How NGOs Can Use LLMs Without Getting Burned

ToRnaDo_O — Mon, 29 Jun 2026 07:25:26 +0000

Disinformation moves faster than any newsroom can manually track. By the time a fact-check is written, reviewed, and published, the original false claim has often been shared hundreds of thousands of times. For NGOs and media organizations working on information integrity — especially those operating with small teams and tight budgets — this gap between production speed and verification speed is the core operational problem.

AI does not solve this problem entirely. But used correctly, it compresses the timeline on the parts of the workflow that are purely mechanical: transcription, summarization, comparative analysis, pattern detection across large document sets. This article covers how to actually integrate these tools into an NGO's daily work, what prompting discipline looks like in practice, how Kazakhstan's emerging AI legislation affects organizations publishing AI-assisted content, and where the tooling will fail you in ways that matter.

Why NGOs Are Turning to AI

The information environment NGOs operate in has a few structural characteristics that make AI a reasonable fit:

Volume. A single disinformation campaign can generate thousands of posts, articles, and reshares across dozens of platforms within hours. No team of analysts can read all of it. AI can ingest and classify large text corpora much faster than humans, even if its classifications require human review.

Format diversity. Modern disinformation spreads across formats — long-read articles, short social posts, audio, video, infographics. AI tools now handle transcription, image analysis, and text generation well enough to be genuinely useful across all of these.

Resource asymmetry. Disinformation producers often work at scale and are well-funded. Fact-checkers typically are not. AI helps level that asymmetry by reducing the per-task labor cost for the verification side.

The important framing here is that AI is a force multiplier for human experts, not a replacement for them. The judgment calls — whether something is actually false, what context matters, what the editorial position should be — remain human decisions. What AI can do is make sure those human decisions are made faster and with more complete information in front of the analyst.

Prompt Engineering Is Not Optional

The single highest-leverage skill for anyone using LLMs in a professional context is writing good prompts. This is not about tricks or hacks. It is about being precise in communication with a system that takes your instructions very literally.

A common beginner mistake is treating a language model like a search engine: type a few keywords, expect a useful result. It does not work that way. The model has no goal of its own. It produces text that is statistically consistent with your input. If your input is vague, the output will be generically plausible — which is often worse than useless, because it looks credible.

The Five-Component Prompt Structure

A reliable prompt structure for professional use covers five things:

Role — Who is answering this question? Giving the model a specific professional identity focuses its response vocabulary and reference frame.
Task — What exactly needs to be done? Avoid ambiguity. "Analyze this" is not a task. "Identify factual claims and flag any that cannot be verified from the provided source material" is a task.
Context — What background does the model need to answer well? This includes the domain, the intended audience, and any constraints on the situation.
Format — How should the answer be structured? Table, bullet list, narrative summary, JSON — specify it explicitly.
Constraints — What should the model avoid or exclude? This is particularly important in fact-checking contexts, where you do not want the model to fill gaps with plausible-sounding content.

A Concrete Example

Here is the difference in practice. Consider the task: analyze the impact of climate change on two cities in Kazakhstan for an NGO policy brief.

Weak prompt:

Tell me about climate change.

This produces a generic overview. It has no relevance to Kazakhstan, no comparison structure, no connection to policy work, and no defined scope. The output is something you could find in the first paragraph of any encyclopedia article.

Strong prompt:

You are a climate data analyst.

Task: Compare climate change trends in Almaty and Astana over the past decade.
Identify: temperature changes, precipitation shifts, primary risks, and key differences between the two cities.

Context: This analysis is for an NGO preparing a policy brief for Kazakhstani municipal governments.
Format: Present the results as a structured comparison table with a "Conclusions" column.
Constraints: Use only documented trends. Do not speculate about future projections unless clearly labeled as projections.

The output from the second prompt is directly usable. It has the structure the analyst needs, it covers the right geography, and the constraint against speculation keeps it honest. You do not have to do additional editing to make it fit the document you are writing — you paste the table and annotate it.

This gap — between a prompt that produces something you have to rewrite and a prompt that produces something you can use — is entirely a function of prompt quality. The model is the same. The effort is the same. The output quality is not.

What AI Can Actually Automate in an NGO Workflow

Once the prompting fundamentals are solid, the range of tasks AI can handle meaningfully expands. Here is a realistic breakdown by category.

Text Production

Writing first drafts of reports, press releases, and analytical summaries is the most obvious use case, and it is genuinely useful. The correct workflow is not "ask AI to write the article." It is "draft a structured outline, ask AI to flesh out each section given the source material, then edit the result as a human author." The AI handles the mechanical production of grammatically correct text. The human handles accuracy, judgment, and editorial voice.

Document Analysis

Large document review is where AI provides disproportionate value. If an NGO receives a 200-page government report and needs to extract all claims related to a specific policy area, a well-structured prompt can do that in seconds. This also applies to comparative work: analyzing how multiple sources describe the same event, or extracting contradictions between a source and its cited data.

