Published on 5th August 2026
Behind the promise of efficiency, a warning. Two authors versed in both AI and humanitarian protection dissect the dependencies, biases, hidden labour and environmental costs that consumer chatbots import into aid work. They argue that no internal AI policy can neutralise tools designed against humanitarian values.
Since its inception in 1956, the term “Artificial Intelligence” has designated a vast field of computer science, encompassing concepts and techniques dating as far back as 1943, as well as a wide array of applications. But more recently, the phrase has been co-opted to focus specifically on a smaller set of techniques (such as large language models or LLMs, trained on vast amounts of non-contextualised data) and applications (“Generative AI”); more to the point, that semantic appropriation benefits a small number of corporations and their commercial products: non-specialised chatbots such as Gemini, ChatGPT, Grok, etc. These chatbots, which underpin professional applications such as coding assistants in the software industry, or target selection in some militaries, drive the recent increase in processing demand and a consequent frenzy of data centre construction. They will constitute the focus of this article.
Introduced to the general public in the early 2020s, these chatbots swiftly impressed by mimicking human writing. The general sentiment, reinforced by aggressive marketing campaigns by vendors relayed by governments and politicians across the board, and combined with a fast deployment in systems and applications used daily by large swathes of the population, was of a major shift to which any resistance was futile. “Adapt or be replaced” became a familiar refrain. The humanitarian sector is no exception, and prospects of substantial savings can be particularly appealing while UN agencies and NGOs alike are being starved of funding and forced to helplessly watch several major crises unfolding. Yet development and humanitarian actors should beware of great promises, as generative AI has several deep flaws. This article aims to summarise the most salient ones for the sector.
ChatGPT is not our friend
The close proximity between the industrial actors of AI and the parties responsible for the collapse of the United States Agency for International Development in January 2025 should come as a warning: their interests and goals are fundamentally distinct from those of humanitarians, if not at odds.[1]“OpenAI’s Greg Brockman is a MAGA donor; here is how much he donated to the Trump campaign”, The Economic Times, 27 January 2026, … Continue reading Most major LLM publishers are based in the United States and can be subject to arbitrary injunctions and ordered to enforce sanctions at any time; this gives the US indirect and deniable leverage on institutions outside their jurisdiction, similar to that are currently targeting several judges and prosecutors of the International Criminal Court.[2]Basile Richard, « Juge français, interdiction de voyage… cinq minutes pour comprendre les nouvelles sanctions des États-Unis contre la CPI », Le Parisien, 21 August 2025, … Continue reading Digital services vendors can even feel compelled to do so by mere pressure or ideological alignment, absent any formal decision. At the moment of writing, the main alternative is Chinese counterparts to these systems, which are known for relaying talking points of the regime and for censoring references to the Tiananmen Massacre or to Taiwan.[3]Donna Lu, “We tried out DeepSeek. It worked well, until we asked it about Tiananmen Square and Taiwan”, The Guardian, 28 January 2025, … Continue reading Relying on these systems therefore creates a dependency on companies based in authoritarian States, introduces information security vulnerabilities, and puts critical operations at risk the day someone decides to pull the plug.
