AI and accountability: trust can't be automated
Picture a humanitarian helpline that never closes. A voice that answers at any hour, in the caller's own language. It takes down needs, points people to the nearest service, logs a complaint. It never loses patience and never judges. Everything suggests someone is listening. But there is no one there.
This is no longer science fiction. Large language models have made possible, in a matter of months, what seemed years away. And aid organisations, caught between soaring needs and shrinking budgets, are understandably tempted.
In its 32nd issue, devoted to artificial intelligence, the journal Humanitarian Alternatives (Alternatives Humanitaires) publishes an analysis by three well-known practitioners: Charly Pierluigi (Groupe URD), Maëve de France (CartONG) and Simon Weiss (Humanity & Inclusion). Their question goes straight to the heart of our work: does AI strengthen communities' trust and participation, or does it risk hollowing them out?
Their answer is nuanced, which is exactly what makes it valuable. Here are its main threads, along with our own reading as practitioners.
Two ways AI meets communities
The authors separate AI use into two families, which raise different questions.
Facing people: chatbots and beyond
The most visible is the direct interface. For someone caught in a crisis, the most tangible change is being able to ask a digital service a question, on a channel they already use, and get a quick answer.
Behind a very similar conversational surface, the technical choices differ sharply:
- Scripted chatbots guide users through menus and pre-approved answers. The American Red Cross chatbot, Clara, works this way.
- Chatbots built on large language models handle freely worded questions in real time and in many languages. Signpost, launched in 2015 by the International Rescue Committee and Mercy Corps, has evolved into Signpost AI, which has run on several language models since 2024.
Both approaches meet real needs: one gives more control over the information delivered, the other more flexibility. But the authors stress a crucial point: a fluent, convincing answer is not necessarily an accurate one. Even when a model is connected to a document base approved by the organisation, that alone does not guarantee the answer is right, or suitable for a sensitive situation.
AI also opens new ways of listening: voice notes, free text, conversations over WhatsApp or interactive voice response. Machine translation and speech recognition can make these channels accessible to people who cannot read, or who live with a disability. They can also, the authors warn, deepen the exclusion of those who lack safe access to digital tools.
Behind the scenes: sorting, translating, analysing
The second family is less visible, and perhaps more transformative. Transcription, translation, classification and structuring now happen almost instantly, which means teams can process volumes of feedback no one could have handled before. Voice messages in local languages, long hard to use, become analysable at scale.
Beyond processing, AI helps with interpretation: spotting trends, prioritising needs, picking up weak signals, connecting sources. It also feeds anticipation. GiveDirectly, for example, uses flood-risk signals in Bangladesh and Nigeria to trigger cash transfers before disaster strikes.
This is the key shift. As the authors put it, "AI does not just process data from communities; it also plays a role in how this data is interpreted, prioritised and translated into action." In other words, it is no longer simply supporting decisions. It is starting to shape them.
Trust: the precondition for accountability
Accountability to affected people rests on the responsible use of power. And that power is only legitimate on one condition: trust. Trust is what leads communities to share their data, their needs and their complaints.
AI can strengthen that trust, or damage it. Here the authors point to the SOLIS Bot in Lebanon: a significant share of users said that, on certain topics, they would rather interact with the chatbot than with a staff member. That result, they note, reflects not a rejection of human contact but a form of trust in the organisation itself, which makes people more willing to share information.
They also highlight a deliberate choice: no generative AI in direct contact with users, but a decision tree whose every message is written and approved by humanitarian staff. We made that choice from the start, and this analysis captures why. In critical situations such as asylum claims, access to essential services or health information, a wrong answer generated by AI can do serious harm and permanently undermine an organisation's legitimacy.
Trust takes a long time to build, and very little to lose.
Genuine participation, or "participation washing"?
If trust underpins accountability, genuine participation underpins trust. This, according to the authors, is where AI raises the hardest questions.
Co-design from the outset. Involving communities from the moment the problem is defined, not just at the testing stage, is the first condition for a legitimate deployment. Otherwise the tool reflects the needs organisations project onto people, rather than the needs people actually have.
Correct for language and cultural bias. AI models are largely trained on Western, English-language content. They risk making people in the Global South invisible and missing the local subtleties of language. Initiatives such as Masakhane, focused on African languages, or Clear Global's "4 billion conversations" project aim to close that gap.
Rebalance power. AI must not become an instrument of extraction: collecting data with nothing in return, deciding remotely on the basis of opaque algorithms, or keeping up the illusion of consultation with no real influence over programme choices. The authors call this risk "participation washing", and it is all the more insidious because digital tools lend it an air of modernity and inclusiveness. They go as far as to warn of a form of "technocolonialism".
Keep a human within reach. AI can free up staff time and handle large volumes. But access to a humanitarian worker must remain possible whenever a request involves a right, protection or a sensitive situation. In their words: "Technology can underpin the relationship, but cannot replace the responsibility to respond."
