About and method
Election Watch helps citizens, journalists and observers see whether official election websites are unchanged, whether fake copies exist, and whether a rumour matches what the official sources say. It is built for the #AI4Elections Hackathon by Watcher-X, an AGXII project.
What it does
- Watches official pages. Every 15 minutes it fetches each listed public page once and compares it with the last good copy. It flags defacement wording, a big drop in page size, new outside scripts, forms that send data to another site, redirect changes, and changes to the security certificate or server address.
- Finds look-alike websites. It generates likely misspellings and swaps of official domain names, checks which exist, and reads public certificate logs for new names. Each suspect is scored. It never enters data on those sites.
- Checks claims. Your text is matched only against the official pages we monitor. An AI model may answer only from those passages, must quote them word for word, and any quote that is not found in the official text is rejected. If nothing matches, the answer is "Cannot verify". Your text is not stored.
- Grades cyber posture. Every few hours it reads each site's security headers, encryption settings and public DNS records (email spoofing protection) and gives a plain-language grade with fixes. It also makes one plain http request, not followed, to see if it redirects to https. A low grade means missing protections, not that a site was hacked. A sharp drop is logged as an alert.
- Keeps tamper-evident evidence. Each alert contains the fingerprint (SHA-256) of the one before it. The button on the home page re-computes the whole chain in your own browser.
What it will not do
- It does not test, scan or probe election systems. It makes one ordinary page request per source, respects robots.txt, and uses a clear identifying user agent.
- It does not publish or predict election results, and it is not an official source. INEC is not affiliated with this project.
- It does not collect personal data. Visitors are rate-limited with a salted, hashed address that is not kept in logs.
Responsible AI
AI is used in one place: the claim check. It cannot browse, it is told to treat the claim as untrusted text, its output must be valid structured data, and every verdict needs a verified quote. Hausa, Yoruba, Igbo and Pidgin summaries are machine translations and are labelled as such; the interface text in those languages is awaiting review by native speakers.
Limits
Changes are not always attacks: sites are redesigned and updated. Alerts say what changed, not why. A quiet board means the pages we watch look unchanged, not that every system is safe.