November 11, 2026

AI Policies for the Humanitarian Sector: What They Are and How to Build One in 2026

Author:
Madigan Johnson

If humanitarian staff have internet access, they are likely already using generative AI: 95% have used or tried AI at least once, and 75% use it daily or weekly. The key question is whether organizations will guide this usage or leave it to chance. In humanitarian work, the risks of unregulated use are significant: data often involves people in crisis, power imbalances between organizations and affected communities are large, and a single careless prompt can cause harm.

An AI policy translates humanitarian principles into practical rules for a rapidly evolving technology. In this article, we explain what an AI policy is, outline key components of an effective policy, and recommend steps to create, launch, and maintain one. We emphasize the importance of ensuring staff understand and follow the policy. Our guidance is based on sector standards and peer-reviewed research, and we note where evidence is limited. As the field evolves, we encourage organizations to adopt best practices suited to their context.

What an AI policy is (and what it is not)

An AI policy is a short, formally approved document that states which AI uses organizations allow, which they prohibit, who decides the gray areas, and who is accountable when something goes wrong.

We find it helps to separate it from its neighbors:

  • An AI strategy outlines the organization's goals for AI implementation. An AI policy specifies what the organization will and will not do to achieve those goals.
  • Existing data protection and information security policies remain in effect. A strong AI policy should build on these, not replace them. We recommend focusing on data rather than tools, and treating AI as an extension of acceptable-use, privacy, and records policies.
  • Guidelines, SOPs, and prompt libraries provide instructions for specific tasks. These documents change frequently and operate under the policy.

A policy also serves a purpose beyond a rulebook. NetHope's 2026 outlook recommends allowing staff to experiment with AI tools while clearly defining usage limits, as individual adoption is outpacing organizational strategy and creating unmanaged "shadow AI" and cybersecurity risks. We recommend the same approach: enable use, but set clear boundaries.

The humanitarian gap

The sector has not yet widely adopted AI policies. In the January 2026 pulse survey by the Humanitarian Leadership Academy and Data Friendly Space (1,700+ respondents in 120+ countries), 75% of humanitarians used AI daily or weekly, but only 23% had a formal AI policy. Local organizations lag behind UN agencies: 13% had a policy, compared to 39%. Only 3% of respondents considered themselves expert users. These figures reflect survey respondents, not organizations. The 2025 survey noted sample bias toward digitally literate respondents and about 75% from the Global South, so these numbers indicate sector trends rather than full representativeness.

The low percentage of expert users informs our policy recommendations. Most readers are capable but not experts, so we suggest creating a policy that enables staff to make informed decisions independently.

‍What the academic research adds

Peer-reviewed research on humanitarian AI is increasing, but few studies directly assess organizational AI policies. We use the following points as a foundation for policy content and recommend adapting them to each organization's context.

  • Recurring concerns include ethics and risks. A 2026 systematic review of 60 studies in Technovation (Lythreatis, Acikgoz and Yassine) found that studies on ethics and risks cited algorithmic bias, data privacy, surveillance, fairness, accountability, and transparency. The review also questioned whether frameworks exist for auditing humanitarian AI systems for interpretability, accountability, fairness, and human oversight.
  • Privacy is the most frequently cited concern, and guidance has not kept pace. A scoping review of 98 studies on personal data and AI in humanitarian crises (Kreutzer et al., 2025, BMC Medical Ethics) found privacy was the top ethical issue (74 studies), deemed current guidelines inadequate, and called for specific guidelines, training, and auditing methods. The review's searches ended in 2020, so it predates generative AI.
  • Human rights law is a firm baseline: Pizzi, Romanoff and Engelhardt (2021) argue in the International Review of the Red Cross that international human rights law provides a globally legitimate and comprehensive framework for humanitarian AI, and that AI is an inappropriate substitute for human decision-making in highly sensitive or high-stakes contexts.
  • Skills and governance influence adoption. A mixed-methods study of generative AI in organizations (Hughes et al., 2026, 14 interviews and a survey of 304 decision-makers) found that staff skills significantly increased AI use, while interviewees expressed concern about insufficient governance. This study is not specific to the humanitarian sector, so applying its findings requires caution.
  • Risk-management frameworks are available at no cost. The US National Institute of Standards and Technology's AI Risk Management Framework (2023) and Generative AI Profile (2024) are voluntary resources. The profile emphasizes governance, content provenance, pre-deployment testing, and incident disclosure, and highlights automation bias and over-reliance as key risks.

What a humanitarian AI policy should contain

Leading published examples agree on core elements. We primarily reference the ICRC's AI policy (November 2024) and the CDAC Network's SAFE AI framework (May 2026), which we find most relevant for the sector. We recommend that a humanitarian AI policy include the following ten elements.

