Aos Fatos using GenAI to surface verified information audiences need
Brazilian fact-checking powerhouse is making finding facts a breeze through FátimaGPT, an AI chatbot that cuts through clutter and delivers clear, concise answers to your questions – all for free.
Written by Becky Horsbrugh
Aos Fatos, a Brazilian investigative news outlet (‘To The Facts’, in Portuguese), faces a challenge that is common to many fact-checking organisations. How can they effectively reach audiences seeking verified information on diverse topics and improve the distribution of their high-quality, fact-checked content?
Some years ago, they had an idea: to create an automated way for users to interact with the news organisation’s content and provide reliable information. That was how the initial version of Fatima was born back in 2019. Through WhatsApp and Twitter, the first iterations of the Fatima engine would answer questions in structured ways, providing links and prepared messages.
However, the emergence and initial availability of Large Language Models technologies could empower new ways to relevant conversations in easier and more effective ways.
The solution: Create their own GPT, incorporating LLM (Large Language Model) technology, which can interpret questions from users, and provide more relevant answers in natural language.
Bruno Fávero, the director of innovation at Aos Fatos, explains: “What we wanted to do is to find new ways for readers to have access to our content and to get our content out to as many people as possible.”
His organisation is just one of a number who have gone further than just testing ChatGPT since its launch in November 2022, by integrating OpenAI’s API (Application Programming Interface) to create their own GPTs.
FátimaGPT is a conversational Q&A chatbot which can be accessed on WhatsApp, Telegram and Twitter.
It’s an advanced version of Fátima, a fact-checking bot they launched in 2019, partly to amplify the news outlet’s work, according to Fávero, and to allow users to interact with their content - and it is free.
Fátima is a play on ‘FactMa,' an abbreviation for ‘fact machine.’
The testing and evaluation phase of the results took around six months.
A team of seven professionals, including developers, designers and a select audience, were involved in its creation.
Functionality and Development
The original Fátima bot had over 75,000 users across platforms, according to Fávero.
But it was limited to keyword searches within Aos Fatos’s archives and returned URLs in replies.
FátimaGPT combines a language learning model with the news outlet’s archives to give users clear answers to their questions, with source lists from the news outlet’s database.
When it came to designing their new model, Aos Fatos were aware of the possible drawbacks.
Fávero says they were excited by the capabilities of ChatGPT, but “...there is a trap of just using those models just because you want to be in the hype.”
So they were very careful with their testing, and thinking how FátimaGPT could actually be useful for their users.
Performance and Accuracy
In the first few weeks of its beta phase, FátimaGPT delivered 94% adequate answers and 6% insufficient responses (meaning the answer was in the database but was not provided.) No factual errors were reported, according to Fávero.
It employs a retrieval-augmented generation technique, which links the language learning model to Aos Fatos’s reliable database.
User Interaction and Transparency
FátimaGPT responds to user queries with detailed answers to their questions.
It also provides publication dates to ensure users understand the timeliness of the data.
As all answers are derived from the Aos Fatos database, the team knows that any information provided is from a specific, reliable source.
This guards against ‘hallucinations’ where ChatGPT can present false information as true.
There are limitations with FátimaGPT. It provides answers from a database that might not always be up-to-date with stories that Aos Fatos covers infrequently.
The news outlet has programmed the chatbot to provide a date range for the published information it has shared to get around this issue.
Readers can press a report error button if they feel the information provided is wrong or the chatbot is not working properly.
How to use FátimaGPT
Visit fatimagpt.aosfatos.org and ask a question
Telegram: Send ‘I want to test FátimaGPT’ to Fátima here
How FátimaGPT has benefited Aos Fatos
Fávero believes a tool like FátimaGPT can help with the fight against misinformation, as it gives users access to the facts in a world where people are bombarded with information from numerous sources.
“It's part of our job to make the content as accessible as possible for people,” adding the tool helps them do exactly that. Though always being aware, AI is “a useful technology that we have to deploy carefully.”
An advantage for Aos Fatos is that the company has strategically decided to invest in technology. Their team is young, as well as the audience, much of their work is focused on.
It has a core team working on the technical side of their work. Out of a newsroom of around 25, they have three developers, one designer and two specialists in linguistics.
Fávero’s background is in journalism, which he studied at the University of São Paulo. He then graduated from University College London with a degree in Political Science and worked as a politics reporter for a leading Brazilian newspaper before joining Aos Fatos.
AI tools and innovations
Aos Fatos has developed other AI tools like Escriba, an audio transcription tool that automatically transcribes large audio and video files in Portuguese, English, and Spanish (though the interface is in Portuguese).
Fávero says they’ve seen success with this tool within and outside the newsroom as it is easy to use, well designed and “when it’s obviously useful, people are very quick to adopt it.” This tool has also been relevant in their business model.
And Radar, a real-time disinformation monitor tailored to Portuguese-language media.
Lessons gained in the process
Aos Fatos is keen to explore how journalists can use technology responsibly and ethically, to improve readers' access to reliable information. However, their approach to AI incorporates teaching the public about its potential risks.
Fávero says he’s concerned that he’s seen many newsrooms trying to use AI “just because everybody's talking about it.” Any use must be intentional, he believes. And newsrooms and outlets need to learn from past experience, and not become hostage to the big platforms. Ten or so years ago newsrooms started building their audiences on Facebook. But when they decided to change their algorithm, traffic decreased by half for many.
On a positive note, though, he feels that the AI world is much more “open” in how knowledge is shared, as seen with the Open Source Models. The feedback they have received from users so far is that they are very excited about the AI tools provided by Aos Fatos.
What we’ve learnt
Have you done a thorough assessment of the current challenges and opportunities of what you already offer your audiences? A great way to start is incorporating incremental optimisations to the workflows already in place.
Prioritise your needs, and analyse thoroughly whether new technologies could be an answer to them.