Part 2: Civil Society Deployment of AI: Hurdles to adoption and navigating ethical questions
Out of 293 surveys received, 63 respondents (21.5 percent of all responses) said that their organisations do not deploy AI at all. Of those that said they deploy AI tools, over half use them in very limited ways, mostly to assist with research tasks. Only 49 respondents reported that their organisations deploy AI tools more broadly, namely to assist them with a variety of tasks (upwards of six to nine tasks) such as research, content creation, data collection, data processing and analysis, and fundraising.
Of the organisations that deploy AI, three quarters use it for research. Just over half use AI tools for content creation, data collection, and administrative tasks such as note-taking during meetings, calendar management, and drafting emails. Slightly over 45 percent of respondents said AI tools provide their organisations with support in data processing and analysis and fundraising tasks such as writing proposals.
While deployment of AI is still limited, its development is even more limited. Only 37 respondents said their organisations have experience developing their own AI solutions such as chatbots, AI models for data analysis or building training datasets.
Examples of development based on survey responses, desk research, and the focus group include:
- (R0014) An internal AI-powered digital engagement and research tool built by the Roberto Save Dreams Foundation in Angola to support policy research and data analysis on climate change, gender equality, and rural development, as well as communication for awareness campaigns and stakeholder mapping.
- In Türkiye, the Centre for Democracy Research (CDR) developed Policyful,110 an AI-powered platform that allows its users to analyse and interpret policy content interactively and access a regulatory archive covering the Turkish Parliament, ministries, regulatory agencies and municipal councils.
- In Jordan, the Jordan Open Source Association (JOSA) built Nuha, 111an open-source model for detecting and classifying online gender-based hate speech in different Arabic dialects.
- Independent Tunisian media outlet Nawaat launched Nawaat Chronicles,112 an AI-enhanced archive platform allowing readers, researchers, students and journalists to easily access and search an archive of over 22 years of independent journalism on Tunisia in Arabic, French, and English.
- Independent media outlet Raseef22 launched Ask Aunty, an AI-powered chatbot that enables readers to access accurate information on sexual and reproductive health and rights (SRHR) in Arabic.113
- In Nigeria, organisation (R0021) mentioned that they developed Value for Money AI, a tool that “automatically analyses audit reports, detects corruption indicators, and drafts legally sound petitions for submission to anti-corruption agencies”. The organisation fsaid that: “Our motivation for developing the solution is to address the inefficiency and complexity of drafting and submitting corruption-related petitions in Nigeria. Many citizens, journalists and activists struggle to transform audit data into legally sound petitions. By automating this process using generative AI, we enhance accessibility, accuracy and efficiency, empowering stakeholders to fight public sector corruption effectively”.
- Ekwere Akpan Foundation, R0036, stated that it developed EAF Insight, “an AI-powered data dashboard designed to analyse community needs and monitor project outcomes in areas such as education, poverty reduction, and women’s empowerment” to “enhance data-driven decision-making, improve transparency in project reporting, and strengthen collaboration with partners such as UN agencies and local governments”.
- South African organisation GRIT built Zuzi, an “African, and trauma-informed” AI chatbot to provide information on the rights and legal services available to survivors in addition to SRHR issues.114
Obstacles to AI deployment
A combination of factors prevents organisations from deploying AI effectively or deploying it at all. Of the 63 surveyed respondents that do not use AI at all in their work, a lack of financial resources and technical capacity were the most commonly cited factors. Slightly over a third of respondents cited challenges such as the lack of models in languages and contexts related to their work as well as human rights and ethics concerns.
Regardless of whether organisations deploy AI or not, the lack financial resources and technical capacity are major obstacles to deployment.
One respondent (R0273) working with a South Sudanese organisation that deploys AI in a variety of tasks—data processing and analysis, content creation and promotion, engagement with target audiences, support with administrative tasks and fundraising— wrote that AI “is literally not familiar to us”, particularly when it comes to development of chatbots and apps for engagement with target audiences. The respondent highlighted the “need [for] complete package training” to be able to deploy AI tools “without difficulties”.
Another respondent (R0158) working with an organisation that deploys AI in very limited ways in research wrote: “The major obstacles preventing our organisation from developing and deploying Artificial Intelligence (AI) solutions are primarily financial constraints and limited technical capacity. As a nonprofit organisation, most of our resources are directed toward programme implementation rather than advanced technological investments. We lack the funding required to procure AI infrastructure, software and specialised expertise. Additionally, our staff have limited exposure to AI tools and applications, making it difficult to integrate AI into our operations without targeted training and technical support. Overcoming these barriers would require both capacity-building and financial investment to harness AI for social impact effectively”.