Where the friction lies
The authors identify several tensions that genuine enthusiasm for digital transformation should not be allowed to hide.
The techno-solutionist trap
Technology does not fix the problems that predate it: the digital divide, unequal access, skewed representation in data. It can even mask them. "Using AI risks shifting attention and resources from the human relationship to the tool," they write, "while creating the illusion of modernised and fantasised accountability."
Listening is more than recording
Automated sorting of large volumes of feedback is genuinely useful. But a machine cannot read between the lines, put a request in context or show someone they have been heard. For the authors, listening "is also an act of recognition." When a chatbot becomes the main, or only, entry point to a feedback mechanism, the organisation makes access to a machine the condition for being listened to.
That tension is sharpest in protection and psychosocial support. An interface can inform, refer or flag an emergency, "but it cannot take on the role of a support relationship."
Who controls the infrastructure?
Computing power and hosting are concentrated in the hands of a few private companies, mostly American. Yet responsible data management means being able to explain, at any moment, how data is collected, stored, processed, shared and deleted. When that chain depends on opaque, hard-to-audit providers, the trust promised to communities rests on infrastructure the organisation no longer controls.
The authors also underline that these are dual-use technologies. According to several investigations they cite, components close to those now used by aid organisations have reportedly been used to generate military targeting recommendations at scale. That proximity is no reason to reject AI outright, but it demands a clear-eyed look at what the sector is handing over, and to whom.
Other costs stay out of sight: annotation and moderation work often outsourced under precarious conditions, the environmental footprint of systems deployed at scale, and ethical safeguards designed first for the US market, which offer uncertain protection elsewhere.
A skills gap
The answer cannot be purely legal. It requires in-house expertise: people able to audit data flows, deploy models, assess their limits and negotiate from an informed position. Alternatives exist. The best open-weight models now catch up with the most advanced closed models within months, and allow self-controlled hosting and sector-wide pooling. The real question, the authors remind us, is not only which tool to use, but "who controls the technical conditions that underpin trust."
According to a March 2026 study they cite, only 35.7% of the organisations surveyed have a formal AI policy. The challenge is even greater for local and national NGOs, whose role is growing just as the sector's technology is becoming more complex.
The questions worth asking
The core of the analysis comes down to one idea: accountability is not the perfection of an always-on service. It is an organisation's capacity to acknowledge its limits, to say what it can do, and above all to answer for what it cannot. A machine recognises neither doubt nor error.
For the authors, confusing the production of text with actual communication reduces accountability to a technical function, "when it is actually a political act." Before any deployment, they suggest asking:
- Why use AI at all?
- In place of what?
- For whose benefit?
- How can communities challenge the decisions that affect them?
They set out clear principles: do no harm, obtain consent, minimise data, and never allocate aid on the basis of an automated decision. Keep communities in the loop from the design stage, as participants rather than data sources. Maintain real access to a human for protection and fundamental rights, accepting that this access has a cost. Invest in in-house skills.
What this means for us
This analysis echoes choices we made early on with the SOLIS Bot, and it helps us put them into words.
No generative AI facing people. Every message an affected person receives has been written and approved by a humanitarian team. We give up some flexibility to guarantee that what is said is accurate, and that we can answer for it.
A complement, never a substitute. The bot works alongside field visits, phone lines and complaint boxes. It absorbs simple requests to free up human time for the ones that need it.
The bot never decides on its own. A crossed alert threshold sends a notification to staff, never an automatic action. No assistance is ever allocated on the basis of automated processing.
Minimise, always. We collect only what is strictly necessary, with consent, and some data never passes through this channel at all. The infrastructure is hosted in France, GDPR-compliant and regularly audited.
A shared resource, not a product. Pooling the tool across organisations, including local NGOs, is also an answer to the question of control: who runs the infrastructure, who decides how it evolves, and who it is accountable to.
None of this settles everything. Gaps in digital access are real, co-designing with communities takes time projects don't always have, and the pull towards more automation will only grow as budgets tighten. But these choices give us a direction.
Deciding together what we hand over to machines
Artificial intelligence has entered the humanitarian sector and it is not going to leave. The question is no longer whether to use it, but to decide collectively what can be entrusted to it. That decision, the authors point out, is itself an act of accountability.
Their conclusion deserves to be quoted in full: "The challenge is not about humanising machines, but about not dehumanising humanitarian action."
We share that conviction. Whether you build these tools, evaluate them or are still deciding whether to take the plunge, we'd be glad to compare notes: solisbot@solidarites.org.
This article draws on "Humanitarian organisations and affected communities: a relationship put to the test by artificial intelligence" by Charly Pierluigi, Maëve de France and Simon Weiss, published on 29 July 2026 in Humanitarian Alternatives, issue 32, "Artificial intelligence: uses, tensions and issues". Read the original article on Humanitarian Alternatives