First published in Digital Transformation and Governance Gaps: The Case of Armenia (DGAP, September 2026). Full report available at DGAP.org.
Introduction
Across Armenian civil society, artificial intelligence (AI) tools have gradually become part of the daily work of drafting grant proposals, conducting and summarizing research, and translating documents, without any rules governing their use. This is happening during a period of rapid, state-backed expansion of the country’s AI sector. So far in 2026, the government has added AI to the name of the Ministry of High-Tech Industry, thus formally elevating the sector, while major private investors have launched large-scale compute facilities. The Ministry of Education is partnering with OpenAI and Firebird to provide 50,000 users with free access to ChatGPT Edu and Codex.
In terms of the legal framework for regulating AI adoption, Armenia’s key external reference points are the Council of Europe Framework Convention on Artificial Intelligence, which it signed in January 2026, and the EU’s AI Act. It agreed with the EU in May 2026 to explore cooperation on AI in line with the act, which is framed as part of broader integration into the EU’s digital single market. This nonbinding commitment, however, is contained within a joint statement focused on infrastructure, investment, and market access rather than on how AI should be used or governed. Armenia has no AI legislation, and the reference to the EU act has not yet been translated into operational rules for organizations using AI tools. That gap is most visible in civil society.
This policy brief argues that the widespread adoption of AI without the institutional capacity to mitigate risks creates serious vulnerabilities. At the same time, through everyday practice, Armenian civil society organizations (CSOs) are developing informal norms around the use of AI tools and becoming de facto co-creators of AI governance, and this should inform the country’s emerging AI regulatory framework.
This policy brief draws on 12 semi-structured interviews conducted with representatives of Armenian CSOs working across the fields of public policy, media, human rights, culture, governance, youth development, and social services. It is also informed by the author’s professional experience in leading digital systems and AI adoption efforts at the APRI Armenia think tank, including the development of internal guidance on AI use and staff training. More broadly, the analysis builds on over 14 years of the author’s work in supporting CSOs in Armenia through digital systems development, digital literacy training, and technology capacity-building initiatives.
A Three-Level Governance Gap
The practices of AI adoption within Armenian civil society can be understood through three connected levels: national regulation, organizational governance, and individual practice.
At the national level, the government has made international commitments and is developing a national AI strategy, but it has not yet adopted domestic regulations. In January 2026, Armenia signed the Council of Europe Framework Convention on Artificial Intelligence, the first internationally binding instrument in the field. The convention – which no signatory except the EU has ratified yet – sets out high-level principles like human dignity, nondiscrimination, privacy, and data protection. As for the EU’s AI Act, for the government it is mainly about market access rather than governance. The Ministry of High-Tech Industry is developing a national AI strategy, with responsible development, data protection, and accountability among its declared objectives. As of July 2026, there was no adopted framework, implementation guidance for civil society, or public timeline.
At the organizational level, most CSOs operate without formal AI policies or internal guidelines. At the time of writing, of the ones included in the research, only two had developed a written AI policy. A small number had verbal agreements within their team about acceptable use, typically focused on content accuracy or avoiding the input of personal data. The majority relied entirely on individual judgment, with no shared understanding of what constitutes acceptable AI use, what counts as sensitive data in their context, and no requirements for internal or external disclosure.
At the individual level, AI adoption is driven primarily by productivity needs. CSO staff members use AI to increase productivity, experimenting with tools (often through personal accounts) based on curiosity or peer recommendations. Levels of AI literacy vary considerably. Many users have a limited understanding of how generative AI systems process and retain data, what privacy and confidentiality risks arise from sharing organizational information with third-party platforms, how to detect inaccuracies or biases in AI-generated content, or when disclosure of AI use may be expected by partners and the public.
Each of the issues observed pushes responsibility downward to the next level: with no national guidelines, organizations improvise; with no organizational rules, individual staff members are left to decide on their own judgment what is safe to share and when to disclose.
How CSOs Use AI Tools
Across the 12 CSOs looked at, generative AI platforms – particularly ChatGPT, Gemini, and Claude – have been used most consistently for three purposes: drafting and editing text (grant proposals, reports, communications materials); translation and proofreading; and research-support tasks such as summarizing documents, transcribing meetings, and generating ideas.
