Artificial Intelligence in Latin America

Items

• Introduction

  Abbi Dicker 16 May 2025 Hits: 3866

  Artificial intelligence (AI) can be described as the next big thing, the technology most likely to drive comprehensive and far-reaching change across Latin America and the Caribbean (LAC) during the coming decade. Along with other technologies and concepts, including the digital revolution, net carbon zero, and the energy transition, the AI concept was first developed in the wealthier countries of the ‘global north’ rippling out from there to impact the ‘global south’ – developing nations including those in LAC. As with previous innovations, much of the discussion revolves around the ability of different countries to capture the beneficial effects of AI while minimising the potential harms it may unleash. There is also justified concern over whether economies closer to the periphery can avoid getting left behind, or see their people being trapped on the wrong side of the digital divide.

  This report will start out by first looking at the way AI has developed in the global north, and in second place, at the way Latin American and Caribbean countries have begun their own journey towards regulation. It may be helpful to note that a simple, lineal form of dissemination and adoption is unlikely to capture the complexities of the process. For a start, the emergence of AI has coincided with a period of radically increased super-power rivalry between the United States and China. Amid a global trade war, LAC countries may find themselves forced to choose between rival versions of AI technologies. There will therefore be limits to their ability to diversify supply chains.

• The slow-burn rise in AI

  Abbi Dicker 16 May 2025 Hits: 2248

  According to The Economist, the term artificial intelligence was first used in 1956 at an academic gathering in Dartmouth College in the US when John McCarthy, a young researcher, wanted to kickstart a discussion about “how to make machines use language, [and] form abstractions and concepts” to enable them to “solve the kinds of problems now reserved for humans”. Despite significant debate, not much happened over the following 50-60 years. By the 1990s however scientists had begun developing software-based neural networks, like those in the brain. These networks can learn by being exposed to vast datasets. Connections between neurons known as weights are adjusted to produce appropriate outputs, a process known as training. The rise of computing power using graphic processing units (GPUs) initially designed for the gaming industry, boosted the efficiency of these neural networks. In 2012 the use of additional neuron layers prompted use of the term deep learning. Trained networks were at this stage being used to identify images with one-word tags, with over 80% accuracy.

  Cousins of AI

  

  Source: Kapur, Radhika. 2018. “Cousins of Artificial Intelligence.” Medium

  With increased training and more neural layers networks began to be used for revolutionary commercial products such as Amazon’s voice assistant Alexa, or text-based translation services. In 2017 a new type of networks known as transformers allowed a form of self-supervised learning using raw content from the internet. These became known as large language models (LLMs). Early versions were far from perfect, sometimes being prone to ‘hallucinations’, moments in which AI makes up facts, provides misleading information, or produces responses that sound plausible but are incorrect.

  In 2019 a start-up company, Open AI, launched GPT-2, a chat bot which responded to user questions, and summarised and translated text. In November 2022 OpenAI launched GPT-3.5, a chatbot which can claim to have enjoyed the fastest-ever take-up of any commercial product. Rivals such as DALL-E and Stable Diffusion also entered the field.

  One of the problems of these AI systems is that they can run out of raw data for training purposes. The general rule is that the more text used, the more accurate the outputs are. A research company, Epoch ai, calculates that by 2028 the stock of high-quality textual data on the internet may have been used up. To avoid hitting this data wall some companies are developing ‘reinforcement learning from human feedback’ (RLHF). Despite credible fears that AI might lead to significant job losses, in the short term at least this has increased demand for human data-labellers who can earn US$100 per hour tweaking data weights. Google has also experimented with synthetic data.

  Perhaps a more challenging supply chain problem concerns silicon chips. In 1958 a chip could hold a single transistor. By 1965 they held 50. In 2023 a single Apple iPhone used chips containing around 19bn transistors. According to Moore’s law (named after Gordon Moore, a founder of Fairchild Semiconductor) the density of transistors nearly doubles every year. This massive rise in processing power enabled the rise of smartphones. But the increasing power needed to operate large language models could even challenge Moore’s law. Demand for ever-increasing video capacity for computer games and other uses lies behind the spectacular growth of US chipmaker Nvidia, which pioneered the graphics processing unit (GPU) chip. GPUs are considered 10-100 times more cost-efficient and up to 1,000 times faster than traditional central processing units (CPUs).

  Put bluntly, maintaining market leadership in AI will require reliable silicon chip supply chains. It is estimated that US digital giants Alphabet, Amazon, Meta and Microsoft invested around US$200bn last year, most of which went into AI and chip supply. Surging demand has led to a proliferation of data centres. According to media reports Microsoft and Open AI are planning a US$100bn data centre/GPU cluster facility in Wisconsin. Data centres require major increases in electric energy; in March 2024 Amazon said it was buying a data centre next to a small 1GW nuclear plant.

  In the context of the current US-China trade war, global GPU supply chains are geopolitically sensitive. US-based Nvidia has seen its capital value grow by a factor of eight since 2022 to reach US$2.8tn. ASML, a Dutch company valued at US$350bn, produces the world’s most advanced lithography machines used to etch designs on chips. TSMC in Taiwan and Intel in the US have acquired these machines. With the US banning some chip exports to China, the Beijing government is investing heavily to build its own supply chains.

  In a scenario in which China were to invade Taiwan (considered unlikely but not impossible) TSMC’s supply of chips would be likely shut down. Perhaps because of this the company is investing US$65bn to build a new chip fabricator in Arizona, which is set to receive some US$6.6bn worth of government subsidies. Nvidia came under pressure this year with the Donald Trump administration threatening to ban sales of its H20 chip to Chinese customers. At the same time Chinese telecoms and technology company Huawei said it was launching 910C, a GPU chip designed specifically for AI applications.

