From Data to Implementation: Scaling Digital Tools for Human Rights Monitoring

Human rights implementation increasingly depends on the quality of information systems, as states must manage thousands of recommendations from UN mechanisms, regional bodies, and other processes. These recommendations need to be clustered, assigned, followed up on, reported on, and translated into concrete actions, which is an enormous administrative challenge. Digital solutions have become increasingly important in overcoming these challenges.

Domenico Zipoli noted that states receive thousands of recommendations per year from UN human rights mechanisms, regional bodies, and other processes, all of which need to be clustered, assigned, followed up, reported on, and translated into concrete actions . He further emphasised that digital solutions are the primary means by which human rights monitoring is overcoming these challenges .

While a growing ecosystem of digital tools exists, it remains fragmented with many tools operating in silos and data that is not always interoperable. Civil society and human rights defenders do not always have equal access to human rights information and the digital space, particularly compared to those in Europe. Technical innovation does not automatically translate into better human rights outcomes.

Zipoli described the ecosystem as fragmented, with many tools still operating in silos and data not always interoperable . He specifically highlighted that civil society and human rights defenders do not always have equal access to human rights information and the digital space, and that technical innovation does not automatically translate into better human rights outcomes .

AI and machine learning offer significant potential for human rights monitoring by helping to process the large volumes of information generated by multiple treaty bodies reviewing governments in a single year. These tools can identify patterns, cluster recommendations, support semantic search, and facilitate multilingual access. However, the design of such tools must be especially careful given the importance of legal nuance, institutional mandates, and the voices of affected communities.

Zipoli explained that AI and machine learning can help process large volumes of information, noting that governments are sometimes reviewed by multiple treaty bodies in one year . He listed specific capabilities including identifying patterns, clustering recommendations, supporting semantic search, facilitating multilingual access, and helping users identify gaps . He also stressed that context, legal nuance, institutional mandates, and the voices of affected communities all matter in human rights implementation .

The key question for the session is not simply whether AI can make human rights monitoring faster, but how digital tools and responsible AI can make implementation more effective, inclusive, transparent, and accountable. AI can support human judgment but should not replace it, a principle consistently heard throughout discussions at AI events in Geneva.

Zipoli framed the central question of the session as how digital tools and responsible AI can make human rights implementation more effective, inclusive, transparent, and accountable, rather than merely faster . He noted that AI can support human judgment but should not replace it, and that this principle is widely echoed in discussions at AI events in Geneva .

Not all member states have established National Mechanisms for Implementation, Reporting and Follow-up (NMRFs), but one pathway to establishing them is through digitising tracking work. By using digital tracking tools, governments are compelled to identify the different key implementing actors within the digital interface, which in turn helps to institutionalise the work of different government bodies.

Zipoli observed that digitising work through tracking tools can help establish an NMRF because, regardless of a government's existing structure, the digital interface requires the identification of different key implementing actors .

Digital tools are not neutral containers, and the design choices embedded in them from the outset will eventually shape how human rights recommendations are implemented. This means that attention to the design of these tools must be absolutely rigorous, as the tool's architecture will influence the entire implementation process downstream.

Zipoli stated that digital tools are not neutral containers and that the design of digital tools from the outset may eventually influence how recommendations are being implemented, meaning the attention on design must be ironclad . He also speculated that the shift towards continuous digital reporting could make the concept of periodic reporting increasingly obsolete within five to ten years .

States face recommendations from a wide range of international and regional bodies simultaneously, creating a fragmented picture that is very difficult to manage coherently. Brazil serves as a concrete example, receiving recommendations from the Inter-American Commission, the Inter-American Court, the Human Rights Council, the UPR, treaty body committees, and multiple special procedures. The digital space can help bring coherence to this fragmented system by showing that recommendations from different bodies often point in the same direction.

