCOM, the Competition Council, the Economic and Social Council, the Fiscal Council, the Legislative Council (deadline 03.12.2026), the Superior Council of Magistracy and the Supreme Council of National Defence, with a request for the Government's view.
That is the whole record. The page links one set of opinions sent by interested persons, which we did not open, and shows no committee report, plenary date or vote.
Which systems it would catch
Article 5 defines a frontier system ('sistem de vârf') by any one of three tests: training compute above a threshold, a critical capability found on evaluation, or designation by the Authority after a reasoned decision. Article 6(1) sets the compute threshold at 10^25 floating-point operations, which it says corresponds to the systemic-risk presumption in Article 51(2) of the AI Act.
Critical capabilities are listed in art. 5(2): help with chemical, biological, radiological or nuclear weapons or pandemic pathogens; autonomous cyberattacks on critical infrastructure or financial systems; autonomous replication, weight exfiltration or resource acquisition; evading oversight or resisting shutdown; and recursive self-improvement. Frontier systems fall into three categories on that basis. Category III systems, those showing the replication, evasion or self-improvement capabilities, could not be put into operation or developed further until the capability is removed.
Article 4 applies the bill to developers with a seat, an operational unit or compute infrastructure in Romania, to operators of frontier systems used in Romania or made available to people there, to compute providers, and to public authorities. A natural person using an AI system for personal purposes would have no obligations, apart from the ban on tampering with provenance marks.
The licence regime
Article 7 would require a developer expecting to train above the threshold to notify the Authority at least 60 days before training starts. The Authority could, within 30 days of a complete notice, add conditions or bar the start if the safety plan is plainly insufficient. Article 8 adds a safety plan with suspension levels, and says reaching one means training stops at once and the Authority is told within 24 hours.
Article 9(1) is the centre of the bill: no frontier system could be put into operation without the Authority's prior authorisation. The decision would fall due within 90 days of a complete application, subject to suspensions, and silence would not count as tacit approval. Authorisation would follow an independent evaluation with full access to the system, weights and training data, and would last at most 24 months under art. 10(1), renewable after a new evaluation.
Article 9(7) offers a simplified recognition procedure for systems lawfully operating in another EU member state under the AI Act, and for general-purpose models with systemic risk, the Authority would rely on documentation given to the AI Office and check only what goes beyond the Regulation. The text adds that this procedure is not a condition for placing a model on the EU market.
Operating duties, cover and penalties
Serious incidents would be reported to the Authority within 24 hours of the developer or operator knowing or being expected to know, with a full report within 15 days (art. 20(1)). Art. 14(1) would require shutdown capability to be maintained and tested at least quarterly, in real conditions, with results kept for at least 10 years under art. 14(6). Art. 17 would require provenance marking, a perceptible mark and an embedded technical mark, on AI-generated audio, image and video that could be mistaken for real material.
Developers and operators would be jointly liable without fault for damage caused by a frontier system (art. 35). Art. 36 sets minimum cover, to be fixed in implementing rules but not below the lei equivalent of EUR 100 million per event for category I systems and EUR 1 billion for category II, with an annual aggregate of at least double. It adds a fallback to lower levels, not below one fifth, if the Authority reports the cover unavailable on the market.
Fines under art. 37(1) run in bands by offence: 1% to 3%, 2% to 5%, 3% to 7%, 5% to 10% and again 2% to 5% and 1% to 3% of prior-year worldwide turnover. The top band, 5% to 10%, covers breaches of the prohibitions listed in art. 37(1)(d), among them art. 14(2), which bars operating a system configured to change its own weights or disable its oversight mechanisms. Where a firm has no turnover, art. 37(2) sets 100,000 to 10,000,000 lei for legal persons. Arts. 38 and 39 add service cessation and network blocking for operators that stay non-compliant, and arts. 40 to 45 set the offences and the rules on who answers for them and who prosecutes, with art. 40(1) at 2 to 7 years for operating without authorisation. We read arts. 40 and 41 in full and arts. 42 to 45 only in outline.
How it sits with the EU AI Act
The bill speaks to this itself. Art. 1(3) says it complements the Regulation and applies without prejudice to it, and in harmonised areas is to be interpreted in line with it. Art. 1(4) says it sets no conditions for placing general-purpose models on the Union market and leaves the Commission's powers intact. For marking duties, art. 17(6) treats the obligations as met where Article 50 of the AI Act is complied with.
