AI Regulation Tracker  /  Advisory opinion on a draft bill

A Competition Regulator Told Its Legislature to Put Algorithmic Collusion in the AI Law

On March 12, 2026 the Dominican Republic's competition authority filed a formal competition-advocacy report with the Senate on the draft Organic Law on Artificial Intelligence Systems. It set out ten lettered recommendations, including a statutory definition of colusion algoritmica and criminal exposure for anyone who manipulates an AI system to run a cartel, and annexed a matrix with the redrafted article text. Nothing here is in force. The report is advisory and the bill is a draft.

By Anthony Guerriero  ·  Reviewed by Anthony Guerriero  ·  Published August 3, 2026  ·  Updated August 3, 2026  ·  Advisory opinion on a draft bill

Key facts

Regulatory briefing

Instrument

Public competition-advocacy recommendation report PR-EXT-PROCOMP-2026-0077 on draft bill 00495-2025-PLO-SE

Authority

PRO-COMPETENCIA, Consejo Directivo, addressed to the Senate of the Dominican Republic

Jurisdiction

Dominican Republic

Status

Advisory opinion issued and stamped received. The bill it addresses is a draft

Bindingness

Non-binding. It is abogacia de la competencia under Article 31 letter n) of Ley 42-08 and imposes no obligation on any regulated party

Effective date

None. No duty attaches today, and none will attach unless and until the bill is enacted in some form

Primary source

PRO-COMPETENCIA report PR-EXT-PROCOMP-2026-0077, March 12, 2026

What the regulator actually filed, and what it is not

Start with the status, because everything else depends on it. The Proyecto de Ley Organica que regula los sistemas de Inteligencia Artificial, file 00495-2025-PLO-SE, was introduced in the Senate on February 26, 2025 by Omar Leonel Fernandez Dominguez, senator for the Distrito Nacional. It was referred to the Comision de Educacion Superior, Ciencia y Tecnologia. It was still a bill in March 2026, when the competition authority wrote to the Senate asking that its recommendations be weighed in the process of study and eventual approval, and we have found no published record of its adoption since. Nothing in this entry creates a duty for anyone, today or on any known future date.

What did happen is procedural and, for a compliance reader, genuinely useful. On March 12, 2026 the Comision Nacional de Defensa de la Competencia, known as PRO-COMPETENCIA, filed a formal written opinion on that bill with the Senate. The document carries reference PR-EXT-PROCOMP-2026-0077 and is signed by Maria Elena Vasquez Taveras as Presidenta del Consejo Directivo. It is addressed to Senate President Ricardo De los Santos Polanco, for the attention of legislative secretary general Jose Carrasco, with copies to Senator Fernandez as the bill's sponsor and to Senator Rafael Baron Duluc Rijo, who chairs the committee holding the file. The Senate's correspondence stamp on page one reads Fecha 12/3/26 Hora 2:20pm, with barcode CI202603829.

Get the shape of the document right before you cite it. The published PDF is nineteen pages. The letter itself is twelve, numbered in the footer from Pagina 1 de 12 to Pagina 12 de 12, and it is followed by an unnumbered annex titled Matriz de recomendaciones de articulos al proyecto de Ley organica que regula los sistemas de Inteligencia Artificial. So a citation to page 11 means page 11 of the letter, which is also page 11 of the PDF, and the matrix begins at PDF page thirteen. Anyone quoting a page number from this file should say which of the two they mean.

The legal basis matters for how much weight to give it. This is abogacia de la competencia, competition advocacy, exercised under Article 31 letter n) of Ley General de Defensa de la Competencia No. 42-08 and Article 12 of its Reglamento de Aplicacion No. 252-20. Article 31 letter n), quoted in the report's own footnote, lets the Consejo Directivo carry out advocacy through recommendation reports and defend competition during the process of forming laws and other normative instruments whose effects may bear on competition. That power does not let the agency amend a bill, and it does not oblige the legislature to do anything. The report guides no regulated party and sanctions no one. Read it as a well-documented position from a public body with statutory standing to hold that position.

