The first rule of cohesion monitoring is that the system must not become a tool against the people it protects. A record that adds up complaints or incidents against a community can be used to frighten or target that community. So good monitoring tracks acts, places and public actors, and it never tracks ethnic groups. The Do No Harm idea applies to data just as it does to aid.
Key points:
- Track what happened and where, and never which community did it.
- Don't infer anyone's ethnicity.
- Keep slurs out of shared lists.
- Avoid a single score that hides what's behind it.
Why can monitoring do harm?
Data outlives its purpose. A list made to protect people can be leaked, stolen or handed to a new office. In 2007-08, attackers went after people because of who they were. Human Rights Watch and later inquiries show how fast that can happen.
A table showing that one community has the most hate speech reports would be quoted in every campaign. It would prove nothing, and it could get people killed.
What should be tracked instead?
Track the act: a threat, leaflet, attack or eviction. Track the place, at the level the evidence supports. Track public actors, such as public figures, media outlets and online accounts. Track institutions and responses.
In TI's cohesion pack, the taxonomy names actors and stakeholders. It doesn't allow an ethnic group to be created as a tracked entity. That's a hard line in the design.
What about ethnicity in complaints?
People sometimes need to say who they are. A person reporting ethnic discrimination may have to name their community to explain what happened. That's their choice. It sits in their own words, in their own account.
Don't turn it into a field that can be counted. Ethnicity should enter only where a complainant states it, and it shouldn't be used to sort records.
Why not infer ethnicity from names?
It's unreliable, and it's unlawful in spirit. Kenyan law treats ethnic social origin as sensitive personal data. A guess based on a surname is still a claim about a person's ethnicity, and people can be harmed by a wrong guess as well as a right one.
Don't guess from names, languages or home areas. Don't let a model do it either.
Should a system use a list of slurs?
A shared list of ethnic slurs is dangerous. It teaches the words, and it can be read as a guide to who's insulting whom. It also dates quickly, as coded language changes.
TI's keyword sets are structural. They cover election violence, displacement, peacebuilding, media reports of hate speech, goonism, ethnic zoning and digital manipulation. The list of terms that NCIC flags is the commission's own, and it should stay with the commission.
Why no composite score?
A single index looks simple, and it hides a lot. It mixes things that shouldn't be mixed, and its weights are choices that someone made. A commission publishes findings that get attacked. Counts and shares can be defended, while an undeclared composite can't.
Report each measure separately, with its size and date.
Who should be able to see what?
Limit access by role. A field officer may need a ward's reports. A national analyst may need patterns. Few people should see names. Keep logs of who viewed what.
Share totals outside the team, and only where numbers are large enough that no one can be identified.
How should people be protected when they report?
Let people report without giving their name, where the channel allows it. Keep identity apart from the content. Ask for consent, and say what will happen next. Never publish anything that could reveal a witness.
Land regulations and the ICC experience show how central witness protection is.
What about public figures?
Public speech by public figures can be tracked. Their words are on the record, and they answer to the public. Even so, keep a report unverified until a person has checked it, and record the source.
What about geography?
Be careful. A ward-level map of tension can stigmatise, or help someone to plan an attack. Hold detail back from public reports. See election hotspots.
What should you do with old data?
Set a time to delete or summarise it. Keep what the law requires, and little more. A complaint file from one election shouldn't sit open during the next.
How do you check your own system?
Ask hard questions. Who could misuse this? What happens if it leaks? Could a new government or a rival party turn it around? Would a community feel safe if it saw how it's described here?
Ask someone outside your team to review it. Note your rules in writing, and report any breaches.
What about working with security agencies?
Be clear from the start. If residents think that a monitoring system feeds the security services, they'll stop speaking. If a real threat to life arises, there must be a route to act. Agree that route, and tell people about it.
What should a county or NGO adopt today?
Write a one-page data rule. It should say that you don't track communities, you don't infer identity, you limit access, and you delete on schedule. Train every staff member and volunteer on it.
What does this cost?
Some useful analysis is lost, and that's a choice worth making. You won't be able to say which community reports most. You will be able to say what's happening, where, and whether it's getting better. For a body that needs to be trusted by everyone, that's the better trade.
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
- Ethnic data and the law
- Verifying conflict reports
- Do no harm
- 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) enforces this in the product: no ethnic group can be created as a tracked entity, and the cohesion pack carries no composite score.
To see it on your own work, book a walk-through.