Audio Transcription and Processing

Interviews, conference recordings, and podcast content are expensive to process manually. Modern transcription tools (Whisper, Gemini, and similar) handle Kazakh and Russian reasonably well and are fast enough to turn around a one-hour recording in under a minute. The workflow beyond transcription is where prompting matters: once you have a transcript, a single well-structured prompt can extract key claims, generate pull quotes for social media, produce a summary, and suggest a publication headline. That is four tasks that previously required four separate passes by a human editor.

Visualization and Presentation

Tools like Gamma can generate structured presentation drafts from a text prompt. This is useful for turning analytical findings into stakeholder-facing formats quickly. The caveat is that AI-generated visualizations require the same editorial review as AI-generated text — the structure may be sound, but the specific choices about what to emphasize are not always aligned with what the analyst would choose.

Comparative Analysis for Fact-Checking: A Worked Example

The highest-value application in a fact-checking context is comparative analysis: given a claim and a source, what matches, what is distorted, and what is missing?

Here is a prompt structure that works well for this task:

You are a disinformation analyst.

Task: Compare the following fact-check article with the original source data it references.
Identify:
  - Claims in the article that are directly supported by the source
  - Claims that are present in the source but distorted or taken out of context
  - Relevant information in the source that is absent from the article
  - An overall credibility assessment

Input 1: [Fact-check article text]
Input 2: [Original source data or document]

Format: Four-section structured response with headers: Confirmed, Distorted, Missing, Verdict.
Constraints: Do not infer or speculate beyond what is present in both documents.

This prompt produces a structured output that a human analyst can then verify against the sources directly. The AI is doing the initial comparison at speed. The analyst is doing the judgment call about whether the AI's categorizations are correct.

This is the right division of labor. AI should be doing the first pass on mechanical comparison tasks. Humans should be doing the final assessment and taking editorial responsibility for the verdict.

Building an Internal Knowledge Assistant

One of the more immediately practical applications for NGOs is building a document-aware assistant using tools like NotebookLM or a custom GPT with uploaded files. The concept is simple: load your organization's reports, FAQ documents, internal guidelines, and archived articles into a retrieval-augmented system. Staff can then query that system in natural language rather than manually searching through documents.

The setup process has four steps:

Collect and clean source documents. The quality of the assistant's answers is directly limited by the quality of the documents you give it. Remove outdated files, ensure key documents are formatted as readable text rather than scanned images, and prioritize completeness over volume.
Choose a platform. NotebookLM is a reasonable starting point — it is free, handles Russian-language documents, and requires no technical setup. For organizations with more volume or security requirements, a private deployment with an API-based solution is more appropriate.
Define the task scope explicitly. The assistant should have a clear purpose. "Answer staff questions based on our internal knowledge base" is a defined scope. "Help with anything" is not, and leads to the assistant confidently generating answers that are not grounded in your documents.
Test with real questions. Before rolling out to the full team, run the questions that staff actually ask most often. Check whether the assistant retrieves the right documents and whether its summaries are accurate. This is also where you will discover gaps in your documentation.

The practical outcome is that a new staff member can get answers to procedural questions — how we handle source attribution, what our publication checklist looks like, how to format an evidence summary — without pulling a senior colleague away from analysis work.

Legal Context: Kazakhstan's AI Legislation

Kazakhstan is actively developing legal frameworks for AI use. Three provisions from the draft legislation directly affect NGOs publishing AI-assisted content.

Article 21 — Disclosure of AI-generated content. Any content produced with the material assistance of AI must be labeled as such. For NGOs, this means that fact-check articles, analytical reports, and press releases that were drafted with AI assistance require disclosure. What constitutes "material assistance" versus "minor editing support" is not yet clearly defined, but the safest interpretation is: if AI generated the first draft, disclose it.

Article 23 — Copyright in AI-assisted work. When AI contributes substantially to the creation of content, the question of authorship becomes legally ambiguous. Current guidance suggests that human contribution — editorial review, corrections, substantive changes — is what establishes human authorship. This creates a practical recommendation: document your editing process. Keep version histories. Record which changes were made by a human editor. This is both good editorial practice and legal protection.

Article 16 — Right to explanation. Where AI is used in decision-making processes (grant allocation, credibility scoring, content moderation), the individuals affected have a right to understand how the decision was made. This applies less to editorial work and more to any administrative or evaluative AI use within the organization.

A Practical Legal Scenario

Consider this situation: an NGO publishes a policy report that was drafted with ChatGPT and contains a factual error.

The organization is liable for the error — not OpenAI, not the model. The disclosure requirement under Article 21 means the organization may additionally face scrutiny for undisclosed AI use if they did not label the content. The copyright question becomes relevant if the report is later adapted or republished.

The practical response to all three risks is the same: treat AI output as a first draft, not a finished document. All factual claims require source verification. The human editor's review must be documented. Disclosure should be standard practice regardless of how much the output was edited.

The Hallucination Problem

No article on AI in fact-checking is complete without an honest discussion of hallucinations, because it is specifically dangerous in this context.

