AMENAZA ROBOTO
ARTIFICIAL INTELLIGENCE IN GOVERNMENT
ARTIFICIAL INTELLIGENCE IN GOVERNMENT
The government is implementing AI without the transparency needed to regulate its use
Monitoring, auditing, and classifying files are some of the tasks for which public agencies use or plan to implement artificial intelligence. Amenaza Roboto asked 24 institutions to identify their projects: Nine provided lists or referred to them. Some responses lacked information about scope or risk; others withheld information, and four denied access entirely.
By: Miguel Ángel Dobrich y Gabriel Farías.
Legal advisor: Matías Jackson.
September 14, 2026
This investigation was produced in partnership with the Pulitzer Center’s AI Accountability Network.
The Uruguayan social security institute, Banco de Previsión Social (BPS), uses predictive models to select taxpayers for audit. The state-owned oil, fuel, and cement company, ANCAP, relies on artificial intelligence to anticipate fuel demand and prepare purchasing decisions. The Ministry of the Interior applies AI to various public security initiatives, but let the legal deadline for responding to Amenaza Roboto’s public information request pass without response.
Artificial intelligence assists public officials in their daily tasks and can influence their decisions. To examine that influence, the public needs to know what information the system received, what it recommended, and how the official used that recommendation. However, the information available from public sources regarding public agencies’ AI developments does not always allow for a reconstruction of that process.
Amenaza Roboto submitted requests for public information to 24 agencies. It asked them for an inventory of their AI projects currently in operation, in the pilot phase, or under evaluation. For each project, it requested the name, purpose, type of development, start and end dates (if applicable), the unit responsible, the estimated number of people affected, and the risk level. Nine agencies provided lists or referred to them.
The agencies that provided lists were ANCAP, ANDE, ANII, BPS, the Central Bank, UTEC, and the Ministries of Economy and Finance, Social Development, and the Environment. Their responses make it possible to identify which government tasks involve artificial intelligence and which agencies are in charge of them.
The Central Bank reported 17 projects and applications; ANCAP detailed nine projects and applications and withheld information on other solutions; UTEC identified five projects and classified two as high-risk. The inventories make it possible to determine which tasks are intended to be automated, which areas are responsible for them, and what scope each institution reports. They also document what the agencies did not report.
In 2024, AGESIC—the regulator for digital government—and the Unit for Access to Public Information (UAIP) recommended that agencies explain what these systems are used for, how they work, and what data they process. Their algorithmic transparency guide notes that this information allows for the examination of potential biases or injustices in decision-making.
Artificial intelligence assists public officials in their daily tasks and can influence their decisions. To examine that influence, the public needs to know what information the system received, what it recommended, and how the official used that recommendation. However, the information available from public sources regarding public agencies’ AI developments does not always allow for a reconstruction of that process.
Amenaza Roboto submitted requests for public information to 24 agencies. It asked them for an inventory of their AI projects currently in operation, in the pilot phase, or under evaluation. For each project, it requested the name, purpose, type of development, start and end dates (if applicable), the unit responsible, the estimated number of people affected, and the risk level. Nine agencies provided lists or referred to them.
The agencies that provided lists were ANCAP, ANDE, ANII, BPS, the Central Bank, UTEC, and the Ministries of Economy and Finance, Social Development, and the Environment. Their responses make it possible to identify which government tasks involve artificial intelligence and which agencies are in charge of them.
The Central Bank reported 17 projects and applications; ANCAP detailed nine projects and applications and withheld information on other solutions; UTEC identified five projects and classified two as high-risk. The inventories make it possible to determine which tasks are intended to be automated, which areas are responsible for them, and what scope each institution reports. They also document what the agencies did not report.
In 2024, AGESIC—the regulator for digital government—and the Unit for Access to Public Information (UAIP) recommended that agencies explain what these systems are used for, how they work, and what data they process. Their algorithmic transparency guide notes that this information allows for the examination of potential biases or injustices in decision-making.