Our reading, marked as inference: those clauses are drafted to keep the bill inside national competence over public safety, national security, civil and criminal liability. A reader should not assume they succeed. We have not tested the bill against the Regulation, against the e-commerce rules the bill itself invokes in art. 39(4), or against any Commission or Romanian government view, and none appears on the record we read.
The same goes for timing. If a text like this were ever adopted, art. 48 would require notification of existing systems within 60 days after the law enters into force and authorisation applications within 90 days after the implementing rules enter into force, and the rules themselves would be due 120 days after publication under art. 49. Each of those periods would run from a law that does not exist yet.
What we did not verify
We opened the Senate's record page for B597/2026 and the initiator's text of the bill (the 26b597FG PDF, 39 pages as extracted by a scraping tool). We read arts. 1, 4 to 11, 14, 17, 20, 31, 35 to 41 and 46 to 50 in Romanian, and checked each quotation and figure above against that text. The extracted text contains optical-recognition noise in places, which we ignored for meaning.
We did not read in full arts. 2, 3, 12, 13, 15, 16, 18, 19, 21 to 30, 32 to 34 and 42 to 45, the covering letter, the Legislative Council request, or the statement of reasons (the extraction is very long and we did not review it). We did not open the one set of opinions the Senate page links. We did not verify the preamble's own factual recitals, including a statement about leaders of AI laboratories on 12 September 2026, and do not repeat them as fact.
We do not claim the bill will pass, will reach a vote, or has any support beyond its initiator. We make no statement about any consultation or plenary deadline beyond the Legislative Council opinion deadline of 3 December 2026 noted above. We do not claim the bill conflicts with, or is compatible with, the AI Act.
Read this as a marker, not a deadline. A single non-affiliated deputy has put a frontier-AI licensing bill on the Romanian Senate's record, with a Legislative Council opinion due 3 December 2026, no vote shown and nothing binding anyone. If you build or serve frontier models, note the pieces that would matter if a regime like this ever travelled: a compute threshold of 10^25, authorisation before operation, quarterly shutdown tests, 24-hour incident reports and insurance floors of EUR 100 million and EUR 1 billion. This is informational analysis for working professionals, not legal advice.
Source File
https://www.senat.ro/legis/lista.aspx?nr_cls=B597&an_cls=2026
Open the Senate page for B597/2026 and confirm the title, the single initiator, the stage înregistrat la Senat pt. dezbatere, the registration entry dated 28-10-2026, the opinion requests dated 03-11-2026 and the Legislative Council deadline 03.12.2026. Then open the initiator's PDF (26b597FG) and confirm art. 6(1) (the 10^25 threshold), art. 7(1) (60 days), art. 9(1), art. 10(1) (24 months), art. 20(1) (24 hours), art. 36(1) (EUR 100 million and EUR 1 billion), art. 37(1) (bands up to 10%) and art. 50 (entry into force).
Niciun sistem de vârf nu poate fi pus în exploatare fără autorizația prealabilă a Autorității. · Propunere legislativă B597/2026, art. 9(1), registered at the Senate 28 October 2026 (original Romanian; text of a proposal, not law)
FAQ
Is Romania's B597/2026 frontier AI bill law?
No. It is a legislative proposal registered at the Senate on 28-10-2026 by one non-affiliated deputy. The Senate record shows opinion requests and a Legislative Council deadline of 03.12.2026, with no committee report or vote.
What would the bill require of frontier AI developers?
As filed, it would require 60 days' notice before training above a 10^25 operation compute threshold, prior authorisation before operation with a decision within 90 days and licences of at most 24 months, 24-hour serious-incident reports, quarterly-tested shutdown capability, provenance marking of AI audio, image and video, and liability cover.
How large are the proposed fines?
Art. 37(1) sets bands from 1% to 10% of prior-year worldwide turnover depending on the offence. The 5% to 10% band covers breaches of the absolute prohibitions listed there. The bill also proposes service blocking and prison terms for operating without authorisation.
Does the bill say how it relates to the EU AI Act?
Yes. Art. 1(3) says it complements the Regulation and applies without prejudice to it. Whether the prior-authorisation gate fits with EU single-market rules is not answered on the record we read, and we do not offer a view.
Related briefings
Sponsored Training
Practical AI training for regulated professionals, built around verification, documentation and a defensible process. See the courses.