Kinan: an Arabic tipline where the machine answers "have we seen this before?" and a person answers "is it true?"

What happened

Fatabyyano, the Amman-based fact-checking organisation, built a conversational bot called Kinan (كِنَان) and attached it to the accounts it already ran on Facebook Messenger, Instagram and X. It runs on Check, Meedan’s open-source verification platform. Someone forwards a suspicious video, image or claim to the bot. Check matches it against the debunks Fatabyyano has already published, using language processing and similarity matching, and returns the existing answer immediately. If there is no match, the request goes to a human fact-checker, who investigates and publishes.

Fatabyyano’s own description of the design:

The intelligent bot named “Kinan” will search for accurate information based on users’ inquiries. If the required information is not readily available, the bot will escalate the request to our fact-checking team for further investigation.

On 23 February 2023 Fatabyyano turned Kinan into a dedicated emergency tipline for the Turkey–Syria earthquake, giving Arabic speakers in the affected areas a route to checked information within weeks of the disaster. It was the first Check tipline built for an Arabic-speaking audience.

The design decision worth copying

The machine answers one question, have we already answered this? A person answers the other, is it true? Automation on retrieval, humans on judgement. That line holds whatever software you use.

Most misinformation during a crisis is the same handful of rumours arriving thousands of times. Answering the ten-thousandth copy is a lookup. Answering a new one is a judgement and cannot be delegated. Put the line in the wrong place and you either drown your team or let a machine tell the public that something is true.

Doing this in a five-person office

1. Build the library first. Two years of your own published debunks, explainers or rumour responses is the asset. The bot only makes it findable.
2. Put one inbox where your audience already is. Fatabyyano did not build an app. They attached a bot to the accounts their audience already messaged. In Baghdad that means Arabic, Kurdish Sorani and English on the channels people actually use.
3. Let the machine handle the repeat question, by matching it against what you have already answered.
4. Route every new question to a person. Say this out loud in your team, because the drift always runs the same way: a matched-answer bot can confidently return a near-miss debunk that does not address what the person asked.
5. Feed every human answer back into the library, so today’s judgement becomes tomorrow’s instant reply. That step is our reading of the design, not something either source states.

You do not need Kinan’s software. A searchable spreadsheet of your published debunks, and one person on rota for new questions, is the same architecture at a scale a small team can staff. Start there and see whether the volume ever justifies more.

What it costs you before it helps

Two warnings neither source raises. An Arabic-language tipline generates far more inbound than a small team can triage, and an escalation queue silently becomes a backlog nobody reads. And a tipline collects sensitive material: photographs, locations, accounts of violence, sent by people in conflict zones who did not think about where it would be stored. That is a data-protection responsibility and it arrives on day one, before any benefit. Decide where those messages live, who can read them, and when they are deleted, before you publish the address.