For CSOs operating in multilingual environments with small teams, AI-powered translation and editing tools provide immediate practical value. One human rights CSO described using AI extensively for drafting reports, grant proposals, and research and translation support, particularly following staff layoffs after losing a significant funding source. A cultural CSO reported using ChatGPT and Grok for tasks ranging from drafting emails and grant applications to creating design mockups of folk costumes. A youth CSO uses AI primarily for participant outreach and communications, while limiting its use to non-sensitive documents.
The level of AI integration varies sharply across the CSOs. At one end, the management of an educational one encourages all staff to use AI and has invested in paid subscriptions. At the other, a youth CSO reported minimal adoption with no organizational position on AI at all. In most cases, adoption depends on individual initiative – a single enthusiastic staff member drives a CSO’S entire AI use, sometimes over management skepticism. This means that AI governance, where it exists, is often tied to a person rather than embedded in the organization.
In most CSOs, staff members make daily decisions about data-sharing, AI tool usage, and disclosure without a shared framework to guide them. Several key patterns emerged from the interviews.
Informal, verbal rules and lack of written safeguards: The most common form of “policy” consists of verbal understandings communicated informally by the CSO’s leadership. For instance, a cultural one has verbally advised caution regarding historical and religious content, while a think tank has verbally banned AI content generation but encouraged it for translation and proofreading. None of these verbal rules have been formally documented. Conversely, a youth CSO has established a strict verbal rule against inputting participant data. However, even the most cautious CSO lacked written protocols or systematic safeguards, leaving management largely unaware of exactly how extensively AI is being used across the organization.
Data privacy blind spots: Several of the CSOs admitted that staff input internal documents, donor communications, and project details into generative AI tools without understanding platform privacy settings. One interviewee noted a total lack of concern about inputting financial data, reasoning that this was not sensitive because their CSO’s work was legal (“We have nothing to hide”).
Inconsistent disclosure practices: Most of the CSOs do not disclose AI use in their outputs. The reasons given for this vary. Some view AI editing as too minor to mention, while others simply have not considered it. A few disclose AI-generated images but not AI text. Exceptions include one think tank with an online disclosure policy, and media organizations bound by the Media Ethics Observatory’s Code of Ethical Principles. No CSO reported receiving donor requests to disclose its use of AI in grant proposals or reports.
Shared account vulnerabilities: To save money, many of the CSOs share a single paid AI subscription (like ChatGPT Plus) among multiple team members. A cybersecurity expert warned that this creates severe security risks: it exposes individual conversation histories, eliminates accountability for what is shared, and increases the likelihood of data leaks if credentials are compromised.
Isolation in AI governance: Awareness of peer practices is very limited, with most of the CSOs developing rules in total isolation. Outliers include a youth network of 40+ organizations co-developing shared guidelines, and a social development CSO that received a draft internal policy from its US head office.
Unequal access to capacity-building: While nearly all of the CSOs express a need for it, access to AI training is uneven. Internationally connected ones and donor-supported media outlets receive regular training. Conversely, smaller, regional CSOs rely purely on self-taught experimentation and lack the time and expertise to build internal guidelines.
Risks of Informal Use
The interview findings reveal three operational risks, each one compounded by the absence of institutional governance frameworks.
Data Security and Confidentiality
The most immediate risk involves the uncontrolled sharing of sensitive information with AI platforms. These platforms typically store user inputs and their free-tier services may use conversation data to improve their models, creating the possibility that confidential information entered by one user could surface in outputs generated for others. A CSO has no control over data once it is entered. For those working with vulnerable populations or on sensitive research, a staff member entering personal information into an AI tool creates an exposure that it cannot monitor, limit, or reverse.
Content Accuracy and Overreliance
AI-generated outputs are often polished and persuasive, which can mask inaccuracies, fabricated references, or misleading summaries. For human rights CSOs producing advocacy reports or think tanks publishing policy analysis, the credibility cost of an undetected factual error is significant. The risk of overreliance is also generational and behavioral: a youth CSO specifically flagged concern that younger team members and participants in projects may substitute AI outputs for original thinking, treating AI as a source of truth rather than a supporting tool. Extensive dependence on AI-generated drafting may weaken internal analytical and writing capacity, particularly among younger staff members.