  Writing for US-based and tech-oriented online publishing platform Medium, Seema Singh seeks to explain the surge in the popularity of AI by noting that in the past we had a very small amount of data on which we could make predictions. Today, there has been an explosion in data availability, coupled with “support for more advanced algorithms, high end computing power, and storage that can deal with that huge data size”. Combining these factors has allowed the emergence of innovative products such as the Tesla self-driving car or Apple’s Siri.

• Case studies: what is AI being used for in LAC?

  Abbi Dicker 16 May 2025 Hits: 2949

  Many companies are using AI in the region to further develop their businesses and capture efficiency improvements.

  Globant/Argentina. Globant is a highly successful software and digital services company, set up by a group of friends in a bar in Buenos Aires in 2003, and which from there eventually turned into a ‘unicorn’ – a start-up with revenues of over US$1bn – in 2015. Today its annual revenue is close to US$5bn. Last year revenues grew by 15%, a pace which, amid current global uncertainties, it expects to drop to 12% in 2025. It is currently transitioning from being a software provider to focusing on helping its 1,000 customers in over 80 counties incorporate AI into their business processes. Chief executive Martin Migoya says AI brings with it a revolutionary change, comparable to the introduction of the assembly line by Ford Motor company in the early 20^th century. The company began investing in AI ten years ago and says that AI-related projects contributed US$350m to its 2024 revenues, more than double the amount in 2023. “Right now, AI is a personal productivity tool, just as the internet in the beginning was a simple tool for sending messages” Migoya says, adding “but in the future AI will be at the core of business processes, with humans overseeing them. It is about getting AI agents into the process and getting the right supervision points for humans inside the process.”

  MercadoLibre/Argentina: Another start-up, MercadoLibre is an online retailer that was launched in Argentina, is now headquartered in Uruguay, and generates most of its revenues from Brazil. The company had a bumper year in 2024: revenue soared by 38% to US$21bn, confirming its position as the region’s most valuable company by market capitalisation. It now focuses on warehousing, the courier business, finance, logistics, and even running its own fleet of aircraft. The big area of growth has been its payments and fintech division known as Mercado Pago, which now accounts for 41% of total revenue. Analysts say it has integrated AI across its operations improving user experience, operational efficiency and financial services. AI is used to personalise product recommendations, with algorithms analysing consumer behaviour and transaction history. It is also used for real-time fraud detection and credit scoring. In 2023 AI technology use was linked to a 22% increase in units sold and a 47% increase in unique active users.

  Solinftec/Brazil: This is a company set up by a group of Cuban engineers who moved to Brazil to provide AI and software-as-a-service (SaaS) to the agribusiness sector. The company has focused on precision agriculture using an autonomous, solar-powered robot which operates around the clock and is known as Solix. Tasks conducted by Solix include real-time plant monitoring, local automated weed control, non-chemical video-based pest control, crop and environmental management based on soil quality and designed to achieve low-impact, high yield agriculture. The Solix system is currently deployed over 12m hectares in Brazil, including crops such as sugarcane, maize, soya, wheat, and cotton.

  Dapta/Brazil: Dapta is using AI-powered diagnostic tools to analyse medical images, allowing early detection of diseases like cancer and heart conditions, particularly in the more isolated areas where medical infrastructure is scant. The system uses machine learning algorithms to automate the diagnosis process: as a result, identifying those at risk of major disease is done at much greater speed. The company is also investigating the use of AI in tele-medicine where doctors can interact with patient in isolated and remote communities.

  Vozy/Colombia: Vozy is an AI voice platform that was set up in Medellín in 2017. It uses Lili, a virtual conversational assistant, that uses Natural Language Processing and machine learning to understand and speak in eight different Latin American Spanish accents, intended to create a more consumer-friendly interaction. The Vozy system adapts to the urgency and complexity of customer requests. Chatbots are used for simple, non-urgent requests; voice assistants for medium complexity request that require immediacy, while human agents are reserved for requests that require intellect and empathy. Explaining its approach, the company says that across the call-centre industry, customers often have to call two times to get problems resolved; the aim of Vozy is to get the issue solved with one call, improving customer satisfaction and reducing costs. It claims it can make customer service up to 40% more efficient.

  Rappi/Colombia: Rappi initiated as a courier company in Colombia but evolved into an on-demand delivery platform available in various Latin American countries. It has grown at breakneck speed, with a current market valuation of around US$5bn. It offers grocery delivery, restaurant food delivery, and courier services. It has recently expanded into financial services through RappiBank, offering digital financial solutions like credit lines and QR-code based virtual wallets. The company uses AI in various ways to streamline and optimise its operation. Algorithms are used to personalise recommendations of additional goods and services offered to existing customers; to choose the best delivery routes; and to use chatbots for customer care in real time.

  Diagnostikare/Mexico: This is a six-year-old primary health start-up, which began by providing a digital healthcare service via smartphone targeting around 500,00 people in poor areas of south-eastern Mexico. The platform includes health counselling, prescription fulfilment, lab test referrals, mental health support and digital health tools. AI is used on the Diagnostikare platform in various ways. First, for symptom analysis: algorithms are used to provide preliminary patient assessments. Second, to provide personalised recommendations – such as health advice, lifestyle recommendations, and treatment plans based on individuals’ medical data. In third place, AI is used for predictive analytics – to identify early-stage risks and facilitate timely medical intervention. Diagnostikare recently announced it is working with German pharmaceuticals company Bayer to improve access to digital healthcare, as well as with a leading Mexican insurance company to offer a health products add-on to almost 9m consumers.