The constraints mentioned above can be interlinked. CSOs often lack financial resources to hire professionals who can support them with exploring the possibilities offered by AI tools, as well as integrate such tools into their workflows with policies and measures to ensure AI adoption does not result in bias, inaccuracies, privacy violations and other ethical and human rights infringements. This reality has been compounded in recent years by large-scale foreign aid budget cuts that have forced many organisations around the world to scale back or shut down operations.115 This has further limited the resources available for employees, making it a challenge to “balance daily programme demands with learning and experimenting with AI…due to time constraints and competing priorities” as one respondent noted. Another respondent, R0200, whose Guatemala-based organisation does not deploy AI at all, stated: “Human rights organisations like GAM are facing a big crisis related to funding, and this generates a big gap in getting more information about how to use AI tools to strengthen staff knowledge”.
In some cases, finding people with necessary technical skills is a challenge. R0202, who works in Ethiopia, noted that “the most significant challenge is limited financial resources, which makes it difficult to invest in advanced AI infrastructure and high-quality computing tools. In addition, there is a shortage of local expertise in AI programming and data science, making it necessary to rely on external consultants for technical support”.
“As a non-profit organisation primarily focused on civic and climate work, we face constraints in terms of in-house AI expertise. Developing and managing AI models requires specialised knowledge in data science, machine learning, and software engineering. These skills are not always readily available within the civic sector,”, a respondent from Kenya said.
Financial resources are also needed to acquire the necessary hardware for organisations to develop their own tools. R055 of Spring NGO in Pakistan said: “Developing or adapting AI solutions requires investment in software, hardware, and skilled professionals, which has not yet been feasible within the organisation’s existing funding structure. Most available grants focus on direct programme delivery rather than technological innovation”. Some respondents mentioned bad overall infrastructure, such as poor internet connectivity and lack of cloud computing and data storage and management infrastructure, which prevent them from deploying AI in their work. R0264, working in South Africa, noted that “reliable internet connectivity, high-performance computing resources and stable power supply are not always available—especially in remote or developing regions where our work is primarily focused”.
Respondents also noted the lack of models that reflect the contexts in which they work as another obstacle. “Many AI models are not trained to reflect local languages, cultural contexts, or the lived realities of marginalised groups such as LGBTQI+ persons in Nigeria, making them less useful or even biased in our setting. We also have ethical concerns around data privacy, misinformation and the possibility of algorithmic bias, which could harm the very communities we seek to protect”, one respondent wrote.
R0114, whose organisation “is in the early stages of developing AI-driven advocacy and research systems” said that “we are cautious regarding ethical implications and language representation. Most AI tools lack training in Arabic and MENA cultural contexts, limiting fair accessibility and representativeness”.
“Our main limitation is access to locally contextualised models trained on African nutrition and behavioural datasets. Current tools often reflect Global North biases and lack sensitivity to local nutrition, and language and gender dynamics”, Nigeria-based respondent R051 said.
R0219 of Centre Stage Media Arts Foundation (CSMA) in Zimbabwe explained that in addition to being constrained by limited human, technical, and financial resources, language is an obstacle: “We are unable to deploy free AI tools since they use English mostly, yet our audiences are more conversant in local languages”.
In addition, the lack of high-quality data availability on which models can be trained is a major hurdle for civil society actors to develop their own models instead of relying on dominant models. This lack of data becomes particularly scarce for low-resource languages and for excluded groups. “For AI to function effectively, it requires large, high-quality datasets. In our case, linguistic and cultural data (especially for under-documented or oral languages) are limited or scattered, which makes model training difficult”, said R0264.
This is a sentiment echoed by disability rights organisation Spring NGO in Pakistan (R055): “access to reliable data remains a barrier. Effective AI deployment depends on large, high-quality and well-structured datasets — something that is often unavailable in the disability and inclusion sectors in Pakistan. The lack of standardised, disaggregated data makes it difficult to build or train AI systems that could meaningfully support advocacy work”.
“In many of the communities we serve, especially in rural or marginalised areas, there is limited access to reliable data. Additionally, existing data may be fragmented, outdated, or even not collected in formats suitable for AI processing”, said R0176, who works in the DRC.
