AI ethics and humanitarian principles: do no harm, bias, transparency

1The humanitarian principles and AI

The four humanitarian principles also apply to digital tools:

• Humanity: the tool must serve affected people, not only internal efficiency. Saving time on a report is useful if that time goes back to programmes.
• Impartiality: AI must not introduce discrimination in access to aid (for example by ranking households according to biased criteria).
• Neutrality: generated text may, unnoticed, adopt the vocabulary of one party to a conflict. Review the names of places, groups and events.
• Independence: depending entirely on one provider for a critical function is a risk; keep the ability to work without the tool.

The do no harm principle requires asking before each new use: who could be harmed if the tool is wrong, if the data leaks, or if the result is misunderstood?

2Bias: where it comes from, how to spot it

Models learn from texts mostly produced in a few languages and a few regions of the world. This results in biases:

• Language bias: lower quality in languages with little presence on the Internet (many African, Asian or Indigenous languages), and sometimes in certain varieties of Arabic or French.
• Perspective bias: a "Global North" reading of crises, standardised solutions, little knowledge of local actors and national NGOs.
• Stereotypes: gendered or cultural representations (the "doctor" is a man, the "beneficiary" is passive) in texts and especially in generated images.
• Sycophancy bias: a tendency to agree with the person asking. If you ask "confirm that our project is a success", the assistant is likely to confirm.

To spot them: ask for the opposite position ("what arguments go against this conclusion?"), have texts for communities reviewed by people from those communities, and compare with local sources.

3Decisions about people: the line not to cross

Decisions that directly affect a person's rights or access to aid must not be entrusted to a generative AI assistant:

• selection or exclusion of beneficiaries;
• assessing the credibility of an account (asylum claim, complaint, protection case);
• shortlisting or rejecting job applicants;
• disciplinary or investigation decisions.

These tools can help prepare (draft an interview grid, structure published criteria, summarise a policy), but the decision belongs to authorised people, following a documented procedure, with the possibility of appeal. In the European Union, the Artificial Intelligence Act (AI Act, adopted in 2024 and applied gradually) classifies certain uses such as recruitment or access to essential services as "high-risk", with stronger obligations.

Test: if the person concerned learned that the decision had been made by an AI, would that be acceptable, and could they challenge it? If the answer is no, do not use AI to decide.

4Transparency and accountability

Trust is the capital of a humanitarian organisation. A few simple rules protect it:

• Human responsibility: every document has a human author responsible for its content, whether or not they were helped by AI.
• Proportionate transparency: disclose the use of AI when it is significant or when the recipient can reasonably expect it (evaluation, research, published content, exchanges with communities). For a spelling check, it is not necessary.
• Respect for donor and partner rules: some donors and calls for proposals regulate or require disclosure of AI use in applications; read the terms (week 6).
• No impersonation: never generate an image, voice or quote falsely attributed to a real person, and never present a generated image as a field photo.

The sector's ethical communication guidelines (dignity, informed consent, no pity-based imagery) apply fully to content produced with AI.

5Copyright, plagiarism and integrity

Three points to watch:

• Other people's content: asking AI to "rewrite" another organisation's report to present it as yours is still plagiarism. Cite your sources.
• Generated content: the legal status of generated texts and images varies between countries and is evolving. The terms of use of Claude and Gemini specify the user's rights over the outputs; for commercial use or a major publication, consult them.
• Academic and professional integrity: for the exams of this Mini MBA, as for a recruitment test, using AI to answer in your place is fraud. The aim is for you to master the skills, not the tool alone.

For photos, the organisation's rule remains: no photo of others without rights, no photo of an identifiable child without consent, and a generated image is clearly labelled as such.

6Environmental cost and equity of access

AI consumes electricity and water in data centres. Estimates per request vary widely between studies and models; the sector's overall impact is growing fast. For an organisation committed on climate, reasonable habits:

• use AI when it brings real value, not by reflex for trivial tasks;
• prefer one well-prepared document to dozens of successive attempts;
• include digital use in the environmental policy.

Equity of access: field teams often have weak connections, old devices and less access to paid subscriptions. An organisation that reserves AI for headquarters and international staff widens inequalities. A fair policy provides access, training and languages for national staff and local partners.