  Nutri Co/Peru: Nutri is a small company producing plant-based nutritional products using superfoods. Typical products are shakes, bars, cookies, and cereals marketed as healthy and environmentally sustainable. The company recently received US$2m of funding to expand into Chile and Mexico. AI is an important component of Nutri’s business. The company uses AI to analyse vast quantities of data on ingredients, nutritional values, and health benefits. This allows the use of optimised recipes tailored to specific dietary needs and preferences.

• Is Latin America ready?

  Abbi Dicker 16 May 2025 Hits: 2670

  In 2023 and 2024 a research team formed by Chile’s National Centre for Artificial Intelligence (CENIA) and the UN’s Economic Commission for Latin America and the Caribbean (Eclac) began publishing an annual Latin America Artificial Intelligence Index (known by its Spanish acronym ILIA) covering 19 of the region’s countries. The ILIA is intended to measure readiness to adopt AI. Countries are given a score on a range from zero to 100, with the higher scores designating higher levels of preparedness.

  In the 2024 ILIA Chile obtained the top spot with 73.07 points, followed by Brazil (69.30) and Uruguay (64.98). These three countries are described as AI pioneers in the region. A second group of countries were described as ‘adopters’ and includes Argentina (55.77), Colombia (52.64), and Mexico (51.40). The report says that on average over the last eight years LAC has doubled its AI talent concentration within the workforce but still lags when compared to the global north.

  The pioneer group stands out because of its better performance in three critical areas, namely enabling factors; research, development and adoption; and governance. CENIA-Eclac says the pioneers “have not only made progress on implementing AI-based technologies but are also orienting their national strategies towards the consolidation and expansion of these technologies in all sectors of their economy and society”. Chile, Brazil, and Uruguay are also praised for creating a “conducive environment” fostering “research, development, and adoption of relevant technologies.”

  Latin America AI Readiness index

  

  Source: ILIA (Indice Latinoamericano de Inteligencia Artificial). 2024. Latin American Index of Artificial Intelligence 2024.

  Looking at recent patterns of AI development in LAC, the report established that Brazil, Chile, Uruguay, and Mexico are ahead of the field in research, development, and adoption. In terms of AI infrastructure, the leaders are Uruguay, Chile, and Costa Rica. Looking at data availability Brazil and Uruguay stand above the rest. In terms of a generally supportive background the two most industrialised economies, Brazil and Mexico, have better rates of patenting, more high-tech workers, unicorn companies, and the manufacturing of cutting-edge technology. Looking at regulation, Chile is seen as a leader in the adoption of AI regulation in “an ethical and responsible way”.

  Aisén Etcheverry, Chile’s minister for science and innovation, who was present at the launch of ILIA 2024 highlighted the intensity of change associated with AI. She said “Soon those who will be voting are not going to get information door-to-door or from what they read in formal media outlets, but instead from social media content. For young people the future they imagine is full of fear about how AI will replace them.”

• The opportunity and risks of AI

  Abbi Dicker 16 May 2025 Hits: 2273

  It is generally accepted that AI brings potentially vast opportunities and increased efficiency, while at the same time posing significant risks of various types of harm.  Eclac Acting Deputy Executive Secretary Javier Medina Vázquez sought to capture the dilemma by commenting that “AI can foster innovation and address key changes in health, education and the environment. AI can optimise government’s administrative processes, improve decision making and better meet citizen’s demands.” At the same time however, he said “it can also widen pre-existing socioeconomic gaps if we don’t act quickly and decisively, especially in terms of investment, infrastructure, education, computing capacities, and regulation”.

  In April 2025 US-based cybersecurity company Palo Alto Networks presented the results of a survey it had conducted among roughly 500 executives in European companies. A majority of 63% of the respondents said AI-related issues presented the most serious operational threats this year. A lower proportion, 42%, said that their companies were making progress to achieve cybersecurity solutions.

  Helmut Reisinger, a Palo Alto Networks manager, said AI-related security threats have become more acute because they are occurring in the context of “fragile” global geopolitics. He noted a changing pattern: in the past cyberattacks could last various days; now data theft can take place in less than a day; in 20% of reported cases, it happened over a few hours.

  Carme Artigas, former secretary for digitalisation and AI in the Spanish government, noted that AI is being used by both sides: those who seek to preserve cyber-security and those who seek to break it. She said AI is key for early threat detection, for automating responses, and to mitigate incidents and detect vulnerabilities. On the other hand, she warned that AI is being used in breach of ethical and legal regulations, and in systems which violate copyright and the privacy rights of users.

  An example of the geopolitical sensitivity of AI was given by Sarah Wynn-Williams, former public policy director of Meta, who left the company and published a book titled Careless People, alleging that company executives had worked “hand in glove” with the Chinese government on potential ways of allowing Beijing to censor and control content, in exchange for access to that lucrative market. Meta denies the allegations and has taken legal action in the US requiring Wynn-Williams to stop promoting and distributing the book.

• AI in LAC, as seen by ChatGPT.

  Abbi Dicker 16 May 2025 Hits: 2530

  One way of illustrating the rapid development of AI capabilities is to ask Chat GPT a question. The specific question, chosen as an example for this report, was “Can you give me a 400-word summary of recent developments in AI in Latin America, with bullet points?” The answer, delivered in seconds, was as follows:

  “Latin America is experiencing a dynamic surge in artificial intelligence (AI) development, marked by significant investments, policy advancements, and regional collaborations. Here’s a concise overview of recent developments:

  Regional Highlights

              •          Brazil: Announced a $4 billion AI investment plan in 2024, focusing on sustainable and socially oriented technologies across sectors like health, agriculture, and education.