Finally, sometimes AI use can be perceived negatively, which may deter deployment. One respondent from Sri Lanka (R0199) wrote: “Normally when we are writing proposals we use AI like ChatGPT to develop our proposals, once we give our own input, it refines our English. But, the issue is, once it is submitted the funders find it as AI generated as the wordings are from AI though the idea is from our side. So, though we like to use AI, the tendency of rejecting our proposal is high due to recognition of AI use”. Another respondent from Nepal (R0151) added that due to the multiple risks they fear public backlash: “Currently there are too many opaqueness about AI models and their algorithms that prevents us from providing the models [with] our data which we do not know how [it] will be used and stored. Also we do not want to misuse other’s copyright materials in our works without proper attribution, and we definitely do not want to lose public trust in our institution due to AI”.
Human rights and ethical concerns associated with AI deployment by civil society
Civil society globally has expressed concerns over human rights and ethical questions related to AI use. Nearly a third of respondents cited human rights and ethical concerns as one of the obstacles that has prevented their organisations from effectively deploying AI or deploying it at all. Respondents were particularly concerned about exacerbating the spread of misinformation and disinformation (60 percent of all respondents) and the risks to privacy associated with the design and use of AI models and systems (56 percent).
When it comes to misinformation and disinformation, concerns include the use of Generative AI models, which can generate outputs that may contain inaccuracies, biased information or hallucinations. Inaccuracies and biases in Generative AI outputs can thus contribute to undermining information integrity and public debate.
“Generative AI models can produce convincing but false or misleading content, which can easily be weaponised to spread propaganda, manipulate public opinion, or discredit CSOs. This is especially concerning in contexts where information ecosystems are already fragile. The speed and scale at which AI-generated misinformation can circulate threaten both democratic discourse and public trust in legitimate information sources”, respondent R0201 noted. “AI-generated content may contain inaccuracies or outdated information, which could be misleading or harmful if disseminated to the public or used in advocacy efforts”, another respondent (R048) said.
An online environment of information disorder makes it challenging for civil society actors to spread their messages. As one anonymous respondent noted, “the growing problem of misinformation generated by AI could weaken public trust in advocacy messages and make it harder for people to distinguish fact from manipulation, a risk that can undermine human rights work and advocacy credibility”.
Respondents expressed concern about the way existing and dominant models were developed—through massive data collection including personal data without informed consent—and thereby potentially exposing personal information, in inputs into AI models. According to R0201, “many AI systems rely on large amounts of data, including personal or sensitive information. This raises serious concerns about how such data is collected, stored and used, especially in environments where digital surveillance is common and data protection frameworks are weak. There is a real danger that AI systems could unintentionally expose or misuse information belonging to activists, journalists, or human rights defenders, putting their safety and privacy at risk. Ensuring data security, consent and transparency in AI operations is therefore a top ethical priority for us”.
“Our main worry is privacy — most AI tools collect and process large amounts of data, and in a country like Nigeria where LGBTQI+ people face discrimination, any misuse or exposure of personal data could put lives at risk”
— Anonymous respondent, Nigeria
Nearly half of the respondents were concerned about exploitation of other people’s work in dominant AI models (for example, the training of models using copyright protected work without consent of authors or compensation) and risks of bias in the design and outputs of AI models and systems. Some organisations also cited the risk of dependence on AI models designed and controlled by private tech players in the Global North (44 percent).
The risk of bias is attributed to dependence on Global North models that do not reflect realities and contexts in which respondents and their organisations work. It is also attributed to existing systems being trained on biased data and reflecting existing inequalities. These risks can particularly affect excluded groups such as LGBTQI+ people and persons living with disabilities.
“The organisation’s foremost concern is that AI technologies, if not developed and implemented responsibly, may unintentionally reinforce existing inequalities and biases—particularly those affecting persons with disabilities. Many AI systems are trained on datasets that under-represent or misrepresent marginalized groups”, said R0055 of Spring NGO in Pakistan.
“Many AI models are trained using data that does not reflect the realities or voices of African or queer communities. This can lead to distorted or harmful representations that reinforce stigma instead of supporting equality. Another ethical issue is the dependence on AI tools developed and controlled by big tech companies in the global north. These systems often lack transparency (“black box” algorithms) and are not accountable to local civil society needs”, another respondent said.