The myth of objectivity
A common misconception around technology is that because it is automated, it is free of subjectivity and bias, and therefore objective. On the contrary, these systems have been programmed by humans who brought their own worldview into their work, and have been trained on data produced by people on the privileged side of the digital divide, i.e., overwhelmingly white men in the Global North.[4]Mirca Madianou, Technocolonialism: When Technology for Good is Harmful, Polity Press, 2024, p. 129–136. Ipso facto, these general-purpose chatbots perpetuate, promote and launder the gender and racial biases that permeate their software and their training sets, because everything an LLM chatbot does amounts to predicting an output based on statistics and recurrence – hence the nickname of “stochastic parrot” they are sometimes given.[5]Emily M. Bender, Timnit Gebru et al., “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?, in FAccT ’21: 2021 ACM Conference on Fairness, Accountability, and Transparency, … Continue reading This also applies to image and video generators, on which some organisations have started to rely as a workaround to image rights issues.[6]Arsenii Alenichev, “AI visuals: A problem, a solution, or more of the same?”, The New Humanitarian, 28 October 2025, … Continue reading In worse cases, there have been deliberate interventions to alter a model in order to push a political agenda. Grok, a chatbot that operates on “X” (ex-Twitter) overtly normalises White supremacist talking points, to the extent that it once called itself “mecha-Hitler”; this same system is the basis for “Grokipedia”, an online encyclopaedia that plagiarises online corpora, but also accepts neo-nazi propaganda for its sources.[7]Harold Triedman and Alexios Mantzarlis, “What did Elon change? A comprehensive analysis of Grokipedia”, https://doi.org/10.48550/arXiv.2511.09685 While this is probably the most grotesque example in that matter, the lack of openness in the functioning of chatbots means that this risk is general in the whole industry.[8]Ashley Capoot and Laura Kolodny, “Musk’s Grok AI chatbot says it ‘appears that I was instructed’ to talk about ‘white genocide’”, CNBC, 15 May 2025, … Continue reading This contributes to normalising hate speech not only amongst its human readers, but also because AI-generated content is now being harvested to train the next generations of LLMs, with the risk of baking racism, misogyny, homophobia and eugenics into future systems.[9]Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao et al., “AI models collapse when trained on recursively generated data”, Nature, no. 631, pp. 755–759, 2024, … Continue reading
Moreover, the goal of service vendors is to keep customers engaged by providing them with answers that will be deemed satisfying, be it at the expense of accuracy. They are programmed not to alienate users, to the point of being notoriously “sycophantic”, and will therefore not offer any criticism or contradiction, which makes them particularly poor and unreliable for decision-making and raises accountability issues.[10]“AI overly affirms users asking for personal advice”, Stanford Report, 26 March 2026, https://news.stanford.edu/stories/2026/03/ai-advice-sycophantic-models-research Because their development is intrinsically opaque and does not follow published quality standards for their outputs, users should not trust them without triangulating with reliable sources. As chatbots rarely cite the sources they rely on to provide their answers, or even regularly make them up, it means any verification will take at least as long as conducting “old-school” research, cancelling out prospective productivity gains.[11]In the field of AI, we call a hallucination a false or misleading response presented as a certain fact, for example a bibliographic reference [Editor’s note]. Ramin Skibba, “Bosses say AI boosts … Continue reading But this methodology turns what used to be a creative endeavour into a soul-crushing race to correct and proofread the content produced at industrial scale by an automated system. This encourages cutting corners and skipping verification entirely, especially given the constant pressure of increasing productivity while resources dwindle, which LLMs help reinforce – a phenomenon dubbed “cognitive surrender”.[12]Steven D. Shaw and Gideon Nave, “Thinking–Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender”, The Wharton School Research Paper, … Continue reading Finally, the proliferation in recent years of LLM-generated websites, which are in turn fed to AI engines to train on, means that it is going to be harder and harder to separate the wheat from the chaff, to the point that LLMs are sometimes compared to asbestos, which still lingers in buildings across the world decades after the risks it poses were identified.[13]Cory Doctorow, The real (economic) AI apocalypse is nigh, 27 September 2025, https://pluralistic.net/2025/09/27/econopocalypse