              •          Microsoft’s Commitment: Pledged $2.7 billion to enhance Brazil’s cloud and AI infrastructure, aiming to train 5 million individuals and expand data centres in São Paulo.

              •          Argentina: Partnered with a U.S. investor to develop a nuclear-powered data centre in Patagonia, intending to support AI-driven operations and position the country as an SMR (Small Modular Reactor) technology exporter.

              •          Chile, Brazil, and Uruguay: Recognized as regional leaders in AI readiness, excelling in infrastructure, research, and governance frameworks.

   

  Regional Cooperation & Governance

              •          Montevideo Declaration (2024): Brought together representatives from 20 countries to discuss AI governance, resulting in a roadmap for ethical AI development and annual summits to monitor progress.

              •          Cartagena Declaration: Adopted by 17 nations, emphasizing responsible AI use and fostering regional collaboration.

              •          Ibero-American Forum of Digital Parliamentarians: Held in October 2024 to enhance legislative capacities in AI regulation and policy design across Latin America and the Caribbean.

   

   Private Sector & Innovation

              •          MercadoLibre: Leveraging AI for personalized financial services, contributing to its status as Latin America’s most valuable firm with a market capitalization exceeding $100 billion.

              •          Globant: An Argentine tech company generating $350 million in AI-related revenue in 2024, doubling from the previous year, despite facing market challenges.

              •          Startup Ecosystem: A surge in AI-driven startups, particularly in fintech, healthtech, and agtech, is evident, with young entrepreneurs leading innovations despite an aging median population.

   

  Challenges & Opportunities

              •          Infrastructure Deficits: The region faces challenges in computing resources and digital infrastructure, potentially hindering AI advancement.

              •          Labor Market Impact: AI could automate up to 5% of jobs, particularly affecting women and younger workers in formal sectors, highlighting the need for inclusive policies.

              •          Digital Access: Expanding internet connectivity and formal employment opportunities are crucial for maximizing AI’s productivity benefits.

  Latin America’s AI landscape is rapidly evolving, with countries like Brazil and Argentina making significant strides. While opportunities abound, addressing infrastructure gaps and ensuring equitable growth remain paramount.”

• Quantifying the gains

  Abbi Dicker 16 May 2025 Hits: 2481

  Rodrigo Durán, a manager at Chile’s CENIA, highlights a study which seeks to quantify AI gains. This shows that the incorporation of AI tools into the work done by people in Chile’s 100 most important job categories, a total of 5.69m workers, could lift the country’s GDP growth by 1.21 percentage points, an amount equivalent to US$3.381bn. For Latin America and the Caribbean as a whole, there are estimates that AI could boost GDP by 5.4% of US$0.5trn by 2030.

  Latin America artificial intelligence market, 2018-2030 (US$M)

  

  Source: Grand View Research. Artificial Intelligence Market Outlook: Latin America.

  But the potential prize could be bigger. An IMF Working Paper published in October 2024 points out that Latin America has been suffering from poor productivity growth for decades. It links this to the large informal labour market consisting of small and unproductive companies. This in turn may reflect the limited diffusion of new technologies. The report argues, however, that AI could trigger productivity growth by “enabling countries to leapfrog technologies and reach a technological frontier”. As an example, it cites the case of Pix, a digital payments system introduced in Brazil, which has by-passed costly credit card technologies and widened financial inclusion in the country.

  The Working Paper makes a distinction between degrees of exposure to AI, on the one hand, and degrees of complementarity on the other. By employment categories, the jobs most likely to prosper are those with high exposure to AI and low complementarity. An example might be doctors, who will benefit from AI-led cuts in administrative costs releasing more time to spend on their core skill set - attending to patients - which is less vulnerable to automation.

  The IMF says many AI benefits may remain out of reach in Latin America and the Caribbean because of factors driving low diffusion (market competition, regulations, industrial structure, and deficits in physical and human capital). High levels of labour market informality will also act as a drag on uptake. On the other hand, it suggests a policy approach that could accelerate adoption and productivity-enhancing operation. These include lowering barriers to entry for new firms; strengthening competition and anti-trust regulations; and promoting cooperation across academia, industry, and government. Among other steps it suggests upgrading supervision, particularly relating to cyber-security and data protection. It also calls for appropriate unemployment insurance schemes to mitigate the negative impact of AI on workers, helping them where necessary to redeploy to jobs which better match their skillsets. 

• Will the bots destroy Latin American jobs?

  Abbi Dicker 16 May 2025 Hits: 2593

  Initial analysis of worldwide adoption of AI has led to worries that millions of jobs currently performed by humans may be automated and, in effect, taken over by AI. According to a UN Department of Trade and Development (Unctad) report published in April 2025, 40% of global jobs will be impacted in some way by AI over the next decade. The report says that four mechanisms will be at work: first, human workers may be replaced; second, what they do may be complemented by AI; third, automation may be deepened; and fourth, new human jobs may be created. The net effect of these different forces is not necessarily clear. Unctad notes that giant digital companies in the global north favour building up capital, which improves efficiency and could undermine the traditional comparative advantage of the global south – cheaper human labour.

  The report says the AI market will grow to US$4.8trn in value by 2033, but that this will be highly concentrated with around 100 companies, mostly based in the US and China, accounting for half global AI research and development spending. Tech giants such as Apple, Nvidia, and Microsoft have a combined market value of US$3trn, rivalling the GDP of the entire African continent. This level of market dominance, it says, “may widen technological divides, leaving many developing nations at risk of missing out on it benefits”.