Its structure is four parts, lettered A through D: the legal basis for the advocacy power, the context and objectives of the bill, the analysis and its impact on competition, and conclusions and recommendations. The operative content is a set of ten lettered recommendations at pages eleven and twelve, and the annexed matrix then maps each one to the article of the bill it would change and, for most of them, sets out the proposed replacement text.

The gap the agency says the bill has

The analytical core sits at paragraph twelve, on page four. PRO-COMPETENCIA states that from its reading and analysis of the initiative it has found that the bill does not contain a regulation covering concerted practices and anticompetitive agreements originating through artificial intelligence systems. That is the whole complaint, and the ten recommendations are the remedy the agency proposes for it.

The theory of harm is not homegrown. At paragraph thirteen the report adopts the OECD's three routes to collusion by algorithm: algorithms that make explicit cartels stable, because automated pricing on observable data detects and punishes deviation; hub-and-spoke configurations, where several competing firms use the same third-party pricing software and that shared tool becomes a channel for information exchange; and autonomous algorithmic tacit collusion, where self-learning systems settle on a non-competitive outcome with no shared information and no explicit coordination. The footnotes cite the OECD's 2023 background note on algorithmic competition and a 2023 Latin American analysis of algorithms and collusion. Paragraph fourteen adds that comparative law already records sanctions against firms that used software and algorithms to agree discounts, coordinate price changes, split markets and customers, and align commercial conditions.

The two most quotable proposals are the definition and the criminal limb. Both appear on page eleven of the report:

B) Adicionar en el articulo 2 de 'Definiciones' la definicion de 'Colusion algoritmica', que podria establecerse como: 'Se refiere al uso de algoritmos para establecer mecanismos de coordinacion, intercambio de informacion y acciones anticompetitivas entre empresas competidoras, con la finalidad de eliminar la competencia'.

G) Modificar el articulo 58 sobre 'Manipulacion o alteracion' para incorporar el siguiente supuesto: 'La persona que manipule o altere cualquier sistema de Inteligencia Artificial (IA) con el objeto de facilitar la ejecucion de acuerdos colusorios o el intercambio de informacion estrategica entre competidores sera sancionada con reclusion y multas segun lo establecido en la presente ley'.PRO-COMPETENCIA, PR-EXT-PROCOMP-2026-0077, page 11 of 12, March 12, 2026

In English, and the translation is mine: recommendation B would add to the bill's definitions article a definition of algorithmic collusion, proposed as the use of algorithms to establish mechanisms of coordination, exchange of information and anticompetitive actions between competing firms, for the purpose of eliminating competition. The annexed matrix shows where it would sit: as numeral 6 of Article 2. Recommendation G would amend Article 58 of the bill, on manipulation or alteration, so that a person who manipulates or alters any AI system in order to facilitate the execution of collusive agreements or the exchange of strategic information between competitors would be punished with reclusion and fines as established in the law.

Note what recommendation G does not say. It does not name a prison term and it does not name a fine amount. The proposed wording defers to whatever the enacted law ends up providing, and the bill's own penalty scale is not reproduced in the report. Anyone quoting a number for a Dominican AI cartel offence is quoting something that is not in this document.

Note also the drafting target of recommendation B. It reaches the use of algorithms to establish coordination mechanisms and to exchange information between competitors. That is wider than the tacit-collusion-by-learning-agent scenario that gets the conference-panel attention, and the report's own hub-and-spoke passage tells you why: it covers the far more common fact pattern where competing firms feed data into a shared pricing tool and the tool does the coordinating. If a definition in those terms were enacted, the compliance question for a pricing vendor stops being whether the algorithm intended anything and becomes whether competitors' information passed through it.

The ten recommendations, and what each would do

The remaining eight recommendations are worth reading as a set, because together they describe an institutional design rather than a slogan.

