Generative artificial intelligence is entering Latin American classrooms as a tutor and as an aid to game design, but the regional evidence has not been synthesized by role. This narrative review summarizes studies set in or about Latin America and published from 2013 to 2025. Fifty-six works were selected. Sixteen core works addressed tutoring, 14 addressed gamified or game-based activities with artificial intelligence, 10 addressed risks, 10 addressed adoption and teacher readiness, and six were reviews. Tutor studies reported faster task completion, feedback at scale and mixed learning gains, including two randomized trials with opposite results. Generative models appeared as game designers mainly in teacher preparation of playful activities, and none of the gamification works compared them with conventionally designed games. The reported risks were plagiarism, lower critical thinking with dependence, inaccurate output, and unequal access by region and gender. Only abstracts were read, and the coding awaits verification by the authors. Designs that separate the generative component, scaffold the tutor and train users in verification are the most direct way to test these claims.
La inteligencia artificial generativa llega a las aulas latinoamericanas como tutora y como apoyo al diseño de juegos, pero la evidencia regional no se ha sintetizado por rol. Esta revisión narrativa resume estudios ubicados en América Latina o sobre ella y publicados entre 2013 y 2025. Se seleccionaron 56 trabajos. Dieciséis trabajos centrales trataron la tutoría, 14 las actividades gamificadas o basadas en juegos con inteligencia artificial, 10 los riesgos, 10 la adopción y la preparación docente y seis fueron revisiones. Los estudios sobre tutoría informaron mayor rapidez en las tareas, retroalimentación a escala y ganancias de aprendizaje mixtas, con dos ensayos aleatorizados de resultado opuesto. Los modelos generativos aparecieron como diseñadores de juegos principalmente en la preparación docente de actividades lúdicas, y ningún trabajo de gamificación los comparó con juegos diseñados de forma convencional. Los riesgos descritos fueron el plagio, el menor pensamiento crítico con la dependencia, la salida inexacta y el acceso desigual según región y género. Solo se leyeron resúmenes y la codificación espera la verificación de los autores. Los diseños que separan el componente generativo, andamian la tutoría y entrenan en verificación son la vía más directa para probar estas afirmaciones.
Chatbot tutors raised learning outcomes in a meta-analysis of 24 randomized studies, with larger effects in higher education than in primary or secondary schooling.(1) A later synthesis of 62 studies and 135 effect sizes reported a large effect that fell to small or moderate after correction for publication bias.(2) Large language models add content creation and personalized interaction on demand, and they require fact-checking and critical judgment from the user.(3) Their tendency to hallucinate, that is, to produce plausible but nonfactual text, is documented.(4) Dialogue-based tutors predate these models: a review of 33 of them found evaluations of learning gains for 90.91% of the systems.(5)
Regional evidence is growing faster than its synthesis. A systematic review of 25 papers published from 2021 to 2023 covered generative AI in higher education across eight Latin American countries,(6) and a scoping review of 84 documents from 2022 to 2024 mapped the competencies teachers need.(7) Student data can be large: 4,811 students at eight Ecuadorian universities answered one plagiarism questionnaire.(8) Structural inequality conditions any deployment, as a review of 28 articles linking the gender digital divide to education in Latin American countries indicated.(9)
Gamification adds a second thread. A meta-analysis reported significant effects on cognitive (g = 0.49), motivational (g = 0.36) and behavioral (g = 0.25) learning outcomes,(10) and a later meta-analysis of 22 experimental studies found a moderately positive effect on academic performance (g = 0.782).(11) Generative models now reach the design step itself. ChatGPT was proposed for brainstorming themes and mechanisms, drafting game components and critiquing prototypes in serious board game design.(12) A language-model tool generated gamified programming exercises, most of which were ready for use with little human effort.(13)
Designers are less enthusiastic than tool builders. In a survey of 42 game design professionals, generative AI accelerated ideation and prototyping, yet respondents raised concerns about originality and creative dependency.
The objective of this narrative review was to synthesize studies set in or about Latin America on generative AI in two classroom roles, tutor and designer of games and gamified activities, and to record the opportunities, risks and conditions of adoption they report. It asks which tutor roles are documented, how generative models take part in designing games and gamified activities, and which risks recur. Its scope differs from earlier contributions to this journal on a prompt engineering game and on ChatGPT in libraries: it concerns teaching and learning in schools and universities, and it neither treats prompting as a game nor addresses library services.
The 56 core works came from Ecuador (n = 17), Peru (n = 14), Brazil (n = 5), Colombia (n = 5), Chile (n = 3), Mexico (n = 2), Uruguay (n = 1) and Guatemala (n = 1), plus two multi-country studies and six reviews. Sixteen works addressed tutoring, 14 addressed games and gamified activities, 10 addressed risks, and 10 addressed adoption and teacher readiness. Seventeen titles were in Spanish or Portuguese. Only two of the 16 tutor works were randomized, and four others were quasi-experimental (Table 1).