Anticipating demand and planning supply
ANCAP forecasts demand for diesel, gasoline, and liquefied petroleum gas using a neural network model. According to its response to Amenaza Roboto, it uses historical demand data, public policies, and market forecasts to support business and planning decisions. The initiative began in 2021, is overseen by the Digital Transformation Department and the Commercial Planning Division, and was classified by the agency as high-risk.
When asked to estimate the number of people affected, it responded, “the business in general.” ANCAP operates the country’s only refinery and has, with legal exceptions, a monopoly on the import of crude oil and petroleum products.
The use of artificial intelligence to forecast demand could have consequences for both ANCAP’s finances and the manufacturing sector, transportation, and households that depend on that supply.
The fuel demand model that ANCAP included in its inventory is also published on AGESIC’s AI Observatory. Its fact sheet details some of the factors used in the projection: rice farming for diesel, tourism for gasoline, and the weather forecast for liquefied petroleum gas.
The fact sheet indicates that the system is already in use and explains that a team makes purchasing decisions using the model’s report as an input. The extent to which the tool influences these decisions remains unspecified, both in the response to Amenaza Roboto and in the AGESIC AI Observatory fact sheet.
When asked to estimate the number of people affected, it responded, “the business in general.” ANCAP operates the country’s only refinery and has, with legal exceptions, a monopoly on the import of crude oil and petroleum products.
The use of artificial intelligence to forecast demand could have consequences for both ANCAP’s finances and the manufacturing sector, transportation, and households that depend on that supply.
The fuel demand model that ANCAP included in its inventory is also published on AGESIC’s AI Observatory. Its fact sheet details some of the factors used in the projection: rice farming for diesel, tourism for gasoline, and the weather forecast for liquefied petroleum gas.
The fact sheet indicates that the system is already in use and explains that a team makes purchasing decisions using the model’s report as an input. The extent to which the tool influences these decisions remains unspecified, both in the response to Amenaza Roboto and in the AGESIC AI Observatory fact sheet.
More state initiatives:
| Institution | Application | Responsible party | Estimated scope | Risk | Status |
|---|---|---|---|---|---|
| Evaluating support and investments | |||||
| ANDE | AI assistance to the committee evaluating applications to the Semilla program | Not disclosed | Not disclosed | Not disclosed | Being defined1 |
| MEF | Chatbot for investment eligibility screening and internal regulatory queries | COMAP | Not disclosed | Not disclosed | Not disclosed2 |
| Reviewing documents and case files | |||||
| UTEC | Review of travel-expense invoices and data entry into the management system | Not disclosed | 500 people | Not disclosed | Not disclosed |
| ANII | Review of expense-report invoices (minimum viable product) | Responsible executive decides approval/rejection | Not disclosed | Low | Minimum viable product3 |
| ANII | Validator for documents attached to applications (new version) | Not disclosed | Not disclosed | Not disclosed | In development4 |
| MEF | AI-based Document Control for the Single Window (Ventanilla Única) | COMAP | Not disclosed | Not disclosed | Not disclosed |
| ANCAP | Assistant for classifying information (public/restricted/confidential) and drafting Board resolutions | General Secretariat and Digital Transformation | 1,500 people | Minimal | Implementation project5 |
| BSE | Document image recognition (Watson-based system) | Not disclosed6 | Not disclosed | Not disclosed | Completed and in production6 |
1 For the pilot, ANDE plans human validation of all responses and the use of anonymized or simulated data. It is one of four projects the agency reported.
2 The MEF reported the project's dates and the responsible unit, but did not specify scope or risk.
3 Part of the inventory of nine applications submitted by ANII, which also includes tools that are in use, in development, and discontinued.
4 Replaces an earlier validator that ANII declared discontinued; that validator could block an application until an error was corrected.
