Four AI Tools for Analyzing Public Documents and Data Under Censorship

Four AI Tools for Analyzing Public Documents and Data Under Censorship

Published on 31 Aug 2026

In authoritarian regimes, opacity is the weapon of choice for stifling investigative journalism. Today, artificial intelligence is turning that equation on its head.

In newsrooms with limited resources, tight deadlines, and data sources that are increasingly difficult to access, sifting through the appendices of a recently passed law or identifying a pattern of corruption in a management report could take weeks of manual labor, which is prone to human error. This is precisely one of the problems that most severely affects independent journalism operating in authoritarian regimes, where media shutdowns, the blocking of news websites, financial strangulation, and the criminalization of journalism are commonplace.

In this context, Large Language Models (LLMs) have become high-speed research assistants for journalists in any country where public information is scarce, fragmented, or outright concealed. They do not replace journalistic judgment or instinct, but they do drastically reduce the technical time it takes to read, organize, and cross-reference data.

That’s why we’re presenting four key tools you can use to conduct research with AI:

For journalists working with confidential sources or sensitive documents, this is a great option. Google NotebookLM functions as a “closed system” that responds only based on the documents you upload, which almost completely eliminates the risk of the AI fabricating data.

It allows you to upload up to 50 different sources (PDFs, Google Docs, spreadsheets, web links, etc.) and ask it, for example, to compare a report with a public budget, generating quotes that link directly to the exact passage in the original document.

Designed for data breaches and large volumes of scanned files, audio recordings, or emails. Its optical character recognition engine (OCR) can read blurry or handwritten documents.

What’s interesting about Google Pinpoint is that it automatically extracts names, companies, locations, and organizations to show how they’re connected to one another within the documents.

When the volume or complexity of the material demands it, the paid version of this tool can make all the difference. Claude, from Anthropic, stands out for its linguistic and logical reasoning capabilities, making it especially useful for processing laws, constitutional amendments, and complex resolutions. Its large context window allows you to work with hundreds of pages at once, and the workspaces feature lets you keep laws and reports accessible for months.

When we talk about ChatGPT, we’re referring to its paid version. Thanks to its advanced data analysis capabilities, it doesn’t “guess” figures but rather runs code in the background to process them. It’s the most reliable option when dealing with poorly tabulated public budgets, as it can extract tables, verify whether official totals add up, and cross-reference the information with data from previous years.

Having access to the right tool is only half the battle; the other half lies in how you interact with the AI. To avoid inaccurate or made-up answers, it’s crucial to give instructions with a zero-tolerance policy for deception and to require that the AI base its responses solely on the uploaded document, cite the exact page or section for each piece of data, and explicitly state when information isn’t available—rather than filling in the gaps with assumptions.

That’s why we’re providing a prompt that can work in this case. Copy and paste this instruction, substituting what applies to your situation based on the result you want to achieve:

Act as a data analyst and forensic auditor specializing in investigative journalism and public policy in restrictive environments. Your sole objective is to analyze the document(s) I have just attached with pinpoint accuracy.

For your analysis, you must strictly follow these quality control rules:

  1. Base your answers SOLELY on the explicit data in the text. If the document does not contain the requested information or if you cannot deduce it with absolute certainty, respond verbatim: “The information is not found in the provided document.” Never use external knowledge.
  2. Whenever you extract a piece of data, a figure, a name, or a conclusion, you must include the exact source in square brackets at the end of the sentence (Example: [File name, p. X] or [Section X, Paragraph Y]).
  3. Present the summarized information in clear Markdown tables or numbered lists, avoiding unnecessary digressions or introductions.

It is important to note that not all of these AI tools are available without restrictions in every country. In countries where access to certain digital services is blocked or restricted, it is common for some of these tools to be inaccessible or to display errors; in such cases, using a VPN is usually the quickest solution. Furthermore, artificial intelligence tools serve as a support to streamline journalistic work but do not replace judgment, rigor, or the final product.

Go to all articles

  • Other related contents

  • 01

    The Nicaraguan Regime and the use of Tweeting Troops

    Nicaragua 2022 Annual Report

    Read more
  • 02

    Citizen Protest vs. State Online Fiction

    How many tweets does it take to manipulate a country?

    Read more
  • 03

    #ProDatos: How to avoid scams on WhatsApp?

    Read more
  • 04

    API access: the Twitter change ignored by most

    Is it possible to further analyze Twitter?

    Read more
  • 05

    Statement on changes to Twitter API access

    ProBox's statement.

    Read more
  • 06

    5 things that have changed on Twitter since the arrival of Elon Musk

    What's next?

    Read more
  • Subscribe

    Subscribe to our weekly newsletter

    Subscribe