Doble Lectura

Every Monday we take a paper or a policy report and read it slowly: first what it says, and then what it means from Latin America. It is the unhurried companion to the daily blog.

Nothing here. Try another topic or keyword.

• Sep 28, 2026 The same GPT-4 that improved the work of 758 consultants made them get another task wrong

  A preregistered experiment with 758 BCG consultants shows that GPT-4 sped up and improved their work on product development tasks. On a business case that fell outside its reach, the opposite happened: those who used the tool got it right less often, and their wrong answers sounded more convincing.

  true Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality, by Dell'Acqua, McFowland, Mollick, Lifshitz, Kellogg, Lakhani et al. · Organization Science, 2026.

• Sep 21, 2026 One person with AI performed like a team of two, and did worse at picking their best idea

  A preregistered field experiment with 791 Procter & Gamble professionals shows that working alone with AI matches the quality of what a cross-functional pair produces without it. The twist is in the next step: those who used AI seem to have been less accurate in choosing which of their own ideas to develop.

  true The Cybernetic Teammate: A Field Experiment on Generative AI and Teamwork, by Dell'Acqua, Ayoubi, Lifshitz, Sadun, Mollick et al. · Organization Science, 2026.

• Sep 14, 2026 They asked the models about the risk of catastrophe, and they answer higher than humans do

  A panel of four frontier models forecasts a 0.47% probability that AI causes a catastrophe killing at least 10% of humanity by 2030, and 6% by 2050, between 4.8 and 6.8 times what superforecasters estimate. The paper's twist: those forecasts become more informative precisely as the risk rises.

  true Automated Forecasts of Catastrophic Risks, by Abaluck, Karger, Merrill, Tetlock, Vivalt and Williams · Forecasting Research Institute, working paper, September 2026.

• Sep 7, 2026 A mental well-being chatbot: supplement, medicine, or yoga instructor?

  Twenty-four experts in clinical practice, ethics, public policy, and health technology design, more than a hundred regulatory documents, and a conclusion that is uncomfortable for the industry: what determines whether a mental well-being AI is well designed is not how good it is at conversation, but what concrete benefit it promises and to whom. A tool meant to serve everyone answers to no one.

  true Framing Responsible Design of AI for Mental Well-Being: AI as Primary Care, Nutritional Supplement, or Yoga Instructor?, by Cooper, Guridi, Hwang, Kolko, McGinty and Yang · CHI 2026.

• Aug 31, 2026 With AI, homework grades go up and test scores go down

  Thirty months of records from 26,811 secondary school students in a Chinese county. After adopting generative AI, homework grades rise 18% and the time spent on it falls by a third, but closed-book test scores fall 20% within six months, and the effect on admission exams takes two years to appear in full.

  true The Generative AI Learning Penalty: Evidence from Chinese Secondary Education, by Strömberg, Lei and Wu · SSRN, June 2026.

• Aug 24, 2026 Almost half of real-world AI use disappears when it is measured only as work

  An academic consortium pooled seven sources of real conversations with AI assistants and annotated them with a single taxonomy. When the occupational filter of the Anthropic Economic Index is applied to them, 48% of the conversations are discarded, and what gets discarded is not noise: that is where health, relationships and much of the sensitive content are concentrated.

  true The AI Observatory: A Public Measure of Real-World AI Use, by Longpre, Reuel, Ki and others · Preprint, August 2026.

• Aug 17, 2026 AI books sell little each, but they lower what every title earns

  Almost 14,500 self-published novels on Amazon, with AI detection on the full text and real daily sales. Books with substantial AI make up 20% of the catalog and just 12% of sales, but the catalog grew much faster than the money available, and books with no AI detected also earn less than before.

  true Generative AI floods and dilutes the market for books, by Chakrabarty, Liu, Ginsburg and Dhillon · arXiv, July 2026.