  Unctad assessed the position of ‘frontier technologies’. It found that the US has the edge in digital technologies like AI, the internet, big data, blockchain, and 3D printing. China on the other hand leads development in 5G cellular networks, drones, and solar photovoltaics (PVs). The geopolitical and technological rivalry between the two may affect the way LAC countries can access these technologies. Unctad calls for developing countries to expand reliable internet connectivity, to secure high-quality data sets so as to train AI models, as well as building digital skills in education. It also suggests the creation of a global tech facility which would “share AI tools and computing power equitably between nations”. This may, however, be a bridge too far for the big tech companies and superpower governments who will be reluctant to cede market control.

  Perhaps unexpectedly, evidence is emerging that AI requires a growing number of human jobs to be done. These humans are needed to accelerate AI model training and recognition abilities, for example tagging images of objects in the databases used by self-driving cars. AI annotators are employed in Europe and the US, but the work is also outsourced to the global south where many well-educated people are available. Indian industry body Nasscom says it expects annotation revenues will reach US$7bn per annum and employ 1m people by 2030. At present US annotators are paid around US$10-20 per hour; those in poor countries get US$4-8 per hour.

  Fei Fei Li, a US computer scientist who created ImageNet, one of the world’s largest image databanks, said she started using US university students for annotation work but then hired workers all round the world using Mechanical Turk, an online gig-work platform run by Amazon. The World Bank has estimated that between 4.4% and 12.4% of the world’s labour force is involved in gig work, including work on AI models. It is likely that, as models are trained, the need for human input may diminish or evolve. But many in the industry believe annotation work will never cease. Sushovan Das, a dentist working on AI-linked medical projects, told The Economist that annotation work will never disappear. “The world is constantly evolving” he said, “So the AI needs to be improved time and again.”

  Human input to test and refine chatbot output is likely to continue being required. After the launch of ChatGPT in 2022 OpenAI said it was using experts to “qualitatively probe, adversarially test, and generally provide feedback” on its models. As a result of that process the models now refuse to respond to some prompts (for example, requests to write social media content promoting al-Qaeda, a terrorist group).

• The big AI obstacle in LAC: labour informality

  Abbi Dicker 16 May 2025 Hits: 2458

  In a recent blog IMF economists identify labour sector informality as one of the major obstacles to AI take-up in Latin America and the Caribbean. A large part of the regional labour force operates entirely outside the formal economy. Typically, they work on their own or in small and medium sized enterprises, paying no taxes and receiving no social security benefits. These companies often struggle to survive and are less likely to attract investment in new technologies such as AI. According to the IMF less than half of the region’s jobs will be impacted by AI. This will protect the individuals and companies involved from the short-term disruption caused by the arrival of AI. But in the longer term it will also shut them out from the pro-growth benefits of the new technology.

  Labour informality is highest in countries like Bolivia, Peru, and Honduras. By job categories those in healthcare are likely to gain from AI-enhanced productivity gains without major job losses. However, those currently in call centres are more likely to be displaced by AI machines. It follows, according to the authors, that governments interested in long-term development should seek to formalise employment, by extending financial support, reducing licensing restrictions, and providing improved education, training, and social safety nets.

  Latin America has fewer jobs highly exposed to AI, partly due to its large informal sector

  

  Source: Werner, Alejandro, and Carolina Villegas-Sanchez. “How Artificial Intelligence Can Boost Productivity in Latin America.” IMF Blog, 20 March, 2025.

• The DeepSeek challenge

  Abbi Dicker 16 May 2025 Hits: 2487

  The sense that the US retains a commanding lead in the AI race suffered a significant blow in early 2025 when DeepSeek, a small Chinese computer lab using open sources, released its R1 LLM model. Unlike its US rivals, DeepSeek made its research approach to LLM-building public; it also revealed it was able to use Nvidia GPU chips more effectively than its rivals. Most significantly. DeepSeek claimed it used only 2,048 Nvidia H800 chips and an investment of US$5.6m to train a model with 671bn parameters.

  Industry specialists say this is only a fraction of the chips and investment the OpenAI would require training similar-sized models. The breakthrough, announced in January, led to an immediate reaction on the US stock exchange, where Nvidia shares dropped by a record 17% in a single day, 27 January. Although they subsequently recovered, the development suggests the superpower AI technology race remains wide open. There has been recent news of Chinese advances in other technology fields, with electric vehicle companies such as BYD and CATL developing batteries that can be charged in five minutes, also ahead of US/European standards.

  The leader of DeepSeek, Liang Wenfeng, is a hedge fund manager with a long-standing interest in trade-focused algorithms. One of his business partners, talking about their first meeting, described him as “This very nerdy guy with a terrible hairstyle talking about building a 10,000 chip cluster to train his own models. We didn’t take him seriously.” But the DeepSeek R1 release has underlined the scope for innovation and more efficient use of chips. Ritwik Gupta, a policy researcher at the University of California, pointed out to the Financial Times that “The first person to train models has to expend lots of resources to get there. But the second mover can get there cheaper and more quickly.”

  From Latin America’s point of view AI-focused technological rivalry between the US and China can be seen as a double-edged sword. On the one hand it may lower the cost of entry into the industry and provide scope to shop around for competitive alliances; on the other, however, it could lead to increased dependence on either China or US-developed models. There are different versions of the AI future, and no-one is entirely sure about the probability of various scenarios. One is a straight win-lose, or zero-sum game between the US and China. In that vision Latin American countries will have to choose which technology to follow. Some go as far as saying that DeepSeek’s R1 model is a geopolitical gambit to embed it in the infrastructure of developing nations, strengthening digital ties beyond Western control.