Recommendation A is housekeeping with teeth: add Ley 42-08 to the bill's Vistas, the list of instruments the law recites, in line with the Manual de Tecnica Legislativa. That is how a Dominican statute signals which other statutes it sits alongside, and the matrix records that the bill as filed contains no Vistas at all. Recommendation C would add a numeral 9 to Article 4, which the bill titles prohibited purposes or uses of AI systems, banning systems for coordinating anticompetitive conduct or practices. Recommendation F would list algorithmic collusion in Article 55 as an infraccion administrativa muy grave, the most serious administrative category the bill uses; the matrix puts it in as numeral 5 of the muy graves list.

Recommendation D is the one I would watch. It would amend Article 30, numeral 6, on the powers of the proposed CENIA, so that where there are indicia of the use of AI systems for collusive practices, the matter is referred to PRO-COMPETENCIA for investigation and sanction under Ley 42-08. That is a referral pipeline from the AI regulator the bill would create to the competition regulator that already exists, and the report justifies it on jurisdictional grounds: investigating and sanctioning anticompetitive practices is the competition authority's job under Ley 42-08. Recommendation E completes the loop by adding prevention of algorithmic collusion as an evaluation criterion in Article 32, aimed at systems built to facilitate price or discount fixing, market or customer allocation, and the alignment of commercial strategies. Taken together, D and E would make the AI supervisor a screening layer feeding the antitrust supervisor, which is a materially different arrangement from two agencies operating in parallel and finding out about each other's cases in the press.

The last three recommendations are not about collusion at all. They are about the bill's own regulatory burden, and they are the reason I would not read this report as a pure enforcement expansion. The report's complaint, at paragraphs twenty-three and twenty-four, is that Article 31 sends every system that may impact fundamental rights to CENIA for prior evaluation and approval, that the criterion is broad enough to catch a customer service chatbot or a content recommender, and that the resulting uncertainty and cost work as an entry barrier favouring incumbents with large legal departments. Recommendation H therefore asks the legislature to delimit more precisely which systems require prior evaluation, adopting a tiered risk approach expressly modelled on the EU AI Act: minimal risk with no specific requirements, limited risk with transparency obligations, high risk with strict conformity assessment before going to market.

Recommendation I asks for maximum deadlines on CENIA's Article 31 evaluations and the Article 48 certifications, giving thirty days for medium-risk systems and sixty for high-risk as the example, with positive administrative silence once the documentary requirements are met, meaning approval is deemed granted if the regulator does not answer in time. The annexed matrix turns that into draft text and is more precise than the letter: thirty and sixty business days from receipt of a complete application, and a separate thirty business day limit on Article 48 certifications. Recommendation J asks that the algorithmic transparency duty in Article 25 be graduated by risk, with general information about purpose and basic functioning for low-risk systems, more detail for high-risk ones, and sensitive material sent to CENIA under confidentiality agreements rather than published, so that mandatory disclosure does not hand a competitor the design.

A competition agency asking for deadlines, deemed approvals and trade-secret protection is behaving consistently with its mandate rather than against it. Slow, discretionary licensing is a barrier to entry, and barriers to entry are a competition problem; the report says as much and cites the OECD's 2012 recommendation on regulatory policy for the proposition that administrative procedures without time limits can become entry barriers even when the underlying rule is justified. Whether thirty and sixty business days are realistic for a body that does not yet exist is a different question, and the report does not answer it.

The annex goes further than the letter: tying, data access, killer acquisitions

Three further competition concerns run through the body of the report, and two of them reach the annex as proposed statutory text even though they are not among the ten lettered recommendations. If you read only the lettered list, you will miss them.

The first is algorithmic tying and bundling, at paragraph twenty-one. The concern is that a firm dominant in one market uses AI to extend that power into adjacent markets through data leverage and unjustified technical barriers, so the report recommends an express prohibition on using AI systems to impose algorithmic tying or bundling that forces users to acquire complementary software and limits interoperability with smaller competitors. The matrix drafts it as a new numeral 10 of Article 4, covering algorithmic tying or bundling that obliges users to acquire unwanted complementary products or services, limits interoperability with competitors' solutions, or restricts consumer choice.