5 A previous pilot reached two people in the General Secretariat and ended in October 2025.
6 Per its entry in AGESIC's AI Observatory (not from a direct response to the access-to-information request, which BSE denied); it runs as an automated process with no human intervention in the final decision for that document-handling task.
Auditing and supervision
The Bank of Social Provision (BPS) uses predictive models to guide the selection of taxpayers to audit. Its 2025 Audit Plan, published on the agency’s website, describes the selection of cases showing signs of tax evasion, including calls for audits of rural contributions based on estimates derived from satellite imagery, as well as the selection of construction projects for tax assessment.
The BPS informed Amenaza Roboto that it uses predictive models in “various business processes” and stated that they are in production—meaning they are in use. However, it did not identify which models or explain their functions. It withheld the requested additional data—such as start dates, scope, and risk—citing cybersecurity concerns.
The response contrasts with the AI security policy that the BPS itself shared: It requires transparency toward the public regarding automated decisions and assigns risk assessment to an internal group. Although it publishes specific uses in its audit plan, in response to the public information request, it offered only a generic description and cited cybersecurity concerns to withhold the additional data. The BPS did not explain what risk disclosing each piece of information would entail.
The BPS informed Amenaza Roboto that it uses predictive models in “various business processes” and stated that they are in production—meaning they are in use. However, it did not identify which models or explain their functions. It withheld the requested additional data—such as start dates, scope, and risk—citing cybersecurity concerns.
The response contrasts with the AI security policy that the BPS itself shared: It requires transparency toward the public regarding automated decisions and assigns risk assessment to an internal group. Although it publishes specific uses in its audit plan, in response to the public information request, it offered only a generic description and cited cybersecurity concerns to withhold the additional data. The BPS did not explain what risk disclosing each piece of information would entail.
Other institutions that use AI for auditing:
| Institution | Application | Responsible party | Estimated scope | Risk | Status |
|---|---|---|---|---|---|
| BCU | Querying and summarizing minutes of supervised entities | Superintendency of Financial Services | 100 people | Medium | Not disclosed1 |
| BCU | Assistance with the authorization process | Superintendency of Financial Services | 50 people | Medium | Not disclosed1 |
| BCU | Reports comparing contracts against regulations (outsourcing arrangements) | Superintendency of Financial Services | 50 people | Medium | Not disclosed1 |
| BCU | Aranda Virtual Agent | GG-TI, as identified by the BCU | 600 people | Medium | Not yet started |
| Ministry of Environment | Strengthening of oversight and inspection capacities | DINACEA/ACDA2 | 20 people2 | Low2 | Project built with Superset; an assistant is planned to be incorporated, pending evaluation2 |
1 The BCU reported start dates, but no individual status for these three projects. Those dates do not confirm that they are in production.
2 For the Ministry of Environment, the responsible party, scope, and risk correspond to the project as a whole, not to an independent assessment of the planned future assistant. The status reflects the response received in May 2026.
The figures are as reported by the agencies; they should not be added together as unique individuals.
Estos usos pueden implicar el tratamiento de datos personales de las personas fiscalizadas o supervisadas. Matías Jackson, abogado especializado en sistemas de información y asesor jurídico de esta investigación, advirtió que, en esos casos, los organismos deben aplicar las garantías previstas en la Ley de Datos Personales, la Nº 18.331.
La ley reconoce el derecho de las personas a "no quedar sometidas a decisiones con efectos jurídicos que las afecten significativamente cuando se basan en el tratamiento automatizado de datos para evaluar aspectos personales", señaló Jackson. Cuando ese tratamiento es el único fundamento de un acto administrativo, agregó, la persona afectada "puede impugnarlo y acceder a información sobre los criterios y el programa utilizados para decidir".
La ley reconoce el derecho de las personas a "no quedar sometidas a decisiones con efectos jurídicos que las afecten significativamente cuando se basan en el tratamiento automatizado de datos para evaluar aspectos personales", señaló Jackson. Cuando ese tratamiento es el único fundamento de un acto administrativo, agregó, la persona afectada "puede impugnarlo y acceder a información sobre los criterios y el programa utilizados para decidir".