  In a rival scenario, however, LAC countries may find they can build a more decentralised AI ecosystem where competition and regulation create more development pathways. A significant point is that DeepSeek can be adapted to work with Spanish or Portuguese LLMs, reducing reliance on English. In this scenario both US and Chinese companies may co-exist in the AI space in LAC.

  Some assumptions about how AI will develop in Latin America may prove to be wrong. One is that only countries with low-cost energy sources, such as those of hydroelectric-rich Brazil or oil-and-gas focused Argentina, will be able to build a sustainable AI eco-system. The counter view is that countries with moderate energy costs but larger investment in digital infrastructure, such as Colomba and Mexico, may lead the way.    

  That said, the US-led drive towards increased protectionism and punitive tariffs for silicon chips and other key technology products could cancel out those gains.

• The promise of Latam-GPT

  Abbi Dicker 16 May 2025 Hits: 2962

  Chile’s National Centre for Artificial Intelligence has been involved in a project over the last two years to create Latam-GPT, a new large language model which specifically focuses on Latin American culture and outlook. CENIA says that although current LLMs are of high quality, “their understanding of the Latin American context could be enriched and refined”. CENIA is also eager to experiment with a more collaborative way of building an LLM, so that the product is “public, open, and inclusive”. Roughly 30 institutions and 60 experts at regional level have been involved in the work.

  Rodrigo Durán of CENIA told the BBC that “The importance of this model lies in the fact that it allows the culture, idiosyncrasies, and world view that stem from the history of Latin America to be captured in a language model, which is truly the most transformative technology we’ve seen in recent decades.” As part of the partnership approach project leaders say they have generated a data corpus of over 50trn parameters, comparable to OpenAI’s Chat GPT-3.5. The University of Tarapacá in Chile is proposing to acquire a supercomputer that will be used to train the LLM. Initial training expected to 35-40 days of continuous processing.

• Rising data centre energy demand

  Abbi Dicker 16 May 2025 Hits: 2461

  Electricity consumption in the world’s data centres will double between now and 2030, according to a report by the International Energy Agency (IEA) which says that the growth will be concentrated in the US and China.  The two superpowers will be responsible for around 80% of the growth in data centre energy demand, with the European Union also growing but still lagging in third place. The agency says that under its central forecast, data centre energy demand will rise from 415 terawatts hour (TWh) in 2024 to 945TWh in 2030, representing a 15% annual growth rate.

  IEA also estimates that data centres currently represent 1.5% of world electricity use, a share which will double to 3.0% by the end of this decade. Here too growth will be strongest in China, the US, Japan, and Europe, in that order. Per capita electricity demand by 2030 will be 1,200KWh in the US, 270KWh in Japan, 200KWh in China, and 165KWh in Europe. Per capita electricity consumption in India will be much lower, at 15KWh.

  Olade, the Latin American Energy Organisation, has similar data for the expansion of data centres. It says there are 455 data centres operating in the LAC region, each using some 50GWh of electric power per annum. Andrés Rebolledo, Olade executive secretary, says AI applications will be competing with residential, industrial, and transport demand for power.

  The essential worry is that the expansion of power-hungry data centres around the world could lead to increased power use and pollution and contribute negatively to climate change. In some areas data centre energy growth has been massive: for example, one fifth of all electricity consumed in Ireland is now used to power data centres. The centres have a multitude of activities, including streaming video and financial transaction processing; the AI share of that work is expected to rise exponentially.  

  Whether or not data centres have a negative effect on the environment depends on the interplay between various factors. A positive is that AI is already playing a role in reducing pollution and optimising energy use across various industries. A negative is that expanding data centre demand may be met by burning carbon-emitting fossil fuels (oil, gas, and coal among them). On the other hand, some analysts describe the abundance of renewable energy sources in Latin America (hydro, wind, solar, and biofuels) as a comparative advantage, since it may give a boost to the AI sector.  

  The idea that Latin America’s strength in renewables gives it an advantage in AI has been backed by Rolando Lemos, a scientist from the Rio de Janeiro-based Institute for Technology and Society. In an article written for the World Economic Forum (WEF) earlier this year, he argued that Latin America “can set new standards in AI” thanks to abundant renewable energy, rich biodiversity, and deep cultural and linguistic diversity. The region, and particularly Brazil, generates most of its electricity from clean sources. In a suggestion echoing a proposal by Unctad (see above), Lemos calls for LAC countries to form a “multi-stakeholder consortium to pool resources, democratise access to AI and promote techno diversity”.

  Lemos says that there is a fork in the AI road ahead for Latin American countries. One route will be for them to “follow global AI trends at a distance”, while the other more challenging one is “to set an entirely new standard for how AI is powered and governed”. Lemos has an optimistic message, highlighting a “wave of innovation in areas such as agritech, industrial AI, and health research, all backed by rising venture capital inflows”. For this potential to be captured, however, he says it will be essential to overcome inequities in computer infrastructure and bridge a staffing gap by recruiting some 1.3m technology professionals.  

• The European strategy

  Abbi Dicker 16 May 2025 Hits: 2575

  Speaking in April, Henna Virkkunen, European Commission vice president for technology sovereignty, said the AI race is far from over. She presented a plan to build five gigafactories in Europe, noting “Of course, the US has the large technology companies, and China has its own vision, but in the EU, we must take advantage of our own strengths.”  Virkkunen said the commission is aiming to mobilise EUR200bn (US$227bn) in public-private partnership investment for AI. This would include spending some EUR20bn (US$22.7bn) on the five gigafactories, a term invented by Elon Musk to describe major battery production plants.