The second is access to essential inputs, at paragraph twenty-two, and it is the most consequential thing in the document that nobody will read. The matrix proposes an entirely new article on interoperability and access to non-personal data. Operators of AI systems classified by CENIA as systemically important would have to maintain open, documented APIs permitting interoperability with third-party systems and competition in secondary markets. Where a non-personal dataset constitutes an essential input for developing AI solutions, PRO-COMPETENCIA, in coordination with CENIA, could set fair, reasonable and non-discriminatory access conditions. Systemically important operators would be barred from restricting user data portability or imposing unjustified technical barriers to migration. That is an essential-facilities regime for training data, drafted by the competition agency, sitting in an annex to a letter.

The third is at paragraph thirty and is the more novel one for a small jurisdiction. The report discusses adquisiciones asesinas, killer acquisitions, of AI startups that record modest turnover but hold algorithms with significant disruptive potential, and which therefore escape merger review built on turnover thresholds. It observes that jurisdictions including the European Union, the United Kingdom and the United States have begun to implement or discuss special review mechanisms for technology-sector acquisitions that go beyond the traditional criteria. Note where the recommendation actually lands: at paragraph thirty-one the agency asks the Dominican legislator to bear the problem in mind when evaluating future reforms to the competition framework and its merger control mechanisms, not to fix it in the AI bill. It proposes no threshold in the lettered recommendations and none in the matrix.

What all of this tells you is that the agency imported an existing analytical frame rather than inventing one. The EU AI Act is cited by name and by Official Journal reference for the tiered risk approach, the OECD supplies both the collusion taxonomy and the regulatory-delay argument, and the merger passage borrows from the EU, UK and US debate. That is usually the right instinct, and it also tells you what the resulting law would look like if the recommendations were adopted wholesale.

What to do Monday

For most US readers the honest answer is: file it, do not act on it. There is no duty. There is no date. The instrument that would create duties is a bill that has been with a Senate committee since February 2025, and I have no basis for predicting whether it moves.

That said, there are four concrete things worth doing if you have any Dominican exposure.

First, if you sell or operate pricing software in Dominican markets, read recommendation B and ask whether your product would meet that definition on its current architecture. The question to answer internally is narrow and answerable: does information from competing customers ever influence the output shown to another customer, directly or through a shared model? You want that answer written down before a definition like this is enacted anywhere, not after. The same question is being asked in US hub-and-spoke pricing litigation, which is why this one is worth ten minutes even if you never sell a licence in Santo Domingo.

Second, if you are advising on an acquisition of a Dominican or Dominican-facing AI business, note that the competition authority has publicly flagged sub-threshold AI acquisitions as a concern and has tied that flag to future amendment of Ley 42-08 rather than to this bill. That changes no filing obligation today. It does change the tone of a conversation you may end up having, and it tells you where to look when Ley 42-08 reform surfaces.

Third, if you are tracking this bill, track the committee rather than the plenary, and track the annex rather than the letter. The file sits with the Comision de Educacion Superior, Ciencia y Tecnologia. The matrix annexed to the report is the document a committee would work from if it decided to adopt any of this, because it already contains article-ready text. Watching for that text to surface in a committee report is a cheaper signal than watching for a vote. It is also worth checking the Senate's own publications page, which carries a proyectos perimidos list, since that is where a file that quietly lapses shows up.

Fourth, and this is the part that generalises beyond the Dominican Republic: check whether your own regulators have filed comparable advocacy opinions on AI bills in the markets you care about. Competition authorities file these routinely, they are public, they are specific, and they are almost never covered. A document naming article numbers and proposed statutory text tells you more about where a law is heading than a ministerial press release does. This one happens to be from Santo Domingo. The pattern is not.

One caution on how to describe all of this internally. The Centro Nacional de Inteligencia Artificial is a body the draft bill would create, and every reference to CENIA in this entry is a reference to a proposal, including the powers the matrix would give it. Nor should this report be attributed to INDOTEL, to OGTIC or to the national AI strategy apparatus. It is the competition authority speaking, on competition grounds, about a bill that is not its own.

Last verified: August 3, 2026 against the primary sources listed above.