You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$14,036 left to reach the funding goal.
AI is already inside most organizations, adopted one account at a time, long before any decision about it. Where governance exists, it is often a document nobody reads. Nobody can say with confidence what data goes into which tool, what comes back, or who answers when the output is wrong. I will measure this in three impact-driven nonprofits in Brazil, with the same baseline and end-of-cycle measures, around one light standardized intervention: a one-page rule set and a two-hour working session. The results go out as an anonymized open dataset, with the protocol and a short guide in Portuguese on everyday risks.
Theory of Impact
The capacity to govern AI well starts in ordinary organizations, and I found very little measurement of what that capacity looks like in practice in Brazil. This study is empirical groundwork: the kind of field-building evidence that AI governance work needs to stand on. It does not claim to be catastrophic-risk research. Open, comparable data on how written rules relate to everyday use gives funders, researchers and the organizations themselves a base that does not exist today.
Baseline and end-of-cycle measurement in three organizations, with the same three measures as a smaller related pilot (one organization) that I am submitting separately.
A qualitative reading of the gap between written rules and real use.
An anonymized open dataset (four organizations if that pilot is funded, otherwise three) and the measurement protocol.
A short guide in Portuguese.
Four months. The organizations are being invited now; I have no formal partnership yet and have not chosen them.
The protocol runs end to end in the three organizations. Each one gets a baseline and an end-of-cycle dataset on the same three measures: how many AI tools are in use and with what data; how long a decision about AI takes; how many people can name who answers for a given output. The fourth result is the qualitative reading. It records, organization by organization, where the written rule and what people report doing diverge, and why. I will know it worked when the anonymized dataset, the measurement protocol and the guide in Portuguese are public under an open license, and the protocol is documented well enough for another practitioner to run it without me. Two limits are stated up front. The measures are self-reported, and the working session itself teaches who answers for an output, so part of any change in that measure comes from the session itself. The study describes change. With no control group, it does not prove what caused it.
~196 h of direct work at US$ 60/h (US$ 11.760): measurement and qualitative reading ~90 h, light intervention ~36 h, comparison and open dataset ~30 h, guide in Portuguese ~25 h, publication ~15 h; plus US$ 1.000 in direct costs (transcription and review). The guide in Portuguese is written and delivered here and covers the everyday risks of AI use; the Emergent Ventures pilot publishes a separate open guide in English and Portuguese, on governance and a first maturity model. They are distinct deliverables and neither covers the other. My time beyond the funded hours is my own contribution.
The minimum covers this full scope. Beyond the goal, extra money would buy more organizations and a follow-up measurement.
Solo. Journalist, data and AI product manager and consultant. Almost seven years at Greenpeace: for almost three I built Greenpeace Brazil's data and digital transformation area from scratch; I sat on Greenpeace International's Technology Council, took part in its decisions and helped write its AI governance guidelines and train people on them; I was Greenpeace Brazil's DPO for a year and a half. I have worked as Maré Consultoria since May 2026 (registered as a company in Brazil in August 2026). Published work in digital journalism and fact-checking (five academic articles). I have not published on AI safety; what I know comes from practice. Applied to BlueDot's Frontier AI Governance course (cohort of 26–31 October 2026).
The first risk is recruitment. The organizations are being invited now and none has confirmed. Three organizations (four if the pilot is funded) is a small sample and not random, and part of the comparative dataset depends on a pilot not yet funded. Organizations may not agree to share data. If that happens, I publish the protocol and the guide anyway, and I report what could not be measured.
None so far. A smaller related pilot at Emergent Ventures, once submitted, will cover a different scope: one organization, an open guide and a first maturity model, with no shared hours. Its guide is a separate deliverable from the Portuguese guide here, and neither budget pays for the other. A first, broader Emergent Ventures request was declined in October 2026.
If another grant ever covers any of these same hours, I will not take payment twice for them. I will report each grant against its own deliverables.