On Saturday, the tech oligarch Dario Amodei (Anthropic), then supported by Sam
Altman (OpenAI) and Elon Musk (SpaceX), invited everyone -despite them owning the
few companies that produce AI frontier models-, called on frontier AI companies and
governments to slow the pace at which they improve AI. Why? Because there are
serious risks humanity could face if we don’t stop now.
But in some ways, that is already happening. Not about killing us, but AI companies,
working with governments, have created tools to control people, and we can see this
from examples in Latin America.
The ultimate aim of technology is to serve a common purpose. In this sense, the
concept of well-being is useful for framing and evaluating the intentions of AI initiatives.
Using the idea of well-being from the capability approach (Alkire, 2005, 2015; Robeyns,
2017; Sen, 1984, 2004) and applying it to AI projects, I define a well-being-focused AI
project as one that aims to increase or protect people’s abilities by clearly improving—or
lowering risks and hardships—in one or more parts of well-being. The project must also
think about how these effects affect different people and groups, what they mean for
future generations, and how they connect with other parts of life.
Well-being has many parts: efficiency, productivity, economic growth, or saving money
alone are not enough. Following the capability approach, the focus is on what
technology allows people to be and do. Each project can be judged by five things: the
ability it affects; the well-being area like health, education, income, housing,
environment, safety, social ties, civic involvement, or how time is used; whether the
effect is positive or negative; who gets the benefits and who faces the risks; and its
long-term effects and side effects.
Using this idea of well-being, I studied a database of public AI projects in Latin America
and the Caribbean. I found that most projects publicly AI government initiatives, control-
related constitute a substantial share of the database and are strongly concentrated in
facial-recognition initiatives classified as potentially negative for well-being.
General findings: AI controls the population
- The geographical distribution is highly concentrated. Brazil accounts for 315/1,048 = 30.1% of the database, Colombia for 265/1,048 = 25.3%, and Mexico for 151/1,048 = 14.4%.
- Facial recognition is not a marginal application: it accounts for 293/1,048 = 28.0% of all recorded initiatives.
- The three most common technologies—facial recognition, chatbots, and anomaly detection—represent 621/1,048 = 59.3% of the database.
- AI deployment is not limited to central governments. Subnational governments account for slightly more than half of all initiatives.
- 34% of AI projects are linked to control, monitoring, and the government’s oversight role.
controlling the population.
Table 1: Descriptive Indicators of general findings
| Indicator | Number / Share | Description |
|---|---|---|
| Initiatives analysed | 1,048 | Individual country-level records |
| Countries covered | 26 | Uneven coverage across countries |
| Brazil, Colombia and Mexico | 731 — 69.8% | Nearly 7 out of 10 records come from three countries |
| Five leading countries | 871 — 83.1% | Brazil, Colombia, Mexico, Chile and Argentina |
| Facial recognition | 293 — 28.0% | Most frequently recorded AI technology |
| Chatbots | 190 — 18.1% | Second most frequent technology |
| Anomaly detection | 138 — 13.2% | Third most frequent technology |
| Three leading technologies | 621 — 59.3% | High technological concentration |
| Executive branch | 753 — 71.9% | Main institutional setting |
| Subnational level | 529 — 50.5% | Almost evenly distributed relative to the national level: 519 — 49.5% |
Findings related to well-being:
Key findings about well-being and AI:
- 45% of the projects can clearly be linked to well-being.
This includes projects using facial recognition and security because they aim
to protect people’s lives and keep them safe from crime and danger. Since the
capability approach starts with being alive, projects about security and
facial recognition are also counted as well-being projects. - 46% of AI projects had no connection to well-being impacts.
11% had a small connection, 22% had a moderate connection, 19% were
connected, and only 2% were strongly connected to well-being. - The database includes 177 facial recognition projects linked to well-being.
Of these, 169 (95.5%) are also connected to control. Among them, 153
(90.5%) were seen as possibly having negative effects. - 26% of the projects might harm well-being.
This is mostly because control-related projects could discriminate against
vulnerable groups and communities already treated unfairly. - El 16,4% de las iniciativas de IA registradas en la región combina
reconocimiento facial, relación con el bienestar y funciones de control.
En Brasil, este cruce representa el 37,5% del total nacional y concentra
118 de los 172 casos regionales. - 153 of the 170 projects linked to well-being and seen as having
negative effects also involve facial recognition and control.
This shows that harmful effects on well-being are not spread evenly across
AI projects. They mostly happen in surveillance and control technologies.