Boyer described Brazil as a country that receives recommendations from the Inter-American Commission, the Inter-American Court, the Human Rights Council, the UPR, committees of experts, and special procedures, with five special rapporteurs visiting in a single year . She noted that from a human perspective it is very difficult to see coherence in this fragmented system, but that digital tools can help reveal that the recommendations are actually largely consistent with one another .

OHCHR developed the National Recommendations Tracking Database (NRTD) in response to the challenge of helping states move beyond receiving guidance to actually implementing recommendations. The starting point was the Universal Human Rights Index, an online platform aggregating all UN system outcomes, but states needed a tool that could track the trajectory of recommendations all the way through to change at the community level. The NRTD was designed to fill this gap by bringing all relevant recommendations together in one place.

Boyer explained that OHCHR first developed the Universal Human Rights Index as an online platform where anyone can access all outcomes of the UN system, searchable by country, theme, group, or SDG . She then described how states needed to move from standard-setting to implementation, which required a tracking tool to follow recommendations all the way to change at the community level, leading to the development of the NRTD .

The NRTD enables states to cluster recommendations by theme, SDG, group, and ministry, which is essential for translating international guidance into concrete government action. By identifying which ministry leads implementation, which co-leads, and which supports, the tool embeds accountability into the tracking process. The reporting functionality then allows ministries and entities to document what has been done to implement recommendations.

Boyer described how the NRTD allows countries to cluster recommendations per theme, per SDG, per group, and per ministry, enabling civil servants to see exactly what is on their table to implement and whether they are leading, co-leading, or supporting . She emphasised that clustering and assigning responsibility is the foundation of accountability, followed by reporting on what has been done .

The NRTD has demonstrated clear demand, with 20 countries actively using it and approximately 40 more waiting for deployment. However, OHCHR lacks the capacity to accompany all waiting states through the deployment process. The tool is continuously improved through feedback from states and users.

Boyer stated that 20 countries are currently using the NRTD and approximately 40 more are waiting for the tool to be deployed in their country, but that OHCHR lacks the capacity to accompany them all . She also noted that receiving feedback from states and users is key to improving the tool, and that OHCHR works on improvements every day with the different states it accompanies .

AI assistance is being integrated into the NRTD to help governments navigate the large volumes of data that exist and identify what is relevant for measuring progress on human rights recommendations. AI can help spot relevant data, assess what has already been collected, and identify gaps where additional data collection is needed. This is done while maintaining human control over the process, in line with the UN's position.

Boyer described how AI assistance is provided within the digital tracking tool to help governments look at the recommendations they need to follow up, identify the data available to report on progress, and determine what additional data needs to be developed . She noted that the UN's position is to keep humans in control while using AI to scale up work and spot relevant data .

The human component behind digital tools is not merely a rhetorical addition but an essential requirement. AI cannot create reality or replace the knowledge of people working directly with communities who understand subject matters and can enter, analyse, and reflect on data. Some tools have shown limitations precisely because there was no human component to feed information into the system.

Boyer argued that some tools have shown real limitations because there was no human component behind them to feed information into the system, and that AI is not going to create a reality that affects people every day . She stressed that people working with communities who know the subject matters are needed not only to enter data but to analyse it and reflect on what works in a given country context, which AI cannot currently do .

While data exists in many contexts, there is a significant gap in disaggregated data that would make invisible populations visible. The challenge is not only the absence of data but also the sheer volume of existing data and the difficulty of processing it to identify what is relevant. AI can help address this by spotting relevant data and identifying where additional collection is needed.

Boyer acknowledged that while data exists, there is a definite lack of disaggregated data because the goal is to make invisible populations visible, and the challenge is processing the large amounts of data that do exist . She described AI as playing a key role in spotting relevant data, measuring progress, and identifying what additional data needs to be collected .

States, particularly small island developing states and least developed countries, face significant data gaps and limited capacity to collect data, meaning they need guidance on the minimum set of data required to track progress on human rights recommendations. OHCHR has begun research starting with recommendations from the Committee Against Torture to identify this minimum data set. Collaboration with SDG indicator frameworks, which are outcome-focused, could help establish this baseline.