• Aug 10, 2026 AI has already reached almost every occupation, but it covers barely a fifth of their tasks

  Google mapped 15 million Gemini conversations against the official US taxonomies of occupations and tasks. Adoption reaches 68% of occupations, but in the median occupation it covers 21% of its tasks, and less than 10% of use in non-routine cognitive work seeks to have AI do the whole task.

  true Google's AI & Economy ATLAS v1.0: Mapping Gemini Usage in the Economy, by Iscenko, Strand and others · Google, July 2026.

• Aug 3, 2026 Governments already use generative AI, but almost none measure whether it works

  The OECD reviewed the official guidelines of fourteen countries and found that almost all of them say how to use AI responsibly and almost none say how to evaluate whether the experiment worked. The list of countries reviewed includes none from Latin America.

  true Generative AI experimentation in government: Learning from emerging guidelines, by Tõnurist and von Knebel · OECD Working Papers on Public Governance No. 93, 2026.

• Aug 3, 2026 Anthropic learns to read what its models are about to say

  A new interpretability technique shows that models maintain a small set of concepts available for reporting and reasoning, on top of a much larger volume of automatic processing. In alignment tests, deliberations that the final answer did not show appeared there.

  true Verbalizable Representations Form a Global Workspace in Language Models, by Gurnee, Sofroniew, Lindsey and others · Anthropic, Transformer Circuits Thread, July 2026.

• Jul 20, 2026 Being exposed to AI is not the same as being at risk of losing your job

  OpenAI cross-referenced more than 900 occupations with three layers of analysis and placed 18% of jobs in the United States in the group at highest risk of near-term automation; the rest get reorganized, grow or change little. Its thesis: AI 'exposure', on its own, does not tell you where it is going to hurt.

  true The AI Jobs Transition Framework: Mapping AI's Near-Term Impact on Jobs, by Richmond · OpenAI, April 2026.

• Jul 13, 2026 AI can be talked into things with the same tricks that work on people

  A preregistered experiment with 126,000 conversations shows that applying classic principles of persuasion, such as authority or liking, raises from 35% to 51% the probability that three models go along with requests they should refuse.

  true Persuading large language models to comply with objectionable requests, by Meincke, Shapiro, Duckworth and others · PNAS, 2026.

• Jul 6, 2026 Brazil debates how to regulate AI, but labor remains a blind spot

  An analysis of 57 AI policy documents in Brazil between 2019 and 2024. The country is moving forward with a bill, a multibillion plan, and contested governance, but the central draft barely mentions labor rights and union mobilization is still incipient.

  true AI Policy Debates in Brazil: Struggles over Regulation, Governance and Labour, by Grohmann, Paraná, Valente and Figaro · Global Political Economy, 2025.

• Jun 29, 2026 Technology and democracy: which is the question and which the answer

  An essay by Iñaki Goñi shows that two fields that study almost the same thing flip the recipe: for deliberative democracy, technology comes to fix participation; for science and technology studies, participation comes to legitimize technology. And he warns that, without dialogue between the two, both sides end up in pure solutionism.

  true Citizen participation and technology: lessons from the fields of deliberative democracy and science and technology studies, by Goñi · Humanities & Social Sciences Communications, 2025.

• Jun 25, 2026 Governing AI in the state is a problem of institutional design, not technology

  A technology director in the Brazilian public sector proposes leaving static compliance behind and governing frontier AI with adaptive risk management. His thesis: since no one knows how fast the technology will advance through 2030, fixed rules age badly, and the state needs to monitor capabilities, scale controls according to signals, and redesign organizations.

  true Governing frontier general-purpose AI in the public sector: adaptive risk management and policy capacity under uncertainty through 2030, by Correa Xavier · Preprint (arXiv), 2026.

• Jun 22, 2026 Announcing that an AI moderates the conversation is enough to make fewer people want to take part

  An experiment with 1,850 Germans finds an 'AI penalty': simply knowing that an automated tool will facilitate the deliberation, without having used it, is enough to lower willingness to participate and expected quality. And the gap does not follow education, but what each person thinks of AI.

  true Artificial Intelligence in deliberation: The AI penalty and the emergence of a new deliberative divide, by Jungherr and Rauchfleisch · Government Information Quarterly, 2025.

