Hate speech monitoring means watching public speech, checking what's found, and passing evidence to those who can act. NCIC told MPs that social media monitoring and forensic investigative tools are essential, as offences shift to digital platforms. In July 2026, NCIC and the Inter-Religious Council spoke of strengthening real-time monitoring. The hard parts are language, context and proof, and software can't do those alone.
Key points:
- Monitor what's public, and keep to the law.
- Context decides what counts as hate speech.
- Every item stays unverified until a person checks it.
- Evidence must be preserved properly.
Why monitor at all?
Early warning is the first reason. NCIC monitoring has detected a resurgence of coded ethnic language ahead of 2027. Spotting that early allows a response before it spreads.
Accountability is the second. Reports say that weak evidence has hurt prosecutions.
What should be monitored?
Public rallies, radio, television, newspapers, public social media posts and public statements. Focus on people with reach: leaders, broadcasters, and accounts with large followings.
Dangerous speech explains what to look for: a credible speaker, a fearful audience, and a message that makes violence seem acceptable.
What about closed groups?
WhatsApp and other closed groups matter greatly, and they're private. The Inspector General named encrypted platforms as a challenge. Don't infiltrate private groups. Rely on people who choose to report what they've seen, and treat those reports with care.
What about radio?
Radio matters most in many counties. It's also the hardest to monitor. Vernacular stations broadcast in many languages, and speech recognition doesn't yet work well for most Kenyan languages. The Media Council's seven-second delay helps stations to stop harm before it airs, but it doesn't help you to monitor them.
For now, trained human listeners who know the language remain the reliable method.
What can keyword tools do?
They help you to find items for a person to read. They're good at volume, and poor at meaning. A word may be a slur in one place, and ordinary in another. Coded language changes quickly, and it's designed to escape filters.
Use keywords built around themes, such as displacement, ethnic zoning and goonism. Avoid publishing a list of slurs.
What can AI do?
It can sort, group and summarise. It can flag items for review. It shouldn't decide that something is hate speech. It'll miss context, and it may wrongly flag Kiswahili or Sheng. An error here could harm a real person.
A person should check everything before it goes further.
How do you check an item?
Note who said it, where, when and to whom. Find the original. Check for edits. Ask what it means locally. See if others have reported it. Then record a status: unverified, corroborated, confirmed, contested or refuted. Verifying conflict reports explains each one.
How should evidence be preserved?
Save the original link, a full capture, the date and time, and who captured it. Keep the source file. Don't edit it. Note the chain of custody. Courts need to trust that what they're seeing is what was said.
Keep the original language along with any translation, because the original is the evidence.
What is the legal test?
Monitoring finds candidates, and the law decides. Section 13 of the National Cohesion and Integration Act has its own elements, and Article 33 protects expression. The law on online speech has shifted, as the Cybercrimes Act page shows.
How should you avoid bias?
Monitor all sides, with the same rules. A system that catches one party's speech more than another's will soon lose trust. Publish your method. The Inspector General named the politicisation of enforcement as a real challenge.
Watch for language bias too. Don't only monitor the languages your staff speak.
How should findings be reported?
Report themes and trends. Name people only where the law allows, and when you've confirmed the facts. NCIC uses a Wall of Shame, as a statutory body. Others shouldn't copy that without legal advice.
Avoid repeating the words themselves.
What should trigger action?
Set a clear threshold. Direct calls for violence, named targets, a specific time or place, and a speaker with power all raise urgency. A threat to life goes to the police straight away. Lesser items can be held for weekly review.
Who should do the monitoring?
Trained monitors who know the language and the area, with supervisors who check their work. Protect them, as they may face threats or be worn down by what they read.
How much will it cost?
The main cost is people. Software reduces the reading load, but it doesn't remove it. Choose a small set of sources that you can really cover. Forty unread feeds are worth less than four read properly.
How do you know it's working?
Check how often an alert led to action, how fast, and whether evidence stood up. Review what you missed. Compare with complaints and with the barometer.
What does a daily routine look like?
A monitor reads the morning's items and marks what needs a closer look. A supervisor checks those by noon. Anything urgent goes to the named contact at once. At week's end, the team writes a short note on themes, places and changes.
Keep it simple enough to survive a busy month.
How do you monitor in Kiswahili and Sheng?
Use people who speak them daily. Coded terms change fast, and machines lag behind. Keep the original words with any translation. Ask local readers what a term means in that place, because a word in Kisumu may differ from the same word in Mombasa.
How do you avoid spreading what you find?
Share as little of the content as you can. Use summaries in reports. Keep the source file safe. Never post a screenshot of hateful content to warn people about it, since that can spread it further.
Read next
- Conflict early warning system for Kenya: the TI pack page for this work
- Cohesion monitoring on TI: all the pages in this section
- Dangerous speech
- Verifying conflict reports
- Online hate speech
- Takwimu Intelligence (TI) home page
- National cohesion in Kenya: the guide to this library
- The TI platform and how we check our numbers
Where TI fits
Takwimu Intelligence (TI) follows public news and online text, groups items about one event, and marks each as unverified until a person confirms it.
To see it on your own work, book a walk-through.