Language models generate text that is statistically coherent, not text that is factually verified. When a model does not have reliable information about something, it does not say "I don't know." It generates a plausible-sounding answer. In a general-purpose context this is inconvenient. In a fact-checking context it is a direct threat to organizational credibility.

Specific failure modes to watch for:

Fabricated citations. A model asked to provide supporting sources for a claim will sometimes generate references that look real — correct author name format, plausible journal name, reasonable publication year — but do not exist. Always verify citations independently before including them in published work.

False statistical specificity. Models will often produce exact figures ("43% of respondents reported...") that have no actual source. The specificity makes the claim seem credible. It is not.

Invented quotations. If you ask a model to provide quotes from a specific person on a topic, it may generate quotations that the person never said. The output will be grammatically consistent with the person's known style, which makes it harder to detect.

The mitigation is not to avoid AI — it is to never treat AI output as a primary source. AI is useful for organizing, summarizing, and drafting. It is not a substitute for source verification. Every specific factual claim in AI-generated output should be verified against a primary source before publication. This is more work than simply using the AI output directly, but it is the correct workflow.

Practical Tips

Start narrow. Pick one task type — transcription, document summarization, or first-draft writing — and develop your prompting workflow for that before expanding to other use cases. Trying to automate everything at once produces inconsistent results and makes it hard to identify what is working.

Version your prompts. Save the prompts that consistently produce good output. A shared team prompt library for common task types (interview transcription, comparative analysis, report drafting) is a genuine productivity asset. It also ensures consistency when different team members are doing the same type of analysis.

Set context length intentionally. When analyzing a long document, be explicit about the scope of the analysis in your prompt. "Focus only on Section 3" produces better results than asking a model to analyze a 50-page document as a whole. Break large tasks into defined segments.

Cross-verify statistical claims. Any specific number, percentage, or statistical claim in AI output is a hallucination risk. Build a habit of flagging every figure for independent verification before it goes into a document that will be published.

Disclose by default. Do not make disclosure a case-by-case judgment call. If AI was used in the production of content, disclose it. This is both the legally safer position under developing Kazakhstani regulation and the editorially honest one.

Common Mistakes

Using AI output as the final product. The most common mistake is treating a well-formatted AI response as a finished article or report. AI output requires human editing, fact-checking, and editorial judgment. It is a draft, not a publication.

Vague prompts for precise tasks. Asking a model to "analyze" something without specifying what the analysis should produce, what format the output should take, and what constraints apply is almost guaranteed to produce a generic response that requires substantial rework.

Overreliance on AI for source verification. AI cannot verify facts. It can only generate text that is consistent with its training data. Asking a model whether a specific claim is true is not fact-checking — it is asking for a confident-sounding guess. Verification requires checking primary sources directly.

Skipping disclosure. Publishing AI-assisted content without disclosure is both a legal risk under emerging Kazakhstani legislation and a credibility risk if the process is later scrutinized. The short-term convenience of not disclosing is not worth the institutional risk.

Choosing the wrong tool for the task. Transcription, document analysis, and image generation are different task types that are handled by different tools with different reliability profiles. Using a text generation model for tasks it was not optimized for produces poor results. Match the tool to the task.

Key Takeaways

The value of AI in NGO work is real but conditional. The condition is discipline: in how you write prompts, in how you review outputs, and in how you manage the boundary between AI assistance and human editorial responsibility.

The specific lessons worth carrying:

Prompt quality determines output quality more than model choice. Five minutes spent writing a precise prompt consistently outperforms twenty minutes editing a vague one's output.
AI is most valuable for high-volume mechanical tasks: transcription, initial document comparison, first-draft production. It is least reliable for anything requiring factual certainty, source verification, or nuanced editorial judgment.
Kazakhstan's AI legislation is actively developing. Disclosure requirements, authorship documentation, and institutional accountability for AI-generated errors are all real legal considerations now, not future ones.
Hallucinations are not a bug that will be fixed in the next model release. They are a structural characteristic of how language models work. Workflow design should treat AI output as unverified by default.
The goal is not to replace analysts with AI. It is to make analysts faster by handling the mechanical parts of the workflow — so that the human judgment calls that constitute actual fact-checking get more time and better-organized information.

Conclusion

Information integrity work has always been resource-constrained against a well-resourced opposition. AI shifts some of that asymmetry in favor of verification, but only if the tooling is used with the same rigor the work demands. A hallucinated citation in a fact-check is not just an error — it is a credibility crisis and potentially a legal problem under Kazakhstan's emerging AI regulation.

The organizations that will use AI most effectively are the ones that treat it as infrastructure: something that handles defined, repeatable tasks within a workflow that has clear human oversight at the points where accuracy matters. That means investing in prompt discipline, building team-shared workflows, maintaining documentation of AI use, and treating every piece of AI output as a starting point rather than a conclusion.

The tools exist now to significantly accelerate the work. The question is whether the workflows exist to use them safely.

DEV Community: ToRnaDo_O