Taina
Constitutional AI co-designed with communities
Write the first three rules of an AI assistant’s constitution — and notice how different they read from the kind of rules a vendor would write.
Reading
Taina is GainForest’s first “regenerative intelligence” assistant — an open-source language model (built on Meta’s Llama family) that lives inside Telegram, runs on community-owned hardware, and speaks English, Portuguese, Spanish, Bahasa, and Swahili. She was co-designed with four Indigenous and local communities around Manaus, who together formed an Indigenous Data Council that governs her ongoing development.
The technical specifications are interesting but secondary. The piece worth studying is how the constitution — the system prompt that defines what Taina may and may not do — was written. It was not drafted by GainForest engineers and presented to communities for approval. It was drafted by the Council in co-design workshops, in the Council’s own languages, and then translated to a system prompt the model could execute. Lesson 04’s six pillars (language use, cultural protocols, sensitive knowledge, data access, correction pathways, withdrawal) gave the Council the canvas. The Council filled it in.
The result is a small concrete example of what sovereign AI looks like in practice: the same model architecture you might use for any chatbot, configured by a constitution that names a specific community, run on infrastructure that community can audit, with a withdrawal pathway the Council can invoke. If GainForest disappeared tomorrow, the constitution and the model weights would stay with the Council. That portability is, again, the test.
Taina is not finished. She is a working example that the framework from Lessons 01–05 can be operationalised today, with current open-source models, on current infrastructure. The point of this lesson is not to admire her — it is to draft yours.
“Communities have actively participated in crafting the system prompt for regenerative intelligence, enabling our initial system, Taina, to develop a personality that reflects their values and earns their trust.”
Handouts for this lesson
Practise
Exercise
Draft the first three constitutional rules
- 01Pick one community context you know well — your team, a school, a partner organisation, a place you have stewardship over. Name it on the page.
- 02Draft three numbered rules an AI assistant working for that community must follow. Each rule should be one sentence the model could read and act on.
- 03For each rule, write one line: 'If a user asks the agent to break this rule, the agent does X.' Be specific — vague answers create operational holes.
- 04Read the three rules aloud. Notice which ones name the community by name (the soil framing) and which ones could apply to any community (the oil framing). Edit until the rules name the community at least once.
- 05If you can, share the draft with one person from that community. Ask them to redline it. The first draft is never the final one.
Knowledge check
Taina is built on an open-source language model that runs on community-owned machines. Why does the 'runs locally' part matter?
What is the role of the Indigenous Data Council in Taina’s design?
A vendor offers your community a multilingual AI assistant with the same interface as Taina, but the system prompt is closed and unmodifiable. By Lesson 02’s definition, is this assistant sovereign for your community?