A chatbot reads. We investigate.
Ask a general AI tool about a viral claim and it will give you a fluent, confident answer, drawn from what it remembers. That is fine for a recipe. It is dangerous for news, because a fluent answer can still be wrong, out of date, or invented.
We do not ask one big model to remember the truth. We make a team of specialists find it, prove it, and date it.
Every claim that reaches News Intelligence is handled by purpose-built models that each do one job, and do it far better than a general-purpose model can. Large language models still have a place on the team, but a smaller one: they write, they translate, they tidy. They do not get the final say on what is true.
Meet the specialists
These are models we trained ourselves for fact-checking in Nigeria. We keep the recipe to ourselves, but here is what they do for you.
Does this source back the claim?
Our stance model reads a claim and a piece of evidence and decides whether the evidence supports it, contradicts it, or is simply unrelated. We trained it on Nigerian claims and on real newsroom reporting, so it understands how Nigerian stories are written.
Which evidence matters most?
Searching returns a pile of pages. Our ranking model sorts them by how directly they answer your exact claim, so a passing mention never outweighs a primary source.
Meaning, not just keywords
Our search understands what a claim means, not only the words in it. A rumour retold in different words, or in Pidgin, still finds the same facts.
Is this worth checking?
Not every sentence is a factual claim. Our claim detector separates checkable statements from opinion and chatter, so effort goes where it counts.
Pidgin, Hausa, Yoruba and Igbo
Built on a language library of more than 180,000 sentences, so claims arrive the way Nigerians actually speak and are understood the same way.
A memory of what is already settled
When a claim has been checked before, we recognise it, even reworded, and reuse the verified answer instead of starting from scratch.
We speak your language, and we measured it
Most fact-checking tools were built in English and stretched to fit everyone else. We built ours the other way round. Our language library holds more than 180,000 sentences, including more than 130,000 in Nigerian Pidgin, with Hausa, Yoruba and Igbo alongside. We used it to train our specialist models and to test our language accuracy, not just to decorate a feature list.
The result is a checker that does not fall apart when a claim is typed the way people really type it.
Where the language models fit in
Large language models are brilliant writers and poor witnesses. So we use them for what they are best at, and keep them away from what they are worst at.
What the language models do
They turn our findings into a clear, plain-language explanation. They translate it into your language. They tidy the answer into a card you can share. We use more than one, so no single model is a single point of failure.
What they never do
They do not decide what counts as evidence. They do not get to invent a source, a date or a quote. Every statement in an answer has to point back to a source our specialists found and weighed.
Llama, our second reader
Meta's open Llama model stands by as an independent second reader on the answer, a safety net that checks the finished work before it goes out and takes over if another reviewer is unavailable. Two readers who do not share the same blind spots catch far more than one. That is why very little gets past us. The only claims we cannot settle are the ones that have not been published anywhere in the world yet, and for those we say so plainly. We would rather tell you "we cannot confirm this yet" than guess.
Sources first. Trusted fact-checkers on top.
Not every source is equal, so we do not treat them as equal. Fact-checkers who have signed the IFCN code of principles, the international standard for non-partisan, transparent fact-checking, sit at the very top of our trust ratings. When one of them has already examined a claim, we read their work first. Official bodies speaking on their own matters come next, then established newsrooms.
When the evidence looks old, we go and look ourselves
If the sources we find are out of date, News Intelligence does not give up and does not guess. It goes to the relevant official body's own website, then to its verified social accounts, and looks back over recent weeks for anything newer. Then it tells you what it found, with dates: what was reported earlier, what the latest word is, and how old that word is. When nothing newer exists, the answer says so: last reported on a given date, no newer update found.
What happens to every claim
- Claim understood in your language
- Evidence gathered and ranked
- Each source judged for or against
- Official channels checked for the latest
- A dated verdict, in plain words
It usually takes about a minute, and every step leaves a trail you can follow back to the source.
The difference, side by side
| A general AI chatbot | News Intelligence | |
|---|---|---|
| How it answers | Writes from memory | Gathers and weighs fresh evidence first |
| Who decides | One large model | Specialist models, with language models as the lighter layer |
| Local languages | English first, the rest approximated | Trained on Pidgin, Hausa, Yoruba and Igbo |
| Sources | Often none, sometimes invented | Rated, ranked and linked, IFCN fact-checkers on top |
| When it is unsure | Sounds sure anyway | Says what it found, with dates, or says it cannot confirm |
Seen something you're not sure about?Check it before you share it. It takes less than a minute.
Check a claim