• Jun 19, 2026 With AI in the mix, the code you hand in is no longer enough to prove you can program

  Simulated programming interviews with AI allowed show that evaluators do not change what they understand as expertise, but they do change the evidence they ask for. The code handed in no longer suffices: what now gives away a good programmer is which tool they choose, how they talk to it, and when they distrust what the model spits out.

  true Evolving Enactions of Expertise: Software Engineers' Evaluation and Demonstration of Coding Expertise with AI Coding Assistants, by Jang, Sakashita, Niinuma and Gupta · CHI 2026.

• Jun 16, 2026 The poorer the country, the more AI is used for learning

  An analysis of 686,000 conversations in 227 countries finds that educational use of AI rises where income falls, the opposite of what happened with the internet. And that English is taking over as the lingua franca of AI precisely where local languages work worst in the models.

  true How Early Adopters Used Generative AI Worldwide: Variation by Country Income and Language, by Daepp and Slaughter · arXiv preprint, 2026.

• Jun 16, 2026 AI writes the code, but the result depends on how much you know about the problem

  Anthropic analyzed some 400,000 Claude Code sessions and found a consistent pattern: the person who directs the AI well is not necessarily someone who knows how to program, but someone who knows the problem's domain. Experience in the field weighs more than profession, and the gap between novices and everyone else is still there.

  true Agentic coding and persistent returns to expertise, by Hitzig, Massenkoff, Lyubich, Heller and McCrory · Anthropic, June 2026.

• Jun 13, 2026 Users, developers, and authorities see Chile as a 'testing ground' for AI

  A qualitative study brings together those who use, build, and regulate AI in Chile. They share an uncomfortable image, that of the country as an experimental subject for the Global North, and show that almost no one demands transparency from AI in the abstract, but people do demand it when a specific system decides on their medical leave or their children's school.

  true Reimagining AI in Latin America: situated narratives of users, developers, and decision-makers on understanding and governing AI, by Correa et al. · Communication and Change, 2025.

• Jun 10, 2026 Not everything called 'AI regulation' regulates the same thing

  A team from Stanford and Harvard builds a taxonomy to compare the artificial intelligence laws of five jurisdictions, including Brazil. The finding: using the same word for voluntary guidelines and for binding laws creates a false sense of protection and opens the door to regulatory capture.

  true Comparing Apples to Oranges: A Taxonomy for Navigating the Global Landscape of AI Regulation, by Alanoca, Gur-Arieh, Zick and Klyman · FAccT 2025.

• Jun 7, 2026 Imperfect AI images open up better design conversations

  A study used image-generative AI as a mediator in interviews with migrant-community residents to redesign a park in Los Angeles. The finding: the 'perfect' images shut the conversation down, and the imperfect ones were the ones that revealed memories and values the design had not yet captured.

  true From Fake Perfects to Conversational Imperfects: Exploring Image-Generative AI as a Boundary Object for Participatory Design of Public Spaces, by Guridi, Hwang, Santo, Goula, Cheyre, Humphreys and Rangel · Proc. ACM HCI (CSCW), 2025.

• Jun 4, 2026 Politicians and civil servants don't want the same things from AI

  Twenty interviews inside government in Chile and Uruguay to understand why governments barely use AI when consulting citizens. The uncomfortable finding: the problem is not the algorithm, but that no one takes responsibility for the decision.

  true Thoughtful Adoption of NLP for Civic Participation: Understanding Differences Among Policymakers, by Guridi, Cheyre and Yang · Proc. ACM HCI (CSCW), 2025.

• Jun 4, 2026 Adopting AI in UX design teams is a negotiation over who decides and who counts

  Workshops and interviews with 15 UX professionals in the United States, Canada, and South Korea show that AI adoption rarely produces the promised efficiency gains: instead, it adds invisible work, redistributes roles, and leaves designers themselves out of the decisions that affect them most.

  true The Values of Value in AI Adoption: Rethinking Efficiency in UX Designers' Workplaces, by Cha, Wieczorek and Wong · CHI 2026.