Access Now: AI is entering humanitarian aid by the backdoor

There's a version of the AI adoption story that gets less airtime than the enterprise wins and model release races, and it's the one Access Now has documented: what happens when the technology quietly shows up inside humanitarian aid without anyone at the top signing off on it. The report, based on more than 70 interviews with actors from across the humanitarian, public, academic, private, and social impact sector, argues that algorithmic systems are entering the humanitarian space "mostly by the backdoor, and often without safeguards."

The concrete example the report leans on is small and telling. A U.S. nonprofit's chatbot, per the report, went off-script following a product update that activated unexpected AI features, leaving vulnerable users barraged with misleading responses. That's the failure mode worth internalising, because it isn't a rogue LLM or a badly written prompt; it's a routine product update, according to the report, activating AI features nobody at the deploying nonprofit expected.

Why this matters if you're not directly funding aid work: the report points at the same procurement gap that shows up everywhere AI is adopted informally. Individual workers reach for large language models to draft and translate, NGOs plug in chatbots because they scale, and the "progressive cloudification of digital systems," in the report's phrasing, is deepening the divide between large organisations with privileged access to Big Tech and smaller frontline organisations. The report notes that Global Majority aid workers and grassroots organisations are at the forefront of experimenting with AI adoption, yet lack adequate legal frameworks and safeguards.

The honest caveat is that this is a policy argument rather than an incident register. It names one anonymised chatbot failure and gestures at broader risks like autocratic surveillance and unlawful targeting of civilians in conflict, without giving you the vendor-by-vendor breakdown a procurement officer would actually need. What the reporting doesn't give you is which tools, which agencies, or who pays for the algorithmic accountability audits and open source infrastructure it recommends.

The forward-looking bet is the one the report itself makes: organisation-led development based on open source systems is presented as the best way to minimise risks associated with algorithmic systems. If donors tie that recommendation to procurement, the winners are the smaller NGOs and the vendors willing to be audited.