Boyer described OHCHR research beginning with recommendations from the Committee Against Torture to identify the minimum set of data that states need to collect, noting that small islands and least developed countries face huge data gaps and capacity challenges . She suggested that SDG indicators, which are all about outcomes, could be leveraged to work towards this minimum data set needed to track progress .

One of the persistent challenges in managing human rights recommendations is the high rate of staff rotation in government ministries, which means that institutional knowledge about recommendations and tracking processes is frequently lost when personnel change. This makes it difficult to maintain continuity in implementation efforts. Costa Rica has responded by introducing a constant training cycle to address this challenge.

Cespedes noted that staff rotation tends to be quite high in some government institutions, with people staying for one or two years, learning how to manage the tools, and then leaving, causing that knowledge to be lost . He described how Costa Rica has incorporated a constant training cycle for people at the Ministry of Foreign Affairs and all ministries on the NRTD to address this challenge .

Costa Rica introduced the NRTD with the help of OHCHR after previous tracking models proved insufficient, and has found it to be a flexible and intuitive tool. The NRTD has given significant visibility to recommendations from treaty bodies, the UPR, and special procedures, and has helped institutions understand where they have opportunities to implement recommendations. Ministries of Education, Health, and the institution for child safety have all found the tool's functionalities useful.

Cespedes described how Costa Rica introduced the NRTD with OHCHR's help after earlier models proved inadequate, and found it to be a flexible and intuitive tool . He noted that the Ministries of Education and Health, as well as the institution for child safety, have found the NRTD's functionalities quite useful .

One of the most practically valuable functionalities of the NRTD for Costa Rica has been the ability to cluster recommendations by topic or theme. This has led entities to recognise that certain recommendations create opportunities for collaboration with other institutions, thereby increasing communication between ministries and ultimately leading to more coordinated action.

Cespedes highlighted the clustering functionality as particularly useful, noting that entities have said they can see that a given recommendation would enable them to work together with other institutions . He observed that this increases communication between institutions and eventually leads to action .

Given the high rate of staff turnover in government ministries, Costa Rica has made continuous training on the NRTD a core part of how it manages the tool. This ensures that new staff can quickly become proficient and that institutional knowledge is not lost when personnel change. The training cycle covers both the Ministry of Foreign Affairs and all other relevant ministries.

Cespedes explained that Costa Rica has introduced a constant training cycle for people both at the Ministry of Foreign Affairs and at all ministries on the NRTD, recognising that rotation tends to be quite high and that knowledge is otherwise lost when staff leave after one or two years .

Costa Rica is planning to extend access to the NRTD to civil society organisations, moving beyond transparency to make civil society an partner in implementation and monitoring. Rather than merely observing from a distance, civil society would be able to see what every part of the state is doing, make suggestions, and work together with the state on proper implementation. Civil society already has institutionalised representation in Costa Rica's Inter-Institutional Commission through two permanent seats.

Cespedes described Costa Rica's plans to allow civil society to access the NRTD in the future, so that they are not only improving transparency but making civil society a partner that builds on implementation and monitoring . He noted that civil society would be able to see what every part of the state is doing, make suggestions, and work together with the state for proper implementation . He also mentioned that civil society already has two institutionalised seats at the table in the Inter-Institutional Commission .

A significant amount of government activity that constitutes compliance with human rights recommendations goes unreported simply because officials do not recognise the connection between their actions and specific recommendations. AI could help bridge this gap by analysing inputted data and identifying whether government actions correspond to recommendations, thereby surfacing compliance. This could then be extended to community-level monitoring.

Cespedes suggested that a lot of things states do go unreported because officials do not know they are doing it, and proposed that AI could analyse inputted data to determine whether a government action was generated by a recommendation and whether it complies with it . He noted that many people do not know they are already implementing recommendations, and that this could be extended to community-level models to monitor what is happening on the ground .