  In the EU’s vision however, the gigafactories would produce not batteries but a range of AI-related high-tech products and services. Each facility would have 100,000 latest-generation silicon chips, radically increasing computing and processing power. Investment would triple data-centre capacity in Europe. Virkkunen estimates that only 13% of European industry is using AI technology, meaning the continent has an as-yet untapped potential “to become more productive and to develop new services using applied AI”.

  The European Commission is also considering whether to amend the AI law which came into force last year. Virkkunen said Brussels remains committed to a risk-based approach to regulation but is also now considering whether bureaucracy and red tape could be further reduced, particularly in relation to AI use by small and medium-sized enterprises. The Commission, she said, was also interested in feedback on the law, to reduce any uncertainty that might be holding back technology development. Various lobby groups say there is a gap in the new law, since companies are not required to assess or mitigate generative IA models for risks, including possible violations of civil and human rights. There is also an open debate on the treatment of intellectual property (IP) and copyright in the content of learning models. Virkkunen acknowledged that there is “a balance to be found” on these issues.

• Latin America faces regulatory challenges

  Abbi Dicker 16 May 2025 Hits: 2420

  A wide range of regulatory and legislative approaches to AI are being discussed in the LAC countries. Some are modelled on the EU’s Artificial Intelligence Act, which came into effect in August 2024 but which itself may be amended (see above). Some countries, such as Brazil and Chile, already have detailed proposals for legislation. Mexico and Argentina are a little further behind. Most of the largest jurisdictions in the region are following a “risk-based approach” meaning that they prohibit unacceptably risky AI applications, while setting out safeguards for higher risk ones, such as specific regulatory approval or additional disclosure requirements. However, some countries such as Argentina are actively discussing a “regulation light” model designed to attract investment.

  A recent paper by Brookings, a US-based think tank, has been critical of some of the regulation proposed for AI in Latin America, noting “In general, bills tend to be quite vague, fail to provide actionable measures or a clear definition of AI, or grant the executive branch with excessive powers to define when and how regulations should apply.”

  Public opinion, however, seems to be clearly in favour of regulation. A poll conducted by Ipsos and lobby group Luminate revealed that 55% of Latin American respondents believe AI should be regulated. Some 40% believe AI can influence electoral outputs while 37% believe the technology could increase inequality. The poll also revealed that 59% say they know how to operate AI applications, and 46% say they use AI tools in everyday life. Respondents also appear to be wary of AI’s effects. 54% said they opposed the use of AI in courts of justice; 51% are against AI systems drafting the text of laws and regulations; and 50% oppose AI systems being used to determine eligibility for social security payments nationwide

  

  White & Case LLP. Latin America Focus: 2024 – AI Regulation. 12 December, 2024.

• Brazil: pushing new legislation

  Abbi Dicker 16 May 2025 Hits: 2432

  Brazil’s AI regulation effort, Bill 2,338/2023, has been working its way through congress. It aims to promote secure systems for the benefit of individuals, democracy and technological development. It says that developers, distributors, and users (which it collectively calls ‘AI agents’) have obligations linked to the risk-classification of the systems they are using. The riskiest systems will be banned completely, with high-risk ones subject to monitoring and extra safeguards. All AI systems must be conducted under conditions of transparency, reliability, safety, and fairness. Developers of high-risk AI products must prepare “impact reports” on how they affect individuals’ fundamental rights.

  Unlike the EU Act, Brazil’s draft legislation imposes some additional obligations on AI agents including a requirement to use energy efficiently, and an obligation to publicly identify all content that is AI-generated. The bill also creates an umbrella AI National Regulation and Governance System to coordinate regulatory bodies. Sitting under that umbrella will be the Brazilian Data Protection Authority, and an AI Authority with power to issue warnings, impose fines, publicly disclose violations, and prohibit the use of certain databases. With a view to fostering innovation the bill also allows for the creation of a regulatory sandbox where new and experimental programmes can be tested. These regulatory innovations come with a background of strong government support for AI development. In July 2024 the government pledged US$4.07bn to boost technology development and competitiveness in the sector as part of a 2024-2028 AI plan. Also significant is the Brazilian Digital Transformation Strategy (E-Digital) launched in 2018. The current version which covers 2022-2026, sets out priorities in infrastructure, research and development, and trust in the digital environment.

• Mexico: slow development

  Abbi Dicker 16 May 2025 Hits: 2415

  As the second largest Latin American economy, sharing a border with a technology powerhouse to the north, it might be expected that Mexico would be a leader in the regional AI race. However, development has been rather slow, suggesting that the government of President Claudia Sheinbaum is not prioritising AI. There has been significant discussion, but no major government-backed AI laws or regulations are in force.

  The most relevant privately backed bill was submitted by a member of congress in April 2024. It states that AI tools must protect human rights and asserts that both Mexican and foreign based developers are subject to its provisions. Although it follows the EU risk-based approach, it is thought to be somewhat more restrictive. It defines three levels of risk - unacceptable, high, and low – and goes on to say regulatory permits must be secured before any new products are launched in Mexico. The bill requires LLM developers to clear all rights held by intellectual property owners ahead of the training stage.

  Interest in the issues posed by AI remains high, however. In April this year the Mexican Supreme Court upheld a law passed by the state of Zacatecas, designed to protect children from digital violence, including that perpetrated by means of artificial intelligence. The state government had challenged the law, saying that it lacked a clear definition of artificial intelligence. Justice Margarita Ríos Farjat said AI presented serious risks since “on the one hand it can be used to generate and distribute highly realistic child abuse content, while on the other it can be used by delinquents to manipulate and mislead”. Earlier in the year the supreme court had also upheld a law in Sinaloa state introducing sentences of up to six years for the use of AI to create sexual content.

