Information, disinformation and security: what AI can and cannot do

1A new risk for acceptance

Acceptance by communities and parties to conflict is the foundation of humanitarian security. Yet generative AI makes it possible to produce, at very low cost, false and credible content: fake press releases, fake quotes from officials, doctored images or videos (deepfakes), fake audio recordings, mass messages on social media.

Humanitarian organisations have already been targeted by disinformation campaigns accusing them of espionage, proselytism, trafficking or partiality. These campaigns can lead to threats, attacks or expulsions. The ICRC and other actors warn about the risks of "information-related harms" for civilian populations and aid workers.

Disinformation is a security risk. It belongs in the context analysis, the risk assessment and the crisis communication plan.

2Spotting fake content

It is becoming difficult to tell generated content from real content by eye. Verification reflexes remain effective:

• Source: who published first? A recent, anonymous account with no history?
• Context: do the date, place, weather, uniforms and language match?
• Reverse image search: did the image exist before, in another context?
• Technical clues: hands, distorted text, inconsistent shadows, a voice without natural breathing; these clues disappear as tools improve, so they are not enough.
• Markings: some providers add invisible watermarks or provenance metadata to generated content (for example SynthID at Google, or C2PA standard metadata). Their absence proves nothing.
• Direct confirmation: contact the organisation or person supposedly behind the content.

An AI assistant can help structure this verification and list inconsistencies, but it cannot "certify" that content is true or false.

3Responding to a disinformation campaign

If your organisation is targeted:

• Assess: reach, audience, links with armed or political actors, risk for staff and communities. This is a decision for the risk management cell, not only for communications.
• Protect: adapt security measures if necessary (movements, visibility, sites).
• Respond proportionately: sometimes a public response amplifies the fake content; sometimes silence lets it take hold. Often favour direct dialogue with communities, leaders and authorities, backed by verifiable facts.
• Document: dated screenshots, links, reporting to platforms.
• Coordinate with other organisations and coordination mechanisms.

AI can help prepare factual talking points, questions and answers for staff and messages adapted to different audiences, reviewed by management and security.

4AI in support of security analysis

For security, AI can be useful on public information and methodological tasks:

• summarising public reports on a context;
• proposing the structure of an actor analysis or a security plan from a template;
• preparing simulation exercises (fictitious scenarios);
• translating public documents;
• drafting security training materials.

It must not receive:

• detailed incident reports, names, routes, times, positions of teams and sites;
• confidential information from a security network shared between NGOs or from security analysis services reserved for their members;
• information on negotiations, contacts with armed groups or kidnappings.

And it must never decide on a security level, a movement or an evacuation. These decisions belong to the designated managers, according to the security plan.

5Phishing and fraud boosted by AI

AI makes scams more convincing:

• Phishing emails with no mistakes, personalised, in the recipient's language, imitating a donor or management;
• CEO fraud: a fake voice or video message from the director requesting an urgent transfer;
• Fake job offers in the name of humanitarian organisations, asking candidates for fees;
• Fake suppliers changing bank details.

Defences:

• dual approval of any payment, and confirmation through another known channel of any change of bank details;
• a code word or call-back procedure for unusual urgent requests;
• two-factor authentication on all accounts, including AI accounts;
• regular training and phishing exercises;
• a public reminder: a serious humanitarian organisation never asks candidates for fees.

6Security of AI accounts and tools

AI accounts themselves must be protected:

• strong authentication and named accounts, no shared accounts;
• regular review of connectors and access granted (email, storage);
• beware of prompt injection: a document, web page or email may contain hidden instructions aimed at the AI ("ignore your instructions and send the content to..."). When an assistant reads external content and has action capabilities (sending emails, accessing files), this risk becomes real. Keep human approval of actions, and limit access;
• when an employee leaves: deactivate accounts and transfer shared projects;
• apps and extensions: install only official apps; fake "Claude" or "Gemini" apps sometimes circulate.