On the other side, LLMs rely on hordes of human annotators, who painstakingly label images, add tags and keywords, or manually categorise texts and media; these annotators are disproportionately citizens of the Global South, to the point of ushering in the slogan “AI is African Intelligence”.[14]Jason Koebler, “‘AI Is African Intelligence’: The Workers Who Train AI Are Fighting Back”, 404 Media, 12 March 2026, … Continue reading Indeed, contrary to marketers who encourage the view that AI is some sort of magic which is going to free people from tedious tasks, the datasets need to be manually cleaned, tagged and formatted before AI engines can train on them. And these tasks are largely carried out by unprepared young graduates, many of whom live in rural communities where economic prospects are low, who endure low pay, a high pace of work, and exposure to content that can be stressful and sometimes traumatising, without being offered proper psychological support.[15]Anuj Behal, “‘In the end, you feel blank’: India’s female workers watching hours of abusive content to train AI”, The Guardian, 5 February 2026, … Continue reading In other cases, human intervention is much more significant than the service provider claims; it is merely hidden from the user.[16]Victor Tangermann, “It Turns Out That When Waymos Are Stumped, They Get Intervention From Workers in the Philippines”, Futurism, 6 February 2026, … Continue reading Furthermore, the low quality of AI-generated content (“slop”) and the frustration it generates is now being associated with the specific mannerisms of the English dialects of Global South nations, reinforcing stigma and elevating opportunity barriers for entire populations.[17]Marcus Olang’, “I’m Kenyan. I Don’t Write Like ChatGPT. ChatGPT Writes Like Me”, this man’s mind, 8 July 2025, … Continue reading This also perpetuates the exploitative North-South relationship, in which high added value created in high-income countries relies on an underpaid workforce on insecure contracts in the South, which is contrary to commitments to equitable partnerships and localisation such as those made by the Grand Bargain.
Major data protection issues
LLMs further complicate data protection, a perennially complex subject in humanitarian settings.[18]“UN Shared Rohingya Data Without Informed Consent, Human Rights Watch”, 15 June 2021, https://www.hrw.org/news/2021/06/15/un-shared-rohingya-data-without-informed-consent; Shafiur Rahman, … Continue reading Because chatbots work in a totally opaque way, subject to change without notice, on servers belonging to a third party, and can record, store and process the content of any prompt typed in their interface, users and organisations have no way to control what is going to be done with the data they feed to them. Licences and end-user agreements often grant service providers the unlimited right to use all inputs as they please: any data entered into a third-party LLM irremediably escapes the control of the user and that of its employer, including the very questions that users ask. Organisations try to mitigate this risk by entering into agreements with vendors providing a certain level of confidentiality and instructing their personnel not to enter confidential data into these tools; yet, anticipating the potential risk of data, even not clearly labelled as confidential, requires specific skills that cannot reasonably be expected from most staff members. Data deemed not sensitive when taken piecemeal can become harmful in bulk: once information fed to the system (over the years and by many organisations all relying on the same handful of providers) accumulates enough, it makes it possible to picture an organisation’s operations and infer the people it works with, paving the way to retaliation by hostile actors. The resulting sensitive intelligence is self-evidently at the disposal of the service provider, but can also fall into the hands of repressive third parties, be they State or non-State actors: LLMs have again and again been shown vulnerable to prompt attacks, and have a tendency to disclose privileged information when asked. This adds to the long-standing US and Chinese practice of leveraging industrial advantage for mass surveillance.
The LLM industry not only fails to mitigate harmful side effects, but can also directly contribute to creating humanitarian catastrophes. The bombing of Minab primary school, imputable to a faulty LLM-powered target selection made upon obsolete information, that the US military failed to double-check,[19]Marcus Weisgerber, Amrith Ramkumar and Shelby Holliday, “U.S. Strikes in Middle East Use Anthropic, Hours After Trump Ban”, Wall Street Journal, 28 February 2026. might be a harbinger of abuses to come. This is all the more worrying as Anthropic, which denied the Pentagon unconditional use of its services, poses as the comparably scrupulous actor in the industry: after Anthropic criticised US ambitions for automated weapons and for the mass surveillance of US citizens, OpenAI jumped on the opportunity to provide its services. It is worth noting that mass surveillance of non-US citizens never came into question, although it is a violation of Article 12 of the 1948 Universal Declaration of Human Rights and Article 17 of the 1966 International Covenant on Civil and Political Rights, by which the USA is bound.