• Chile: leader in regulation

  Abbi Dicker 16 May 2025 Hits: 2465

  Chile is widely seen as the regional leader in AI regulation. In 2024 the government submitted a comprehensive AI bill to congress. Its stated aim is to promote the development and implementation of AI systems while respecting democratic principles, the rule of law, and human rights. It is seen as applying to developers in Chile but also to foreign developers, when the output data becomes available in Chile.

  Using the risk-based approach, the bill defines four categories. “Unacceptable” risk applies to AI which is incompatible with human rights and might include social scoring or subliminal manipulation. “High risk” includes content that affects health and safety but could be necessary in certain situations such as experimental data training. “Limited risk” includes content that is free from potential manipulation or error and is permitted as long as it meets transparency requirements. The fourth category is “no risk”, which is unregulated. Chile’s bill is also aligned with Unesco and OECD recommended guidelines.

• Argentina: seeking low regulation to attract investors

  Abbi Dicker 16 May 2025 Hits: 2304

  In December 2023 Argentina transitioned from a centre-left Peronist (Partido Justicialista – PJ) government to a far right, pro-market administration under Javier Millei. Some AI bills and draft regulations initiated before that date may not progress further for political reasons. But Millei’s vision of a hands off, lightly regulated sector may also face obstacles, since no party currently controls an outright majority in congress.

  In 2023 the government issued a policy document titled “Recommendations for a reliable Artificial Intelligence” which incorporated ethical principles, AI transparency, and accountability. There are also draft laws awaiting consideration, proposing an AI System based on diversity, inclusion, peace, and justice, as well as one proposing the creation of a Federal Council of Artificial Intelligence. Bills have meanwhile been submitted for consideration by the senate, using European-style risk assessments. One version identifies limited, minimal, high, and unacceptable risk levels. Some also call for a national registry of AI systems with associated impact assessments.

  In June 2024 Demian Reidelan, a key adviser to Milei, said the president was pitching to turn Argentina into the world’s “fourth AI hub,” promising to reduce regulation and attract top US technology companies to invest. The comments came after the Argentine president had met with chief executives from OpenAI, Google, Apple, and Meta. Reidel told the Financial Times that investing in Argentina would provide such companies with a hedge against rising regulatory risks in the US and Europe. He said: “Argentina has a president who is actually putting forth the ideas of freedom, low regulation, and free enterprise.”

• Colombia: more than half the country’s companies are investing in AI

  Abbi Dicker 16 May 2025 Hits: 2751

  In late 2019 Colombia’s National Policy for Digital Transformation and Artificial Intelligence was launched. In 2020 a task force for “development and implementation of AI” in the country was appointed. Among various initiatives under discussion in Colombia, Bill 059 sets out legal guidelines for AI development and use, incorporating provisions for human rights protection. A National AI commission will assess and authorise operations. There are also plans for a compliance framework with fines and penalties to be charged if it is transgressed.

  A business survey carried out in Colombia by global software company SAP gave an optimistic assessment of AI uptake in the country, with 69% of companies consulted saying AI will have a “significant impact”, and 54% saying they plan to invest in AI systems this year. Medium and large-scale companies are leading the pace of adoption. AI systems are having greatest impact in customer care, marketing and communications, technology/automation, and operations, in that order. 44% of Colombian companies say they have already captured tangible benefits from using AI.

  Colombian vice-president Francia Márquez addressed the United Nations in New York in April 2025 during a special forum on Afro-descendant communities. An Afro-Colombian herself, Márquez referred to recent research that shows data used as inputs for LLMs can continue to reflect racist and discriminatory content. The vice president warned that AI can “aggravate” racism and called for a “global forum” to promote a more ethical and inclusive approach to the technology.  She said that during her first year in office she had received around 12,000 racist insults and messages. Márquez said AI outputs are not “neutral nor objective” and that they can “intensify racial profiling and stereotypes that affect young afro-descendant community members”.

• Conclusion – a game of high stakes

  Abbi Dicker 16 May 2025 Hits: 2160

  This report has highlighted some of the key issues around the introduction of artificial intelligence in Latin America and the Caribbean. Three major points can be made. The first is that AI has begun, and will continue, to revolutionise the economies of the region. It will impact the way companies do business, the way governments run public services, and the way millions of people go out to work and earn a living. Those impacts will not be uniform. The adoption process will be bumpy. Negative or harmful effects, ranging from job losses to invasions of privacy and political manipulation will be present.

  The second point is that over the next few years the AI wave will be particularly intense and uncertain because it is coming at a time of heightened technology-based super-power rivalry between the US and China. LAC governments have to adopt AI policies and regulations at a time of a major trade war and rising protectionist sentiment. Both Washington and Beijing are seeking to control silicon chip supply chains. Both are seeking to capture AI technologies and deny their benefits to their rivals. The current situation includes the application of punitive tariffs on US-China trade and outright bans on sensitive products and services. This is causing great uncertainty for the LAC countries.

  Finally, in third place, it is important to acknowledge that the LAC countries are not without agency. In fact, they have a valuable “window of opportunity” as they are drawing up rules and regulations for the nascent AI industry. AI adoption is inevitably going to have a mix of benign and harmful impacts. But a considered and well-written regulatory framework will help move that mix in a more benign direction. Governments can also tackle structural issues – and in particular investment in digital infrastructure and education, as well as the need to formalise the large pool of informal labour – in a way that optimises the benefits of AI.