Table 1. Synthesis of the 56 core works by axis.
Axis | Works (n) | Countries | Designs | Main reported result | Main limit |
|---|---|---|---|---|---|
Tutor and assistant | 16 | Peru 7, Chile 3, Ecuador 2, Mexico 2, Brazil 1, Colombia 1 | 2 randomized, 4 quasi-experimental, 1 pre-experimental, 9 other | Faster task completion and feedback; mixed learning gains | Samples of 22 to 347 where reported; short interventions |
Game and gamified design | 14 | Ecuador 9, Brazil 3, Colombia 2 | 4 quasi-experimental, 1 action research, 1 game-data analysis, 1 rapid review, 6 surveys, 1 qualitative | Teacher-side preparation of playful activities; motivation reported | AI and gamification confounded; no comparison with conventional games |
Risk | 10 | Peru 3, Ecuador 3, Brazil 1, Colombia 1, Uruguay 1, three-country 1 | 6 cross-sectional, 3 qualitative, 1 framework with a planned RCT | Plagiarism, dependence, unequal access | Self-report; samples of 9 to 4,811 |
Adoption and teacher readiness | 10 | Peru 4, Ecuador 3, Colombia 1, Guatemala 1, five-country 1 | 1 quasi-experimental, 8 surveys or mixed methods, 1 qualitative | Adoption follows perceived benefit and peers; teacher training gaps | Perceptions, not learning outcomes |
Review | 6 | Latin America | 3 systematic, 1 scoping, 2 literature reviews | Opportunities with ethical and infrastructure limits | Overlap with core studies not assessed |
Note: RCT = randomized controlled trial.
The strongest tutor design was a randomized trial of 162 diplomacy students in Colombia, in which ChatGPT-assisted debates produced higher combined scores (multivariate analysis of covariance) and gains in understanding of complex concepts, critical thinking and argumentation.
Programming courses supplied two comparisons. In Chile, 71 industrial engineering students learned web programming with Microsoft Copilot or with instructional videos; both groups improved, and the video group gained more knowledge and showed more curiosity and stronger intention to continue.
Two Peruvian designs tested how students use the tool. A prompting method (role, objective, context, examples and format) taught to 30 of 60 university students raised critical thinking scores, with a median of 10 against 8 in the appropriability dimension (d = 2.042).
Two comparisons qualify the gains. In a Mexican course, 96 students formed two groups; the ChatGPT group obtained better grades on practical than on theoretical tasks, whereas the group without ChatGPT did better in a final exam taken without technological assistance. The ChatGPT group reported higher satisfaction and motivation.
Tutor prototypes were small. A 2 × 2 experiment with 22 engineering students in Mexico crossed voice or text interaction with the presence of an avatar: voice lengthened active time, text produced more messages, and the avatar changed none of the measured variables.(24) A chatbot for secondary students in Lima was described through its cloud architecture and a planned pilot, so no learning outcome was available.(25) In Peru, 53 first-year students who used a conversational agent improved in the social dimension of adaptability (p < 0.001) but not in the institutional one (p = 0.124).(26)
An intelligent tutoring study of systems engineering students at a Peruvian university reported R² values of 0.746 for student emotion and 0.723 for the effect on academic performance.
Feedback on written work was a second use. In Chile, ChatGPT-4 with human validation (a process named RECaP-GPT) produced descriptive feedback on 578 responses to pedagogical cases written by 347 students in seven teacher education programs. The feedback followed criteria known to the students, and the authors listed implementation limits.
Accuracy and prompting set the ceiling of any tutor. ChatGPT scored 8.25 against 7.35 for 110 medical students on four clinical cases in Ecuador (p = 0.0293), and the authors warned about inaccurate information.(30) Among 102 Chilean students writing Spanish prompts, explicit context raised response quality more than stating the purpose alone.(31) For English speaking, a review of Latin American work from 2024 and 2025 described ChatGPT as a supplementary tool for overpopulated classrooms.(32)
Three patterns recur in the tutor axis. Speed and completion improved more consistently than unaided learning, as the Brasilia and Mexican courses show. The largest reported effects came from structured interventions, the prompting method and the 15-session nursing program, with groups of 60 and 100 students. School-level evidence was limited to the Lima chatbot design and the card-based pilot, and the other 14 tutor works involved university students.
None of the 14 game works compared a game built with a generative model against a conventional game. The closest was a workshop in which 24 prospective physics teachers in Ecuador used ChatGPT to design playful activities. Of them, 79% valued its adaptation of activities to learner levels and 83% its reduction of preparation time.
An eight-week quasi-experiment with 60 early-childhood teacher trainees in Ecuador gave the experimental group (n = 30) adaptive AI feedback, educational chatbots, badges, levels and H5P mini-games created with an AI tutor. More than 70% of them increased voluntary participation, sustained attention and collaboration.