  1. Purpose and principles: We recommend starting from the humanitarian principles of humanity, impartiality, neutrality, and independence, and anchoring them in international human rights law. The ICRC adds proportionality, precaution and do no harm, safety and security, transparency, and accountability. SAFE AI adds a right to know, a right to understand, and a right to contest AI-influenced decisions.
  2. Scope: We recommend saying clearly who and what is covered. For example, the ICRC applies its policy to all staff, activities, internal services, operational responses, and partnerships, as well as to generative AI, machine learning, computer vision, and predictive modeling.
  3. Acceptable and prohibited uses: We recommend naming both. The ICRC lists prohibited uses such as manipulative techniques, social scoring, and untargeted facial-recognition scraping. Each organization's list will reflect its own mandate and context.
  4. Risk tiers: We recommend distinguishing between low- and high-risk activities, such as drafting a meeting summary versus deciding who receives assistance. The NIST AI Risk Management Framework also uses a risk-based approach.
  5. Data rules: Clearly state what information must never be entered into unvetted tools. A simple guideline is: if staff would not email the data, they should not input it into an AI tool. We also recommend conducting data protection impact assessments for higher-risk systems, as required by the ICRC. This section requires special attention.
  6. Human oversight and accountability: The ICRC states that responsibility for AI-supported decisions rests with the organization and its staff, not the AI system. We recommend the same. SAFE AI advises assigning clear internal accountability instead of delegating it to suppliers, and including right-to-audit clauses in AI procurement contracts; we support both recommendations. NIST's Generative AI Profile highlights automation bias and over-reliance, so we also recommend specifying what reviewers must check.
  7. Disclosure: Require staff to disclose when AI has informed an assessment, targeting decision, or report.
  8. Bias and context checks: EvalCommunity's summary of DFS/HLA research notes that commercial tools are mainly trained on data from OECD countries and may perform worse in crisis settings. We therefore recommend conducting contextual bias assessments for high-risk uses.
  9. A right to refuse: SAFE AI introduces "responsible refusal," and DFS and HLA's pulse-survey commentary emphasizes the need for a right to say no. We recommend including this right. The ICRC allows affected people to opt out and use complaint mechanisms.
  10. Roles, incident reporting, and review: Name a policy owner, establish a process for reporting issues, and set a review date. NIST's Generative AI Profile identifies incident disclosure as a key focus area.

How to create one in 2026

Three trends are notable in 2026. Staff now use more than chatbots: 30% of HLA pulse survey respondents reported using custom-built AI agents. Donor expectations are increasing, with NetHope noting that requirements for safe and responsible AI use are becoming clearer and stricter. The sector also has established reference frameworks, so organizations seldom need to start from scratch. We recommend the following sequence, based on sector guidance and research reviewed for this article. Keep the process lean; a concise, readable policy is more effective than a lengthy document that is ignored. Organizations should adapt these steps to their specific context.

  1. Name a sponsor and a working group: Appoint a senior executive to own decisions and approve high-risk uses, supported by a cross-functional group including programs, protection and safeguarding, MEAL, IT and security, legal, HR, and country offices. The group should meet regularly to gather feedback and review the policy, with legal and HR leading approval and distribution.
  2. Find out what people are already doing: Inventory current uses, not just planned projects, and base your approach on existing practices. Offer amnesty first to encourage staff to disclose their use rather than conceal it.
  3. Borrow before building: Start with the ICRC policy and the SAFE AI framework. NetHope's 2026 outlook also references a Humanitarian AI Code of Conduct, which we have not reviewed here. Use the NIST AI Risk Management Framework to cross-check risk and governance, adapt the structure to your organization's mandate, and align with existing data protection, safeguarding, and acceptable-use policies.
  4. Plan the consultation: Define for each stakeholder group what input is needed, at which stage, and which policy sections affect them. Include country teams and local partners from the outset, as local actors should co-design rather than only test.
  5. Draft short and in plain language: Include principles, red lines, risk tiers, and roles in the policy. Move tool lists, prompt tips, and how-to steps into separate, updateable guidance.
  6. Test it on real cases: Use five or six actual examples from the inventory to test the draft policy and check if staff reach consistent conclusions. If not, revise unclear wording.
  7. Approve it formally: Obtain executive or board approval, record the version and date, and assign a policy owner.
  8. Set the review date before publishing: Decide in advance when the policy will be reviewed and what events, such as new donor requirements or significant incidents, would trigger an earlier review.

Getting people to know it, understand it, and use it

A policy staff hasn't read cannot be effective, and policies are easily forgotten. Provide training early and regularly, and plan for rollout and ongoing training from the outset.

The HLA's June 2026 podcast with Data Friendly Space highlights that a policy "doesn't end up helping" if practitioners do not understand it. Make AI literacy as standard as safeguarding training. The goal is not to make everyone an expert, but to equip staff with enough understanding to ask key questions, such as whether to trust an output and who needs to be informed of its use.

Each rollout will differ. Announce the policy through top leadership in accessible, non-technical language, using multiple channels such as intranet, email, and staff meetings, and provide a way for staff to ask questions. In humanitarian organizations, include channels teams actually use, such as WhatsApp groups or all-hands calls with country offices.

Make it easy to find and easy to ask:

  • Maintain a single location for the policy, FAQ, and approved-tools list.
  • Designate a contact for uncertain cases, so staff consult before entering data into a tool.
  • Provide a space to share good practices, such as a shared prompt library or chat group, which also gives the organization visibility into staff activities.
  • Ask staff to sign or formally acknowledge the policy.