“We are concerned that most AI tools are trained on data from contexts that do not reflect African realities, which can perpetuate bias in advocacy analysis and misrepresent local voices. Additionally, privacy and data protection remain critical challenges in South Sudan where regulatory oversight is weak”, stated R0035.
For some, the dependence on dominant models is seen as an ‘inequality issue’ that only advances the interests of very few powerful companies. For R0208, based in Honduras, “one of the main ethical concerns about the use of Artificial Intelligence relates to the displacement of human interests and rationalities by an algorithm that ultimately responds to the interests and priorities of a small group that often does not even know the priorities of the majority”.
“Exploitation occurs when creators’ copyrighted work is used to train AI models without consent or fair compensation. The high cost of developing these systems is an inequality issue, concentrating power and access within a few wealthy organizations, which in turn reinforces biases and compromises safety”, noted R0234.
“The high cost of developing AI systems is an inequality issue, concentrating power and access within a few wealthy organizations”
— R0208, Honduras
The climate and environmental impacts were also a concern for 115 survey respondents, nearly 40 percent of all respondents.
R0266 said: “Artificial Intelligence (AI) systems have significant environmental impacts throughout their lifecycle. Training large models requires massive energy, often thousands of megawatt-hours, leading to high carbon emissions—sometimes exceeding the lifetime emissions of several cars. Continuous operation for [AI] inference116 also consumes large amounts of electricity. Data centres use vast quantities of water for cooling, contributing to resource stress, especially in arid regions. The production and disposal of AI hardware creates electronic waste and demand for rare minerals like lithium and cobalt, which have their own environmental and social costs. Additionally, data storage and transmission add to energy consumption. These impacts contribute to climate change, resource depletion, and environmental inequality.”
Another respondent working with the World Diplomacy Organisation (WDO), R0108, wrote in response to the survey that “as an institution that promotes sustainable development, responsible governance, and ethical innovation, WDO believes that the rapid expansion of AI technologies must be balanced with environmental responsibility. The organization advocates for greener, energy-efficient AI models and greater transparency from technology providers regarding the carbon footprint and resource consumption of AI infrastructure”.
There are also concerns about the lack of human oversight and accountability in decisions around AI systems and their outputs. Increased reliance on AI to make decisions and find information could erode human judgement,117 agency, empathy and creativity and other human qualities, which are necessary to the work of civil society. One survey respondent who chose to remain anonymous wrote: “Civil society is often the last line of defense for marginalised groups. Its strength lies in human empathy, contextual understanding, and the ability to challenge unjust systems. Our deep concern is that an over-reliance on AI could lead to the ‘automation of compliance’, where human workers simply follow the AI’s recommendation without question. This erodes the essential human judgment and moral reasoning that is the bedrock of effective civil society work. When a flawed decision is made, who is accountable? AI? The developer? The CSO that deployed it? This accountability vacuum is a serious ethical problem. An AI does not ‘know’ it is replicating past redlining or gender discrimination; it just finds patterns in the data it was given. The ethical failure occurs when this output is accepted as fact without critical scrutiny”.
“There are cases where people blindly trust the answers provided by AI without analysis or comparison with reality”, said a respondent from Bolivia, R0192. R056 stated that “our organisation has not yet deployed AI primarily because of concerns about its limited ability to authentically connect with human emotions and relationships”, adding that: “We place high value on genuine human interaction, empathy and contextual understanding—qualities that current AI systems often struggle to replicate with originality and depth”.
Whether organisations deploy AI or not, respondents who took part in the survey generally showed an awareness of the risks. They highlighted the need to deploy AI responsibly and with sufficient safeguards to prevent any harm.
One respondent from Kenya noted the following: “We believe that civil society has a responsibility not only to adopt new technologies but also to advocate for safeguards, policies, and practices that ensure AI serves the public good, especially in areas like climate justice, civic participation, and social equity. We are particularly interested in collaborative approaches that centre community voices in the design and implementation of AI tools, ensuring that technology amplifies, not replaces, human agency and democratic processes”.
In some cases, organisations decide not to deploy AI tools due to ethical and human rights concerns.
