Superintelligence for Newsrooms: AI-Assisted Fact-Checking and Sourcing
A practical guide for newsrooms using AI to find sources and check data, with human verification, decision logs and clear limits for reporters.
Want to apply this in your company?
Book a call with Juan →Superintelligence often comes up in debates about the future of AI, but it does not describe the tools a newsroom uses today. In practice, what works is using generative AI systems to widen the search for sources, cross-check documents and flag data points that need verification, always under the responsibility of the journalist who signs the story. The AI finds and organizes leads. The editorial team decides what gets verified, published and corrected.
Key takeaways
- AI widens the search for sources and the checking of data, but it does not decide what is true.
- Every published claim needs a primary source identified by the reporter, even when the AI found it.
- A workflow with defined steps, a log of queries and clear stopping rules keeps checking from turning into automatic acceptance.
- Personal data and source confidentiality require written rules before anything is sent to external tools.
- Starting with one desk and one agreed metric lets you decide, based on evidence, whether the use should grow, change or stop.
What AI does well in fact-checking and what it does not
A newsroom deals with volume: press releases, official documents, meeting minutes, reports, social media posts and audio messages that arrive through WhatsApp. In this setting, language models are useful for summarizing long documents, extracting names, dates and amounts, comparing versions of the same text and suggesting questions the reporter has not asked yet. This screening work often consumes time that should go to field reporting.
What AI cannot safely do is guarantee that a piece of information is true. Models can produce references that do not exist, attribute a quote to someone who never said it and merge data from different sources as if it were one dataset. For this reason, every AI output should be treated as a working hypothesis. The right question is not whether the answer sounds plausible, but where the original source is and whether the reporter has opened it.
There is a useful distinction for leadership: AI works well at locating and organizing, and poorly at certifying. When the task is finding where a figure was published, the tool helps. When the task is deciding whether the figure is correct, the answer depends on comparing it with the primary document, and that comparison is a human job.
How to build a source-gathering workflow
A simple workflow is usually enough to start. The order matters, because it prevents the model from defining the scope of the investigation before the reporter does.
- Define the journalistic question before opening the tool, including what you need to confirm and what stays out of scope.
- List the primary sources you expect: public agencies, documents, people with direct access and official databases.
- Use AI to search for and summarize these sources, always asking for the link or original document behind each claim.
- Open each indicated source and confirm the exact passage. If the passage does not exist, record the failure and discard the claim.
- Record what was confirmed, what remains pending and who is responsible for each item.
- Move on to the text only when every central fact has a primary source or a clear attribution to an identified person.
This workflow has an explicit stopping point. If the primary source cannot be accessed, the information does not enter the story as fact. It may enter, where it makes sense, as an attributed statement or as an open item. This rule applies to every desk, including opinion content.
Illustrative scenario: a story about city hall contracts
Illustrative scenario: a reporter receives a tip that a city government signed amendments to urban cleaning contracts. She asks an AI tool to locate the cited contracts on the public transparency portal, extract values, validity periods and scope, and compare them with the dates of the city council sessions.
The tool finds three amendments and organizes the data in a table. While checking, the reporter notices that one of the amendments does not appear on the portal, and that the date of one of the values differs from the official gazette. She logs the discrepancy, requests the documents directly from the agency and publishes only after confirmation. The AI saved her screening time. The reporting remained hers, with the official gazette as the reference.
The central point of this example is that the error was caught because there was a human step checking against the primary document. Without it, the neat table would have reached the public with a false figure.
Editorial responsibility and decision records
Responsibility for the story stays with the newsroom, and the best way to support that is to record the verification path. A minimum log should state which tool was used, for which task, which primary sources were checked and who approved the final claim. This history protects the journalist if the story is challenged and makes corrections easier.
It is also worth defining which decisions cannot be delegated. Choosing whom to interview, deciding whether information is in the public interest, writing the headline, choosing the image and setting the angle of the piece are editorial decisions. AI can suggest alternatives, but the team signs the decision.
A good practice is to separate, within the working file itself, what is confirmed fact, what is an attributed statement and what is a hypothesis. This separation often reveals gaps that would go unnoticed in a continuous draft.
The tool can point to where to look, but whoever publishes the information is accountable for it.
Personal data, source confidentiality and LGPD
Before pasting any document into a tool, the newsroom needs to know what could leave its control. Documents containing third-party personal data, information under source confidentiality, material under embargo and investigation files require extra care. As a general rule, the safest policy is not to send such material to external services without a prior assessment and formal authorization from editorial leadership.
In Brazil, the LGPD applies to the processing of personal data, including when AI is used as a support tool. Before defining what can be sent, it is worth consulting the ANPD guidance and assessing where the data is processed. If the use involves international transfer, the topic deserves its own review, and this article on international data transfer when using AI covers the requirements in detail.
A clear internal policy helps the team decide quickly. It should list what may be sent, what must never be sent, which tools are approved and who answers questions. For a set of rules that already works across the whole company, this piece on company policy for ChatGPT use can serve as a starting point.
How to start with a small pilot
The safest way to adopt this practice is to start with one desk, one type of story and a controlled volume. An economics desk, for example, can test checking of official documents before any use in police or health coverage, where the risk of harm is higher.
Our working method follows four steps: map the current process and its baseline, limit data and actions, pilot a real workflow with logs and an agreed metric beforehand, and then hand the use over to the team and decide whether to expand, adjust or stop. In a newsroom, the metric can be simple, such as the number of claims corrected after publication or the time needed to check a standard story, as long as the team agrees on it before the pilot begins.
After the pilot, the decision has to be explicit. If errors dropped and the team trusts the workflow, the use can grow into other desks. If the log showed many verification failures, the workflow should be adjusted before any expansion. If the tool did not reduce the work or increased the review burden, stopping is also a valid result.
For leaders starting to discuss this topic, this superintelligence glossary without jargon helps align terms. It is also worth reading how the discussion about media and blogs changes when the subject is sources and trust, which covers the same credibility risks from another angle.
Frequently asked questions
Can AI replace the fact-checker or the reporter in verification? No. AI can gather sources, organize data and flag inconsistencies, but checking against the primary document and the decision to publish remain human tasks. The person who signs the story is accountable for its content, no matter how many tools were used along the way.
How do you know whether a source suggested by AI is reliable? Open the source and confirm the exact passage cited. If the link does not exist, the document does not contain the information or the date does not match, the reference should be discarded. Secondary sources can guide the search, but they do not replace official documentation or direct statements from someone with access to the facts.
Which data should never be sent to AI tools in a newsroom? Sensitive personal data of third parties, information under source confidentiality, ongoing investigation material and embargoed content should stay out, unless there is formal authorization and a prior assessment. The internal policy should list these categories objectively, so the team does not have to decide case by case under deadline pressure.
Where should you start without disrupting the daily routine? Choose one desk and one type of story with controlled risk, define the checking workflow and the log, agree on a simple metric before starting and set a date to review the results. At the end, decide whether the use should be expanded, adjusted or stopped, based on what was recorded.