Table 2: Well-being indicators
| Well-being indicator | Ratio | Percentage |
|---|---|---|
| Initiatives related to wellbeing | 457/1,048 | 43.6% |
| Positive direction among wellbeing-related initiatives | 287/457 | 62.8% |
| Negative direction among wellbeing-related initiatives | 170/457 | 37.2% |
| Facial recognition initiatives related to wellbeing | 177/293 | 60.4% |
| Wellbeing + facial recognition + control | 169/1,048 | 16.1% of all initiatives |
| Control among facial recognition initiatives related to wellbeing | 169/177 | 95.5% |
| Negative direction within the triple intersection | 153/169 | 90.5% |
| Triple intersection among all negative wellbeing-related initiatives | 153/170 | 90.0% |
Table 2: Well-being indicators
Key country-level findings:
- Brazil is the clearest outlier. It accounts for 118/169 (69.8%) of all initiatives in the wellbeing–facial recognition–control intersection.
- Brazil also accounts for 109/153 (71.2%) of initiatives in this intersection with a negative direction.
- In Brazil, almost all negative wellbeing-related initiatives are linked to facial recognition and control: 109/110 (99.1%).
- Chile and Brazil have similar proportions of wellbeing-related initiatives—57.5% and 56.8%, respectively—but markedly different technological profiles. The triple intersection represents 37.5% of Brazil’s initiatives, compared with 9.2% of Chile’s.
- Colombia has the second-largest number of initiatives in the database, but the lowest proportion of wellbeing-related initiatives among the countries with larger samples: 70/265 = 26.4%.
- Mexico and Chile have nearly identical proportions of initiatives in the triple intersection—9.3% and 9.2%, respectively—although Mexico has more cases in absolute terms.
- Among countries with smaller samples, Panama records 3/13 (23.1%) and El Salvador 3/6 (50.0%) in the triple intersection. Interpret these percentages cautiously because of their small denominators.
| Brazil | 315 | 179 | 56.8% | 118 | 37.5% | 110/179 | 61.5% |
| Colombia | 265 | 70 | 26.4% | 9 | 3.4% | 14/70 | 20.0% |
| Mexico | 151 | 57 | 37.7% | 14 | 9.3% | 14/57 | 24.6% |
| Chile | 87 | 50 | 57.5% | 8 | 9.2% | 13/50 | 26.0% |
| Argentina | 53 | 26 | 49.1% | 3 | 5.7% | 4/26 | 15.4% |
| Peru | 46 | 23 | 50.0% | 4 | 8.7% | 4/23 | 17.4% |
| Uruguay | 35 | 13 | 37.1% | 1 | 2.9% | 1/13 | 7.7% |
What should governments ask before deploying AI?
Yesterday, UN human rights chief Volker Türk warned that AI is a greater existential
risk to every aspect of our lives. And yes, it is true. The autonomous AI drones that
killed three people in Ukraine should warn us about the power of this technology
to kill us.
But it is also true that the same technology has enormous power to improve
people’s lives, expand our possibilities as humans in interdependency with
Nature, and create better options to live better. So, it is not only about danger,
but it is also about the opportunities we are missing with AI.
We live in a time when democracy is rare. According to the latest V-Dem Institute report
(2026), only 26% of people live in democracies. We all need to protect democracies
because they are the only system so far where people can freely choose their own life
paths in a peaceful society, which is essential for a good life. But if people keep feeling
that governments focus more on controlling them than on helping them live well, how can
we expect to keep supporting democracies?
Government needs to be aware of these trade-off decisions, because so far, in regards to AI,
governments should identify the capability it is expected to expand, the groups most likely to
benefit, and the groups that will carry the risks. This also changes what counts as innovation.
A system should not be judged successful merely because it processes more information or
enables the State to see more. Public innovation should be evaluated by whether it gives people
greater freedom, security, and real opportunities to live the lives they value. That is the real
mechanism to ensure democracies.
Methodology
I turned the capability approach into four parts: whether a project involved watching or control;
whether it was linked to well-being; how strong that link was on a five-point scale; and the
possible effect direction from -2 to +2.
I tested the coding rules using cases from Chile. ChatGPT then used the same rules on the
other projects based on their names, organizations, sectors, keywords, and descriptions.
I later checked and confirmed the AI-assisted coding by hand, fixing any differences,
especially in surveillance and security cases.
After leaving out two OAS records, the analysis includes 1,048 projects from 26 countries.
The scores show possible links and effects based on available information; they do not prove
real impact. The database is not complete or equally balanced across countries.