Vigilar e identificar
El Ministerio de Transporte y Obras Públicas respondió parcialmente al pedido de Amenaza Roboto y no compartió el inventario con los campos solicitados. Según el Observatorio de IA de AGESIC, el MTOP reconoce rostros para controlar el acceso a su sede central y habilitar vehículos oficiales a sus conductores. También reconoce matrículas. El sistema reporta eventos y una persona determina el procedimiento a seguir. Abarca a los equipos de trabajo y a quienes ingresan a los edificios del ministerio.
Por su parte, el Ministerio del Interior dejó vencer el plazo previsto por la Ley de Acceso a la Información Pública y no respondió al pedido de Amenaza Roboto. Según explicó el abogado Matías Jackson, en esos casos opera el llamado “silencio positivo”: vencido el plazo, el organismo debe entregar la información.
En la práctica, sin embargo, esa consecuencia rara vez se concreta sin intervención judicial. “Este instituto rara vez se materializa en una entrega”, señaló Jackson. Los solicitantes suelen recurrir a la Justicia, donde, agregó, los jueces han entendido que la entrega no es automática y que corresponde analizar cada caso.
El Ministerio del Interior tampoco tiene desarrollos publicados en el Observatorio de IA de AGESIC, según la consulta realizada el 8 de setiembre de 2026. Sin embargo, en publicaciones en su web oficial, el Ministerio describe herramientas capaces de identificar personas, seguir vehículos, analizar conductas y orientar intervenciones policiales. Son usos especialmente sensibles por sus posibles efectos sobre la privacidad, la libertad de circulación y el trato que una persona recibe de la Policía.
Amenaza Roboto reconstruyó esos usos y proyectos de IA y reconocimiento automatizado a partir de comunicados, memorias de gestión y una compra estatal. Las fuentes permiten conocer parte de la actividad del ministerio, pero no el conjunto ni aspectos que el pedido buscaba precisar, como los responsables, las fechas, la cantidad de personas alcanzadas y el riesgo asignado a cada aplicación. El estado consignado para cada caso corresponde a lo informado por las fuentes citadas.
Por su parte, el Ministerio del Interior dejó vencer el plazo previsto por la Ley de Acceso a la Información Pública y no respondió al pedido de Amenaza Roboto. Según explicó el abogado Matías Jackson, en esos casos opera el llamado “silencio positivo”: vencido el plazo, el organismo debe entregar la información.
En la práctica, sin embargo, esa consecuencia rara vez se concreta sin intervención judicial. “Este instituto rara vez se materializa en una entrega”, señaló Jackson. Los solicitantes suelen recurrir a la Justicia, donde, agregó, los jueces han entendido que la entrega no es automática y que corresponde analizar cada caso.
El Ministerio del Interior tampoco tiene desarrollos publicados en el Observatorio de IA de AGESIC, según la consulta realizada el 8 de setiembre de 2026. Sin embargo, en publicaciones en su web oficial, el Ministerio describe herramientas capaces de identificar personas, seguir vehículos, analizar conductas y orientar intervenciones policiales. Son usos especialmente sensibles por sus posibles efectos sobre la privacidad, la libertad de circulación y el trato que una persona recibe de la Policía.
Amenaza Roboto reconstruyó esos usos y proyectos de IA y reconocimiento automatizado a partir de comunicados, memorias de gestión y una compra estatal. Las fuentes permiten conocer parte de la actividad del ministerio, pero no el conjunto ni aspectos que el pedido buscaba precisar, como los responsables, las fechas, la cantidad de personas alcanzadas y el riesgo asignado a cada aplicación. El estado consignado para cada caso corresponde a lo informado por las fuentes citadas.