R024, who works with an Ethiopia-based organisation that does not deploy AI, said that while the organisation STEM Plus “recognises the transformative potential of Artificial Intelligence, deployment is hindered by unresolved ethical, legal, and human rights concerns”. The respondent further added that: “we actively consider potential ethical and human rights implications in all our digital initiatives. Our approach is precautionary and principled, ensuring that when AI becomes part of our programs, it is introduced responsibly, transparently and inclusively. We believe that sustainable technology must not only advance innovation but also protect and uplift humanity”.
These human rights and ethical concerns do not necessarily equate to disinterest from CSOs in deploying AI but rather show that integration should be done responsibly. Colombia based R0154 stated the following: “We are interested in exploring artificial intelligence, but we are also concerned about how it might reinforce existing inequalities. Most models reflect perspectives and languages foreign to our contexts, which limits their value for working with local communities. There are also concerns about data protection, the transparency of algorithms, and the environmental cost of these technologies.
“The question is not just whether to use AI, but how to do so without losing autonomy or ethical purpose.”
The next section explores how organisations address ethical issues and human rights risks in their deployment of AI.
How CSOs attempt to navigate human rights risks and ethical questions in their deployment of AI
Despite the concerns noted above and general awareness among respondents about the risks of AI deployment, not enough actions are taken by CSOs to implement policies and measures to ensure that AI use does not infringe on human rights. Of the 230 respondents who said that their organisations deploy AI, only 38 percent clearly stated that they consider and attempt to address ethical concerns and human rights risks.
There are two factors that could explain this trend. First, for those organisations that deploy AI tools, its usage remains very limited, while only a small fraction of organisations have developed their own models. Another factor is the lack of technical knowledge and resources CSOs need to address the identified risks and adhere to ethical guidelines when deploying AI.
R0293 stated that: “AI is a relatively new technology, about which we have little information regarding its nature and risks; however, the issue of privacy security is one of the biggest fears.” R030, whose organisation does not deploy AI, explained, “What is preventing our organisation from using AI is mainly the lack of knowledge on the subject and the necessary funding. We want responsible AI that does not violate human rights, an AI that is useful for our work and our community. The biggest challenge for us is the various risks and impacts associated with using AI. We won’t know how to prevent them because we have very limited knowledge”.
Overarching policies and principles guide deployment of AI for some of the organisations that took part in the survey. In particular, these emphasise respect for data privacy, protection from bias and non-discrimination, fairness, inclusivity and inclusion, transparency and accountability, informed consent, human dignity, protection of vulnerable populations and marginalised communities, gender-sensitivity and respecting copyrights.
R0233, based in Bangladesh, said: “We approach all forms of AI use through a strict ethical lens. Our primary focus is to ensure data privacy, avoid algorithmic bias and maintain respect for human dignity. Even in early stages of AI integration, we consult relevant international human rights frameworks and prioritize inclusive, safe and accountable practices. If we adopt or develop AI solutions in the future, a dedicated ethics oversight mechanism will be established”.
Several organisations mentioned that while they still do not deploy AI, if they would in the future, they plan to ensure adherence to ethical principles and human rights standards. “As the organisation explores integrating AI into program design and monitoring, it remains committed to aligning with ethical and human rights standards. This includes ensuring data privacy, obtaining informed consent, promoting gender-sensitive algorithms and avoiding bias or harm to vulnerable populations. WGHEN intends to collaborate with ethical AI experts and adopt global best practices to ensure that any future AI solutions uphold transparency, accountability, and inclusivity in line with human rights principles”, stated R0158.
Of respondents whose organisations deploy AI, only 28 percent clearly stated that such deployment is guided by specific human rights frameworks or ethical guidelines. These include the African Union Artificial Intelligence Continental Strategy,118 OECD AI Principles,119 organisations’ internal policies, the UNESCO Recommendation on the Ethics of Artificial Intelligence,120 and to a lesser extent the UN Guiding Principles on Business and Human Rights.121 Some organisations also listed local laws and policies such as national AI strategies and data protection laws. To help implement overarching policies and principles, organisations mentioned putting in place practical measures. The following sub-section provides an overview of these measures based on survey responses and insights from the focus group discussion.
Practical measures for responsible AI deployment
Facial Recognition Technology (FRT)
Protection of privacy and personal data is of utmost importance for organisations deploying AI. To protect personal data, some CSOs said they refrain from inputting identifiable and sensitive information into AI models, whether public systems or internally developed ones, or they anonymise such data.
