.jpg)
Shadow AI, in simple terms, is the unsanctioned use of AI tools by employees without formal organizational approval or oversight. In humanitarian work, this is the norm, not the exception. The 2025 Humanitarian Leadership Academy and Data Friendly Space report surveyed 2,539 practitioners in 144 countries and territories. 93% had used AI tools and 70% used them daily or weekly. Yet only 8% said AI was widely integrated in their organization, and just 22% reported a formal AI policy.
Shadow AI grows where need is high, and support is low. Humanitarian teams face rising needs and shrinking funding, and free tools such as ChatGPT, Claude, Copilot, and Perplexity are often one browser tab away. 69% of respondents use these commercial tools, mostly for reports, proposals, translation, and email, and their usage is not because of malicious behavior or bad intent, but rather because oftentimes they are overstretched.
None of this is unique to the humanitarian sector. Microsoft and LinkedIn's 2024 Work Trend Index found that 78% of AI users bring their own tools to work, and just 39% of those surveyed had received training. Humanitarians are following a pattern the tech and private sector are also grappling with.
Training has not kept pace. In the humanitarian sector, 64% reported little or no organization-led AI training, so people taught themselves: 52% through self-directed learning and 21% through informal experimentation. One practitioner interviewed for the “Artificial intelligence beyond the promise: the unseen costs of humanitarian actors’ ungoverned adoption of artificial intelligence” article in Alternatives Humanitaires put it plainly: "There is no training being provided. I think there is an assumption that everyone knows how to use it, which is not true."
Academic research points to a second driver: official tools that fall short. In a 2026 AMCIS survey of 340 employees, corporate AI inadequacy bred frustration with internal resources, one of two emotional pathways the authors link to shadow generative AI use. Vendor research makes the same point more bluntly: a big reason for shadow AI is often the lack of official tools. Organizations are slow with uptake, oftentimes for understandable reasons, but this leads organizations to depend on AI skills they never built and cannot see.
The humanitarian sector faces even more extreme risks due to the sensitive nature of humanitarian work. Some of these reasons are:
Exposed data. Practitioners described routinely putting protection data, health information, and community feedback into commercial tools without guidance. In one team handling confidential information, unvetted platforms were used for over a year before any risk assessment. For a sector built on Do No Harm, a pasted case note is a data protection incident waiting to be discovered. IBM's 2025 Cost of a Data Breach Report puts a price on this: one in five organizations surveyed attributed a breach to shadow AI, and those with widespread shadow AI paid an average of $670,000 more per breach. WFP’s data breach has led to significant concerns about how we are using humanitarian data and the AI systems that we employ. The New Humanitarian podcast stated from Marwa Fatafta, MENA policy and advocacy director at Access Now, “We're talking about a huge honeypot that can be exploited, hacked, and used potentially even to kill people,” says Marwa Fatafta. Exposed data puts the populations we are supposed to be supporting at risk.
Lost knowledge. When learning is self-directed, expertise lives in individuals. If they burn out or leave, it goes with them. Leaders also cannot tell a colleague with an advanced workflow from one who occasionally drafts emails, so support is designed for beginners who may not be the real users.
Silence. Some staff feel that using AI looks lazy or that they are not doing their job properly, so the most capable may hide their practice. The tech sector sees the same: in Microsoft and LinkedIn's survey, 52% of AI users were reluctant to admit using it for important tasks. Many in the HLA/DFS research did not want to speak on behalf of their organizations, as they did not feel comfortable or have authority to use AI in their work, as it was unsanctioned.
Inequality. Training is not reaching everyone: 33% of INGO and 38% of local NGO respondents reported none. Access, not willingness, drives the gap.
Research is coming out that argues these harms share a mechanism they call assurance erosion, and name three routes: data crosses the boundary of approved environments, tools bring capabilities nobody has assessed, and leaders lose sight of what is being used. Each is visible in the humanitarian evidence above.
A tool that leaks a spreadsheet loses money and trust; a humanitarian organization that leaks a protection case can endanger a person. Three features make shadow AI weightier in our sector. The first is the data: humanitarian data comes from people who rarely had a real choice about sharing it. The second is power: the communities affected by humanitarian decisions remain largely uninformed about systems that may increasingly shape those decisions. The third is representation: general-purpose tools are trained on data that underrepresents much of the world we work in, and as one survey respondent put it, "There is not enough data on Sub-Saharan Africa on online platforms." The HLA/DFS report found local NGOs, closest to affected communities, voiced the strongest concerns about data misuse and bias while also reporting the least training.
Efficiency is the promise that makes shadow AI hard to argue with. Practitioners report real gains in drafting, translation, summarization, and reporting. One humanitarian director told the Alternatives Humanitaires study: "If we weren't in crisis, efficiency would be about giving more to beneficiaries. Right now, efficiency is about continuing to exist."
But another practitioner asked whether saved time means more time for family or sport. The answer: "It means more time to work even more." If AI speeds up reporting while the reporting system stays excessive, little has changed except how long teams can endure it. Shadow AI makes this worse because gains captured privately are invisible. An organization cannot redirect time it cannot see toward analysis, judgment, and communities.
Restrictive policies tend to reduce the visibility of AI use rather than the use itself, a point made in the Digital Humanitarian Network's September 2026 paper Humanity in the Loop. Blocking alone is a weak defense. Governance that enables, rather than just restricts, looks more like this:
Shadow AI is not a story about rule-breaking staff. It is a signal that people are solving real problems with the tools in front of them, faster than their organizations can respond. The goal is not to switch it off but to bring it into the light, so good practice spreads, risky practice is caught, and the time saved returns to communities rather than disappearing into more reporting.