| Use or project | What the agency published |
|---|---|
| Camera behavior analytics | Software to recognize movements and patterns that the ministry associates with crimes. Presented in September 2024, with an initial rollout planned across 2,000 cameras. Official source. |
| Detection of people experiencing homelessness | In June 2025 it reported that it was already using AI to detect prolonged presence and people lying down in public spaces. Operators forward the alerts to police officers, who assess the situation and coordinate assistance. Official source. |
| License plates, vehicles, and faces on highways and access points | SAIT allows vehicles to be searched by their characteristics and flags possible cloned license plates. In April 2026 the ministry described gantries being installed and pilots at toll booths combining license-plate reading with facial recognition. Official source. |
| Facial identification platform | The purchase of a platform and its maintenance was awarded in February 2020. The award confirms the acquisition; the current scope of its use remains pending a response. Official source. |
| Facial recognition at sporting events | The 2024 management report states that Sports Security uses mobile devices with facial recognition to control access. Official source. |
| Biometrics in immigration control | The Presidency reported in October 2018 the incorporation of facial biometrics along the boarding route at Carrasco Airport, from Immigration to the aircraft. Both public and private actors are involved. Official source. |
| Biometric monitoring of house arrest | In June 2024 it presented an app with facial verification, voice recognition, and GPS location. When checks go unanswered, it is designed to notify the monitoring center and potentially dispatch a patrol unit to the residence. Official source. |
| Civil identification via faces and fingerprints | In June 2026 the National Directorate of Civil Identification reported that it was evaluating ABIS biometric systems and a version of AFIS to modernize personal identification. Official source. |
| ShotSpotter | The ministry describes it as acoustic AI that detects gunshots and sends alerts to the Unified Command Center. In February 2026 it reported that it has been operating in Uruguay since 2023. Official source. |
| Classification and prioritization of 911 calls | In July 2026 it announced AI to classify calls and prioritize those requiring immediate intervention. The publication describes what the tool will be able to do. Official source. |
| Risk assessment for victims | The same July 2026 publication describes a project that will combine AI, alerts, and assessments to identify situations of vulnerability and life-threatening danger. Official source. |
| Fire detection | The 2024 management report describes a Fire Department system that recognizes columns of smoke in images and generates location alerts. It was reported operational with seven cameras. Official source. |
| Selective cellphone blocking in prisons | In November 2024 it presented an AI-based system to analyze usage patterns and block devices considered illegal. It announced a phased rollout starting with a pilot. Official source. |
The request sought to pin down the purpose, those responsible, the dates, the number of people affected, and the risk assigned to each application. The ministry did not provide that data. The publications make it possible to reconstruct part of its activity; still missing is the inventory that would show the full picture and allow scrutiny of how these tools are used.
The lack of transparency is not just a procedural issue. Carolina Aguerre, a specialist in technology governance, associate professor at the Catholic University of Uruguay and affiliated researcher at the Center of Technology and Society Studies (CETYS) of the University of San Andrés, warns that the use of AI without accessible information on how it is implemented “can erode public trust in government institutions.”
Teaching and evaluation
The Technological University of Uruguay (UTEC) is experimenting with TutorIA as a way to support Physics I students and develop an assistant to recommend learning tools, which is still in trial stages. It classified both projects as “high-risk” in the inventory of five initiatives it submitted to Amenaza Roboto.
For TutorIA, UTEC estimated a reach of about 100 people and identified the Bachelor’s Program in Food Analysis and the Center for Digital Transformation as the responsible entities. For the educational technology assistant, it also reported a reach of about 100 people and cited responsibility to the Center for Digital Transformation.
Ceibal, which leads the implementation of technology in public education, responded to the inventory request with links to its projects page and the AGESIC survey. These publications provide information on applications used for various educational tasks.
Ceibal generates work feedback through a virtual tutoring service on its learning platform CREA, aimed at teachers, teacher education students, and administrative teams. The course instructor must approve or modify these suggestions. The service also includes a chatbot with which course participants interact directly, according to the publication by AGESIC’s AI Observatory.