Environmental impact
Finally, LLMs run in data centres that consume massive amounts of resources such as water, fossil fuels and rare earths. The race to more and more computational power far outpaces the capabilities of electric grids and power generation, leading to data centres being powered by generators and even old aviation jet engines repurposed as gas turbines.[20]Kaif Shaikh, “Data centers turn to old jet engines to power AI’s soaring energy demands”, Interesting Engineering, 21 October 2025. In the USA, these installations are disproportionately built in poor neighbourhoods[21]“The Real Safety Risks of Data Centers: What Local Communities Need to Know”, Conimby Foundation, 12 September 2025. which are ill-equipped to object, and where fire safety and environmental standard enforcement is lax.[22]Simmone Shah, “Community Backlash to AI Data Centers Is Growing Across the U.S.”, Time, 22 July 2026, https://time.com/article/2026/07/22/community-backlash-ai-data-centers”; Cecelia … Continue reading The fire safety issue is especially relevant since the backup Li-ion batteries are prone to fires,[23]Adam Barowy and Jasen Dodson, “Challenges to Incident Response at Secured Data Centers”, Security Technology, 1 June 2023, … Continue reading and since data centres concentrate large amounts of toxic, combustible material in well-ventilated enclosed spaces, forming an ideal mixture of fuel and oxidiser inside what functionally acts as an oven.[24]Iain Hoey, “Data centre fires raise questions over safety measures”, International Fire and Safety Journal, 21 October 2024. Organisations that have committed to reducing their carbon footprint should therefore think twice before deploying tools that are going to reverse the efforts they have made to reach emission reduction goals. This is particularly true for humanitarian and development actors, who regularly warn that the loss of land and livelihoods to climate disasters will make crises more frequent, more widespread and more severe. In this context, carbon footprint calculation models need updates to better model indirect emissions, such as those resulting from the use of off-site and outsourced services like data traffic and processing.
Your AI policy will not be enough
Many NGOs have introduced AI usage policies to control their use by staff members and partners. At a crucial juncture where the resources afforded to humanitarian NGOs are collapsing, this may seem to be a reasonable compromise: keeping a hand on AI-augmented work, while harnessing the increase in production that automation promises. This illusion shatters when we realise that chatbots deliver only the appearance of meaning, and that they miss the contextual connections that make information relevant. LLM output features a distinct cheapness that undermines not only their own credibility, but also the standing of their users. To name just a couple of examples of false economies, the apparent quality of popular AI outputs such as translations and generated images badly conceals a lack of understanding of the relevant environment that prevents them from delivering the prompter’s intended message, in addition to diverting work and remuneration traditionally entrusted to local actors.
Rather than cutting corners in the face of resource depletion, NGOs need to nurture their human capital before investing in subscription-based chatbots whose costs could soon spiral out of control against low returns on investment. The local knowledge that humanitarians rely on to conduct their everyday work in increasingly unsafe environments and the requested granularity level are largely out of reach of models trained on massive amounts of data, most of it was irrelevant to the context at stake. Also, the fact that the instigators of the brutal contraction in humanitarian budgets are connected to LLM vendors cannot be ignored: they are indeed sometimes the very same individuals.
Instead of chasing shadows in LLM rabbit holes, NGOs should double down on their proven strength: the know-how, intelligence and topic expertise of their human workers. While their work features an abstract formality and professionalism that LLMs can mimic, it is most of all built on empathy and contextual relevance that LLMs cannot hope to reproduce. These are the qualities that make the value of humanitarian work, and are what invites the recognition that ultimately yield funding. While an AI policy is essential to limit the most immediate threats posed by an indiscriminate use of generative AI by unsuspecting and pressured workers, the safeguards that can realistically be put in place may be mere patches on tools designed to do exactly what humanitarians are up against.
Picture credit: Sasun Bughdaryan (@sasun1990)