A survey respondent who requested to be anonymous wrote: “We try to be mindful of ethical and human rights risks whenever we use AI tools, even though we have not developed our own AI model yet. Our organisation works closely with vulnerable LGBTQI+ individuals, so we prioritize data privacy, confidentiality and consent in everything we do. For instance, we avoid uploading sensitive or identifiable information about beneficiaries into AI platforms and only use trusted tools for research, learning and communication”.
For its Ask Aunty chatbot, Raseef22 minimised personal data collection from the start. “We do not require any accounts or people to log in [to use the chatbot], and we do not ask them to give us any information. The only information that we store is the country – just for analytical purposes. We do not collect any personal identifiers, and we do not share the information with any third parties. So, our team is the only one who can see the queries and the questions on the chatbot”, a Raseef22 employee explained during the FGD.
Raseef22’s Ask Aunty: privacyfirst design for sensitive health information
For its Ask Aunty chatbot, Raseef22 minimised personal data collection from the start. “We do not require any accounts or people to log in, and we do not ask them to give us any information. The only information that we store is the country — just for analytical purposes. We do not collect any personal identifiers, and we do not share the information with any third parties.”
One FGD participant whose organisation deploys AI internally to gather and organise data mentioned that they limit the tool’s access to the organisation’s systems and private information. “The thought we used there is, would we give a junior staff person the same level of permissions? If we would, we can think about it with AI. If we would not, [then] we should not give AI that many permissions in our system to view information and work with information”.
To build Nuha, an AI model that can classify hate speech against women in different Arabic dialects, JOSA collected data from social media. “We went through a lot of processes on how we should handle this data, what kind of data we should include when we are monitoring social media and how we can ensure privacy through social media accounts we collected data from, and it was very important that no personal information was to be involved”, said a JOSA employee during the FGD. She further explained: “we do not store any personal information, and in the data collection process, the data collectors remove the data manually of any personal information”.
Informed and prior consent of data holders, ensuring they understand how their information is used and that they can refuse processing are essential. One organisation stated: “We will avoid deploying AI where consent, safety and data sovereignty cannot be guaranteed”. Another respondent based in South Africa, R0264, stated that “when working with communities — especially in language documentation, translation, or storytelling — we make sure participants understand how their voices, stories, or recordings will be used. We never collect or share data without permission”.
Measures to protect personally identifiable information include secure data management and storage practices. “We prioritise the principles of data minimisation, informed consent, and confidentiality. Personal or sensitive information is handled with strict access controls and encryption. Where possible, we use anonymised or aggregated data to protect individuals and communities we serve”, said Tanzania based R0201.
Ethical reviews and data audits
In their deployment of AI tools, some respondents mentioned that their organisations assess risks to identify potential harms associated with bias, privacy, and misinformation. Assessments were emphasised not only by those that developed their own models but also organisations that use existing models. R0031 in Bangladesh wrote: “We consider and address potential ethical and human rights risks by implementing rigorous safety evaluations, fairness assessments and ongoing monitoring to ensure responsible AI deployment”.
A number of organisations mentioned that they conducted these assessments prior to AI adoption or they make these assessments on a regular basis. “Before implementing any AI tool or system, we conduct an internal review to identify possible risks to privacy, security and fairness. This includes assessing how data is collected, stored and processed to ensure compliance with human rights and data protection standards”, stated R0201, based in Tanzania.
Auditing datasets for biaseswas specifically highlighted for those that created or are interested in creating their own tools during the design and development stage.
“We carefully review datasets and AI outputs to identify and minimize biases that could disproportionately affect marginalized groups, especially women and rural communities”, said a survey respondent whose organisation developed their own model (R0014). A respondent who requested anonymity stated the following: “If and when we do deploy AI tools, our ethical and rights-based approach would include… reviewing datasets for potential bias and ensuring that AI models do not reinforce harmful stereotypes as well as exclude underserved populations. Where necessary, we will work with technical partners to audit data and algorithms for fairness”.
During the data annotation process to develop Nuha, women annotators working in the human rights field were given guidelines to ensure that the model would be free from biases and would not infringe on freedom of expression. “The thing that we had challenges with is that there is a very thin line between hate speech and freedom of speech; how do the annotators know which is which, and how to differentiate them both because we do not want model that classifies everything as hate speech”, explained a JOSA representative in the FGD discussion.