Moving beyond tracking government inputs and actions, continuously monitoring community satisfaction and linking it to what governments are doing on the ground represents a promising future direction for measuring human rights outcomes. This approach would make it easier and cheaper to monitor whether government actions are actually making a difference at the community level. There are already established models for how to properly conduct such satisfaction monitoring.

Cespedes argued that constantly asking people what they think is key and something that needs to be improved upon, noting that there are many models for how to properly do this . He suggested that it will probably become easier and cheaper to start monitoring satisfaction and then linking it up with what is being done on the ground .

Academic projects leveraging natural language processing and generative AI can build tools that enable paragraph-level search across large human rights databases, making research far more efficient than traditional document-level keyword searches. When a keyword search returns hundreds or thousands of results requiring individual tabs to be opened, paragraph-level search dramatically reduces the time needed to find relevant information. This approach has been applied to databases of UN treaty body jurisprudence, general comments, special procedures reports, and European Court of Human Rights judgments.

Szoszkiewicz described building a database originally covering general comments that expanded to jurisprudence of UN treaty bodies and thematic reports of special procedures, as well as a European Court of Human Rights database allowing paragraph-level search . He noted that when a keyword search returns 200 or 1,000 results requiring all tabs to be opened, it is not efficient, and paragraph-level search makes the process much more efficient .

Rather than uploading datasets directly to large language models for analysis, a more reliable approach is to use AI to generate deterministic code scripts that perform the analysis. These scripts work exactly the same way every time, eliminating the probabilistic variability and hallucination risk associated with direct LLM use. The generated code can also be handed to IT professionals for inspection of cybersecurity and privacy issues before deployment.

Szoszkiewicz explained that instead of uploading datasets to language models or browser versions of tools like ChatGPT, one can generate scripts that work deterministically and deploy them online, ensuring they work exactly the same every time with no room for hallucination . He summarised this as using LLMs to build tools to analyse data rather than using LLMs to analyse data directly, with the code then inspectable by IT professionals for cybersecurity and privacy issues .

While hallucinations cannot be entirely eliminated from AI systems, careful design choices can reduce them to very low levels. One effective approach is to use large language models for semantic similarity while invoking source paragraphs verbatim rather than generating new text. Studies testing different AI models in legal and non-legal domains show significant differences in hallucination rates, and the reduction achievable through careful design is substantial.

Szoszkiewicz noted that there are great studies testing different AI models in legal and non-legal domains, showing a huge difference in hallucination rates and how persuasive hallucinations are . He described his own approach of using large language models for semantic similarity while invoking paragraphs verbatim without changing them, so that everything is a decision of the designer and hallucinations can be reduced to a very low level .

The traditional software development process required domain experts to hand technical specifications to IT teams, wait for development, and then test and iterate, which was often slow and costly. Generative AI has dramatically reduced the burden of code generation, allowing domain experts to build functional prototypes directly and test them iteratively. This means much more time can be invested in the decision-making layer, where domain expertise is most valuable.

Szoszkiewicz described the traditional software development process as involving a decision layer, an execution layer where IT teams generated code, and a delivery layer for testing and tweaking, which was often very long and costly . He explained that with generative AI, code generation has become extremely reliable and easy, allowing domain experts to develop full applications rather than non-clickable mock-ups and arrive at solutions that do exactly what they want .

The design and delivery layers of software development cannot be performed by IT professionals who lack domain expertise in human rights law. A human rights lawyer with years of experience knows that paragraph-level search is needed rather than document-level search, even if the latter is easier to implement and more performant. This domain knowledge is irreplaceable in determining what a tool needs to do and in assessing whether it works as intended.

Szoszkiewicz, identifying himself as a lawyer rather than an IT professional, explained that an IT professional who does not know human rights law will not know what a lawyer needs, such as paragraph-level rather than document-level search . He argued that the space in software development for domain experts like human rights experts is expanding while the relative role of IT is decreasing .