Another application is PowerBuddy, integrated into CREA to support lesson planning. The National Administration of Public Education (ANEP) announced its availability in June 2025. Based on the teacher's instructions, it proposes activities and content that the teacher must review, adapt, and validate.
Ceibal also seeks to automate part of the oral assessment for its English program. The report published by the Observatory, based on information reported in March 2024, describes an ongoing project—developed in collaboration with the Faculty of Engineering at the University of the Republic—that uses audio recordings while retaining human involvement in the final decision.
For TutorIA, UTEC estimated a reach of about 100 people and identified the Bachelor’s Program in Food Analysis and the Center for Digital Transformation as the responsible entities. For the educational technology assistant, it also reported a reach of about 100 people and cited responsibility to the Center for Digital Transformation.
Ceibal, which leads the implementation of technology in public education, responded to the inventory request with links to its projects page and the AGESIC survey. These publications provide information on applications used for various educational tasks.
Ceibal generates work feedback through a virtual tutoring service on its learning platform CREA, aimed at teachers, teacher education students, and administrative teams. The course instructor must approve or modify these suggestions. The service also includes a chatbot with which course participants interact directly, according to the publication by AGESIC’s AI Observatory.
Another application is PowerBuddy, integrated into CREA to support lesson planning. The National Administration of Public Education (ANEP) announced its availability in June 2025. Based on the teacher's instructions, it proposes activities and content that the teacher must review, adapt, and validate.
Ceibal also seeks to automate part of the oral assessment for its English program. The report published by the Observatory, based on information reported in March 2024, describes an ongoing project—developed in collaboration with the Faculty of Engineering at the University of the Republic—that uses audio recordings while retaining human involvement in the final decision.
High, medium, or low risk—depending on who’s answering
Knowing what tool an agency uses is just the beginning of citizen oversight. Carolina Aguerre warns that the “terms of use and implementation of said system” must also be examined, as well as the risks that these represent for government responsibilities and for citizens.
The agencies did not adopt a common evaluation methodology. Instead, each project is labeled according to how each agency assesses them; there isn’t an established way to put them in order according to their risk.
The AGESIC and UAIP guide recommends documenting where data comes from, who is involved in its development, and what groups are affected by the system. It calls for specifying the extent of automation and which actions still require a human being. It also proposes keeping records of decisions and detailing the audits, impacts, and risk management process, with special attention to human rights.
Some responses do not allow us to determine how many people are affected. Some count internal users; others, individuals or organizational units. Yet these figures cannot be added together to obtain a total number of affected citizens.
The agencies did not adopt a common evaluation methodology. Instead, each project is labeled according to how each agency assesses them; there isn’t an established way to put them in order according to their risk.
The AGESIC and UAIP guide recommends documenting where data comes from, who is involved in its development, and what groups are affected by the system. It calls for specifying the extent of automation and which actions still require a human being. It also proposes keeping records of decisions and detailing the audits, impacts, and risk management process, with special attention to human rights.
Some responses do not allow us to determine how many people are affected. Some count internal users; others, individuals or organizational units. Yet these figures cannot be added together to obtain a total number of affected citizens.
The information obtained by Amenaza Roboto through freedom of information requests does not match what the agencies publish in the Observatory on Artificial Intelligence in the State. The latest published update from the Observatory is from April 2025.
To understand which systems the government uses, where they are applied, and to what extent, we still need to build the puzzle piece by piece to see the bigger picture. There is currently no public source that provides a comprehensive and up-to-date overview of how the Uruguayan government is incorporating artificial intelligence.
To understand which systems the government uses, where they are applied, and to what extent, we still need to build the puzzle piece by piece to see the bigger picture. There is currently no public source that provides a comprehensive and up-to-date overview of how the Uruguayan government is incorporating artificial intelligence.
For more details on this research, see the Amenaza Roboto repository at GitHub. There you will find the methodology and supporting documentation for the article.
Amenaza Roboto