In some cases, respondents acknowledged that the capacity of their organisations is limited when it comes to conducting these types of assessments. “We also discuss ethical implications within our team when exploring AI-assisted work such as possible bias, misinformation and data security risks. However, we recognize that our current capacity to conduct deeper ethical assessments is limited. We are interested in building stronger digital literacy and AI ethics skills to ensure that as we integrate technology into our advocacy, we do so in a way that protects human rights and maintains community trust”, said a survey respondent.
Transparency and accountability
Respondents noted the following as important factors, namelyttransparency about AI tools through the documentation of their development and purpose; informing users of potential risks; and making it clear to users what the AI tool does.“If and when we do deploy AI tools, our ethical and rights-based approach would include …making AI tools understandable to users and stakeholders, with clear explanations of how decisions are made”, stated one organisation.
AI outputs and decisions also need human oversight. Some of the mechanisms put in place to ensure this oversight include Human-in-the-loop (HITL) validation, feedback mechanisms, and human review and verification of any AI-generated content “for accuracy, fairness and sensitivity before publication”, as respondent R051 stated.
CDR (R020) in Turkey, which developed Policyful, said it ensures transparency of its AI-assisted outputs and human oversight. The organisation explained in a survey response:
“Our organisation develops AI tools to strengthen civic engagement and transparency in policymaking. We prioritise the ethical use of AI by ensuring data integrity, minimising hallucinations through human-in-the-loop validation and maintaining accountability in automated analyses.…We integrate ethical review and human rights considerations into every stage of our AI projects. This includes careful data selection to avoid bias, transparency about AI-assisted outputs and continuous human oversight in model training and deployment. We also provide public explainers and open datasets to ensure that our tools contribute to democratic accountability rather than replace human judgment”.
“Our organisation develops AI tools to strengthen civic engagement and transparency in policymaking, prioritising ethical use through humanin-the-loop validation”
— Centre for Democracy Research (CDR), Türkiye
One Tanzania-based organisation, R0201, stated that: “We are committed to transparency in how AI systems operate and how decisions are made. We provide clear communication to stakeholders about the purpose and limitations of AI tools, and we maintain accountability through regular audits and feedback mechanisms”. Another organisation based in South Africa, R0264, wrote: “We openly communicate when AI tools are involved in our processes, ensuring that human reviewers, translators and leaders maintain oversight and final responsibility for all content produced”.
When AI is deployed in research, one respondent mentioned that they ensure outputs are verified with scientific sources and adequately cited. Another based in Kenya, R0168, said that “we do not rely solely on AI outputs for public advocacy without verification”.
Emphasis was put on using AI to assist humans and not replace them, particularly in important decisions and when working with excluded groups. According to R0201: “AI tools are used to support, not replace, human judgment. Critical decisions, especially those affecting rights or reputations, remain under human review to ensure ethical alignment with our values and mission”.
R055 said: “If civil society groups adopt AI tools without understanding how algorithms function or how decisions are generated, it can lead to a loss of human oversight. Spring NGO highlighted that AI should support, not replace, human judgment — especially in matters affecting vulnerable populations”.
“AI should support, not replace, human judgment — especially when working with marginalised groups”
— Spring NGO, R055, Pakistan
Stakeholder engagement
Some respondents underlined the need to consult with and involve key stakeholders, and particularly experts on AI ethics and risks, as well as affected communities and local authorities in relation to civil society’s deployment of AI.
R0219: “If CSMA is to deploy AI models, ethical and human rights considerations and risks will be of utmost importance. Adhering to ethical issues will ensure that we respect local cultures, involve local communities in our deployment of AI models and also ensure that our people, the primary beneficiaries of our programmes, harvest the fruits of our AI inspired work.”
JOSA conducted consultations with feminist organizations and human rights defenders in Jordan to build Nuha. “Feminist organizations in Jordan had a really big role in the model to be inclusive because it was important for us to build something that is localized that is similar to what we think violence looks like in our own countries, and so it was very important for these feminist organizations to decide on the terminologies on the types of violence”, said an FGD participant from JOSA.
Prioritising accessibility and contextual adaptation
When AI tools are deployed, it is essential that they are accessible to people and communities they are supposed to serve and that such deployment do not further result in inequality and marginalisation, for instance, due to disability, language obstacles, or lack of internet access.
“If and when we do deploy AI tools, our ethical and rights-based approach would include…ensuring that the benefits of AI are distributed equitably and do not widen the digital divide. This includes building capacity and digital literacy among our beneficiaries and partners”, a respondent who requested anonymity said.