Accountability and Transparency Toward Partners
None of the CSOs have received any AI-use guidance on the likes of disclosure requirements and acceptable practices from donors or international partners. They assume their current approach is acceptable, and donors may not realize how extensively AI tools contribute to the reports, proposals, and deliverables they receive. As AI-generated content becomes harder to distinguish from human-written text, this gap in awareness poses a growing reputational risk for both sides.
Conclusion
AI governance in Armenia’s civil society is shaped more by informal adaptation than by institutional rules. That flexibility lets CSOs experiment and adopt tools quickly, but the absence of formal policies and disclosure standards puts transparency, accountability, and public trust at risk. As AI use grows, the shift from improvised norms to deliberate governance will become central to organizational credibility.
A comprehensive national AI regulatory framework is unlikely to materialize soon. Armenia’s commitments remain statements of intent rather than enforceable rules that cover civil society. For now, the only concrete governance many CSOs encounter is sectoral self-regulation, most visibly through the Media Ethics Observatory’s code, which a large share of the country’s outlets has signed. CSOs will keep using AI tools, driven by the same resource constraints and productivity pressures, regardless of whether formal governance catches up. The relevant question is whether the informal norms CSOs develop in the interim will be recognized and built upon when regulation arrives.
Recommendations
For German and EU policymakers
Ensure CSO consultation is built into AI governance technical assistance. When providing technical assistance on AI legislation to Armenia, its European partners should require that consultation with civil society – including CSOs working with vulnerable groups, in the regions, and in the media field– is built into the process. Technical assistance should also include developing practical guidance documents, toolkits, or templates that organizations without legal or IT departments can use.
Facilitate peer learning between European and Armenian CSOs. Armenian CSOs have almost no awareness of how their peers elsewhere handle AI governance. European CSO networks that have already developed AI policies could share their experience through structured exchanges, particularly through existing networks like the EaP Civil Society Forum.
For donor organizations
Provide practical support. Armenian CSOs’ donors and international partners should communicate clear expectations regarding the acceptable use of AI in the production of grant proposals and reports, but they must also match those expectations with resources. CSOs would benefit from donor assistance in three areas. First, access to paid AI tools, which offer stronger privacy protections and more reliable outputs than free versions but remain out of reach for many CSOs. Second, access to expert guidance, whether through dedicated consultants, peer-learning programs, or partnerships with organizations experienced in AI governance, which can help CSOs develop internal policies suited to their context. Third, structured capacity-building programs that go beyond basic introductions to AI and address practical questions of data security, risk assessment, and verification. Some of this support already reaches media organizations and CSOs connected to international networks, but it has not extended to the sector as a whole.
For Armenia’s policymakers
Include civil society as partners in AI policy design. As Armenia develops its national AI policy, CSOs that have accumulated practical experience should be included as partners in the process, rather than be treated just as subjects of future regulation. They bring operational experience with AI tools and, importantly, the perspectives of the communities they serve. Human rights CSOs can flag the sensitivity of the data they handle and the specific risks that poorly designed regulation could create for their work. Media organizations bring practical experience with self-regulatory approaches to AI governance.
Structured consultation with CSOs – whether through the e-draft platform, dedicated working groups, or targeted dialogues – would help ensure that the resulting framework is practical enough to be useful and proportionate enough to encourage voluntary adoption in civil society.
For Armenia’s CSOs
Develop clear internal AI guidance. CSOs do not need to rewrite their charters or to draft comprehensive formal policies immediately. A short internal memo or practical handbook is a more realistic and manageable starting point. Such a document should cover approved tools, acceptable and prohibited use cases, data-protection rules (particularly regarding personal and sensitive information), human oversight and verification expectations, and basic disclosure principles. It should be written in plain language, integrated into existing procedures, and treated as a living document that is reviewed periodically as technology and circumstances change. The Media Ethics Observatory, which incorporated AI-specific provisions into its Code of Ethical Principles, offers a practical model of sector-level self-regulation without heavy bureaucratic infrastructure.