Organizations should train by roles and then repeat. Offer role-specific sessions, interactive formats such as quizzes and case studies, and regular refreshers. Train new hires within their first month, refresh annually, and hold quarterly live Q&A sessions by group, such as IT, legal, and communications. For humanitarian teams, use real scenarios to build case studies and budget for this work. The HLA notes that convening and peer learning should be a dedicated budget line, not absorbed into overhead. Include these costs in proposals.

Make it safe to speak up and close the loop. Avoid punitive rollouts, as they drive use into “shadow AI”. Start with amnesty, emphasize non-punitive incident reporting with a short reporting window, provide anonymous reporting channels with clear processes, and apply proportionate consequences consistently. Focus on clarity, education, and risk reduction rather than relying solely on prohibition or enforcement. Options to collect feedback include surveys, open forums, and suggestion boxes, and inform staff how their input influenced the policy.

A potential dissemination plan at a glance

  1. Leadership announcement in plain language: Announce the policy to all staff at launch, using accessible, non-technical language to explain its purpose and expectations.
  2. One central location for the policy, FAQ, and approved tools: Make these resources accessible to all staff from a single location at launch and keep them updated as changes occur.
  3. Role-based training with real cases: Provide practical, role-specific training for programs, protection, MEAL, communications, IT, and management teams. Include training during the first month for new hires and annual refresher sessions.
  4.  Live Q&A sessions by team: Organize quarterly question-and-answer sessions for individual functions and country offices to address concerns and clarify how the policy applies to their work.
  5. Signed acknowledgment: Ask staff, contractors, and volunteers to formally acknowledge the policy during onboarding.
  6. Amnesty and non-punitive incident reporting: Introduce these measures from the policy launch to encourage staff to disclose existing AI use and report issues without fear of disproportionate punishment.
  7. Feedback survey and a "You said, we changed" update: Collect feedback from all staff after each policy review and communicate how their input has informed revisions.

For humanitarian teams, translate the policy and training into the working languages of country offices, and share them with implementing partners at the same time as staff. This ensures local organizations are not left to interpret the rules on their own.

Keeping the policy alive

Treat the policy as a living instrument. The ICRC describes its policy this way, and SAFE AI refers to its Transparency Card as a central living document. Documented AI incidents increased from 233 in 2024 to 362 in 2025, according to the Stanford AI Index figures cited in HLA's 2026 commentary. Staff tools will also change frequently. We recommend the following practices:

  • Scheduled review: Conduct an annual or bi-annual review by a dedicated committee or task force, with dates set in advance.
  • Triggered review: SAFE AI lists model updates, scope changes, and newly discovereTriggered review: SAFE AI lists model updates, scope changes, and newly discovered failure modes as reasons to reassess a system. Add new donor requirements and any serious incidents as additional triggers.th the policy lead responsible. It is also the easiest way to see which uses sit in which risk tier.
  • Named roles: Assign an executive sponsor, a policy lead, and department liaisons.
  • Incident disclosure: NIST's Generative AI Profile lists incident disclosure as a key focus area. Decide who must be informed about an incident, the required timeframe, and whether affected individuals or donors need to be notified.
  • A short annual report to leadership or the board: Cover the tools register, approvals, training completion, and incidents. These four items provide a solid starting point.

Record organizational learnings. Each question to the named contact, incident, or exception request indicates areas where the policy may be unclear or outdated.

Pitfalls to avoid

  • Writing a good policy and stopping there: An unenforced policy is ineffective. Budget time for training and follow-up, not just drafting.
  • Making it too long or too technical: If staff cannot determine whether a use is allowed, they may guess or stop asking. Keep the policy concise and place detailed information in guidance documents.
  • Rolling it out as a crackdown: Fear of punishment leads to hidden use. Start with amnesty and non-punitive reporting.
  • Leaving it to IT or legal alone: Silos and overlooked practical details are common pitfalls. Involve program, protection, and country teams, as they understand the real risks.
  • Designing at headquarters and exporting it: Local organizations often lag behind UN agencies on policies, but smaller organizations may adopt AI more quickly. Involve them in policy design and share the policy with partners.
  • Handing accountability to a vendor: SAFE AI says to keep accountability internal and secure right-to-audit clauses when buying AI services, and we agree.
  • Treating the policy as a substitute for data governance: NetHope warns that AI will not resolve poor data practices, and SAFE AI states it is not a data governance framework. Ensure the AI policy and data protection rules are aligned.
  • Writing it once: A policy that is not regularly reviewed becomes outdated, which is a significant risk.

Where to start this month

Do not wait for a perfect policy. Organizations can start by learning how teams already use AI, appointing a sponsor and a small cross-functional group, and drafting a concise policy based on the ICRC and SAFE AI structures. Invest as much effort in communicating and ensuring understanding as in drafting. In a sector where most staff use AI regularly but few organizations have a formal policy, a clear, concise, and well-communicated document provides a real advantage for those served.

Sources

← Back to News