R0055: “accessibility and inclusion remain at the core of Spring NGO’s ethical framework. AI tools must be designed to be accessible to persons with disabilities in terms of both interface and content. Without inclusive design principles, technological innovation risks deepening the digital divide rather than closing it. In summary, Spring NGO advocates for an ethical, inclusive, and rights-based approach to AI, where technology amplifies human dignity, promotes equality, and ensures that no one — especially persons with disabilities — is left behind”.
Further, some respondents said in their deployment they seek to adapt AI to their local contexts to prevent risks of bias and exclusion.
CDR mentioned that it has a “multidisciplinary team of AI developers, researchers and legal engineers [that] ensures that model outputs remain contextually accurate and rights-respecting”.
“We recognize that most AI systems are trained on data that reflect Western or dominant cultural worldviews. Therefore, we work carefully to contextualize AI-generated outputs, especially when translating Scripture or creating educational content, ensuring cultural and linguistic accuracy”, said a South African organisation, R0264.
Awareness raising among staff
To help implement their AI policies and measures, organisations also take steps at raising awareness among their staff, partners and beneficiaries on how to use AI ethically and responsibly and about the human rights risks and implications of such technologies.
“Wardil Organization in Iraq, R0018: stated that it “considers the ethical and human rights risks related to the use of AI. Although we have not developed our own AI model, we use AI tools in limited ways—mainly for communication, translation, and data organisation. We address potential risks by promoting transparency, privacy protection and fairness, and by raising awareness among our staff about responsible and ethical AI use”. While R0201 wrote: “We regularly train our staff and partners on responsible AI use, data ethics and human rights implications, ensuring ongoing awareness and adaptation to emerging standards and best practices”.
Limiting scope of AI use or size of training datasets
Further, to minimize human rights risks, some organisations deliberately choose to deploy AI in limited ways, for instance, in administrative tasks and data processing
“We take potential ethical risks into account, which is why we only implement artificial intelligence for administrative processes and data processing to reduce manual work”, said a respondent from Ukraine. Another, R0166, working in Ethiopia stated: “We take precautions not to impact the rights of others. This may be through not completely relying on AI as it may create misinformation. Particularly we apply caution while using AI for research and content creation”.
An FGD participant from Uganda said his organisation is “not actually jumping onto the [AI] train
is not driving core programs at his organisation due to the sensitivity of the topic it is working on—gender-based violence— and the lack of context sensitive tools. “The biggest concern is that AI does not really reflect our realities. Tools do not understand our languages, our culture, our context, which can really lead to harm, especially for organizations like us that are dealing with issues that are related to gender-based violence”, he explained.
Others choose to develop models based solely on small datasets or data that they control to minimize risks associated with large language models, which are built on massive amounts of data scraped from the internet. This is the case for several AI tools developed by CSOs and highlighted above in this report.
113 Kamel, R. “Ask Aunty bridges “taboo’’ conversations in the Middle East”. JournalismAI. 2 June 2025. https://www.journalismai.info/blog/ask-aunty-bridgesnbsp-middle-east-health-info-gap (accessed 1 May 2026).
114 GRIT. “Our Tech”. Undated. https://www.grit-gbv.org/technology (accessed 1 May 2026).
115 Human Rights Funders Network. “Funding at a Crossroads: Foreign Aid Cuts And Implications For Global Human Rights.” September 2025. Https://Www.Hrfn.Org/Wp-Content/Uploads/2025/09/Funding-at-a-Crossroads-HRFN-Sept-2025.pdf
116 The process “of using a trained AI model to make predictions on new data”. See more at: https://www.ibm.com/think/topics/ai-inference
117 “A nuanced process that draws on context, trust, and human insight” to make decisions or form opinions.
118 African Union. Continental African Intelligence Strategy. July 2024. https://au.int/sites/default/files/d
119 OECD AI Principles. https://www.oecd.org/en/topics/ai-principles.html (accessed 3 May 2026).
120 UNESCO. Recommendation on the Ethics of Artificial Intelligence. 16 May 2023. https://www.unesco.org/en/articles/recommendation-ethics-artificial-intelligence (accessed 3 May 2026).
121 OHCHR. UN Guiding Principles on Business and Human Rights. 2011. https://www.ohchr.org/sites/default/files/documents/publications/guidingprinciplesbusinesshr_en.pdf (accessed 3 May 2026).