Academic researchers occupy a unique role as experimentalists who can explore new approaches and prototype tools without the institutional constraints faced by governments or international organisations. The results of this experimentation can then be handed over to other stakeholders, where risks can be properly considered before tools are deployed at scale. This division of roles between academic experimentation and institutional deployment is valuable for the responsible development of human rights technology.

Szoszkiewicz described his role as an experimentalist who hands over results to other stakeholders where risks should be properly considered before tools are deployed at scale . He described his journey as resembling an RPG game where you explore territory without seeing the full map, sometimes finding that approaches work and sometimes not, and perceiving his role as a character exploring this territory with the hope that some of what he develops can be used by institutions at international and national levels .

Assessing whether AI tools work as they are intended to work requires the development of benchmarks, which is a task that can only be performed by domain experts who understand the human rights context. IT professionals who do not understand human rights cannot determine whether a tool is producing the right results for the right reasons. This makes domain expertise essential not only for design but also for quality assurance.

Szoszkiewicz argued that developing benchmarks to measure whether something works as intended is the most important thing in the context of AI, and that this can only be done by domain experts and cannot be done by IT professionals who do not understand the human rights context .

While AI can relatively easily gather data on country commitments and processes, the outcome field presents a major challenge because it requires validated user feedback and documented evidence of what is actually happening for the people who are supposed to be protected. Without this end-user analysis, AI enters a huge field of hallucination when attempting to assess outcomes. This is a structural limitation that affects the reliability of AI-assisted human rights monitoring.

Leblois, drawing on G-State's 20 years of monitoring the CRPD across 143 countries using the DARE Index, explained that AI can fairly easily gather data for the first two legs of their structure (commitments and processes), but that in the outcome field, without user feedback and documented end-user analysis, AI enters a huge field of hallucination . He asked how valid outcome data can be obtained given this challenge .

UN treaty bodies operate on long review cycles of several years, which creates a significant misalignment with the rapid pace of change happening on the ground in countries. This temporal gap means that monitoring systems may be capturing a reality that is already outdated by the time it is reviewed. This structural issue compounds the challenge of obtaining valid outcome data.

Leblois noted that human rights treaty bodies at OHCHR have a very long cycle of several years for the Council to come back, whereas things are going very fast in the field . He framed this as an additional challenge alongside the difficulty of obtaining valid outcome data .

In sensitive areas such as torture, data is often deliberately by governments, making civil society organisations working at the national level essential for gathering reliable information. The Global Torture Index works with 100 NGOs across 39 countries to collect this data, with plans to scale further. The challenge is not only gathering the data but also identifying which data is relevant for assessing torture risk, visualising it, and communicating it to the broader public.

De Armas described the Global Torture Index as working with 100 NGOs at the national level to implement monitoring in 39 countries, with plans to keep scaling the following year . She noted that in terms of torture specifically, data is quite on purpose many times by governments, making civil society data collection essential . She also highlighted the challenge of identifying which data is relevant to identify the risk of torture and how to communicate it to communities and the general public .

Michaela Lissowsky opens the session by highlighting a key concern repeatedly raised in recent discussions: that AI development and governance is concentrated in the hands of a few powerful technology industry leaders. She contrasts this with the principle that human rights belong to all people worldwide, framing this tension as the motivation for the session's focus on digital tools for human rights monitoring.

Lissowsky stated that AI is dominated by a couple of tech bros, which was described as one of the repeated learning lessons from recent days . She immediately contrasted this with the principle that human rights belong to all people worldwide , using this tension to justify the importance of the day's discussion on scaling digital tools for human rights monitoring .

Lissowsky situates the current session within an ongoing series of discussions, noting that the previous year's cooperation with OHCHR and the Geneva Human Rights Hub focused on introducing digital tools and discussing their implications. The current session moves beyond that introductory stage to examine the actual real-world application of these tools in human rights monitoring practice.

Lissowsky noted that the previous year's session introduced digital tools and discussed their implications, whereas the current 45-minute session would focus on real-world application . She described this as a continuation of cooperation with the Office of the High Commissioner and the Geneva Human Rights Hub .