Pupils with dyslexia or language disorders (n = 80) who used AI-driven game platforms gained in vocabulary, reading and writing, and overall performance rose 27% after the intervention, while the control group received traditional instruction.
A mixed-methods study of 40 public institutions in Ecuador, with 400 teachers and 800 students, found high teacher acceptance of AI and student motivation toward gamified activities. Technology gaps were most severe in rural areas and the Amazon.
In Brazil, a business budgeting game for accounting students used k-nearest neighbors, random forest and artificial neural networks as opponents, and four semesters of student decisions were compared with those of the algorithms.(39) These are predictive models, not generative ones. A rapid review of seven Portuguese-language articles found that AI could optimize question creation in quiz games and sustain support after assessment through chatbots.(40) Qualitative evidence on barriers came from a systematic review and a focus group in Ecuador: resistance to change, administrative workload, training needs and institutional limits hindered gamification, while AI was seen as a way to build didactic resources in a short time.(41) In Colombia, 78% of 50 accounting students rated AI as highly relevant and 66% valued gamification, and the barriers were lack of training and resistance to change.(42)
Teacher-design work was the closest regional match to the designer role. A training plan in an Ecuadorian school combined ChatGPT for pedagogical text, Curipod for real-time interactive presentations and MagicSchool.ai for differentiated materials and formative assessment. After seven sessions, a quasi-experimental evaluation showed gains in technical appropriation, interactive design capacity, evaluative use of the platforms and willingness to adopt them.
In secondary language and literature classes, a quasi-experiment with control and experimental groups combined project-based learning, universal design for learning, gamification and AI. It reported better intrinsic motivation, reading comprehension and text production, and a 27% rise in communicative performance over traditional techniques.
Teachers used ChatGPT as a content source. Among 86 Brazilian computer science professors, it served to generate exercises, slides, assignments and tests.
The game works split by who designed the activity. Teachers used AI to prepare activities in four works,
Adoption studies explained why the tools spread. Among 792 undergraduates in Peru, Chile, Bolivia, Argentina and Colombia, relative advantage (β = 0.247) and compatibility (β = 0.246) had the largest effects on attitude toward ChatGPT, and attitude predicted intention to use (β = 0.777).
Use varied by group. Among 256 students from six Ecuadorian universities, female and engineering students showed higher adoption, and four user clusters emerged.
Teachers lagged students in some settings. At the University of Córdoba in Colombia, students showed high familiarity and frequent use, while faculty adopted generative AI more cautiously.
Ten writing teachers at a private university in Lima used ChatGPT mainly for the final revision of texts and little for planning, in a rather unreflective manner.
Perceived benefits were similar in language classes. Among 120 EFL students in Ecuador, ChatGPT was seen as helpful for organizing ideas, managing time and improving writing, with concerns about overdependence, lack of personalization and occasional inaccuracies.
Two literature reviews addressed school levels. For secondary education, generative tools served mainly to search for information, draft texts, organize ideas and obtain feedback, and substitution was linked to less verification and more passivity.
At eight Ecuadorian universities, a logistic model explained 20.2% of the variance in plagiarism level (Nagelkerke R²), and trust in ChatGPT-generated information (OR = 1.176) and reporting its use to instructors (OR = 1.223) were associated with higher plagiarism levels. Knowing more AI tools (OR = 0.911) and perceiving that AI improves performance (OR = 0.945) were associated with lower levels.
Among 56 Ecuadorian EFL students, worry about hindering their own writing development exceeded worry about penalties. Submitting chatbot text was judged dishonest, and fewer participants judged machine translation from Spanish dishonest.
Use of generative AI was moderate and mostly supportive in 200 first-semester students in Trujillo, yet dependence was negatively associated with overall critical thinking, more strongly for inference and evaluation.(61) In Lima, more than 250 STEM students showed no academic benefit from frequent AI use, and some showed lower performance, whereas higher technological self-efficacy went with better results.(62) In turn, among 833 Ecuadorian students, emotional (β = 0.413) and cognitive (β = 0.567) dimensions predicted ethical behavior, while frequency of use did not (β = −0.128; p = 0.058).(63)
In the Ceibal program in Uruguay, female teachers and students engaged more with preparation and structured practice tools, and male counterparts dominated exploratory platforms. The authors designed a randomized trial of a chatbot structured for teacher mediation.(64) Qualitative work in Peru recorded concerns about dehumanization, the digital divide and inequity in access.
Nine upper-secondary teachers in Brazil saw gains in teaching quality and student interest, and named plagiarism and weak digital competence as obstacles.
Risk works differ in what they measured: integrity in three, cognition in two, ethical behavior in one, and equity or teacher and institutional constraints in four. Nine of the 10 relied on self-report or interviews, and the Uruguayan study used platform data and classroom observation. None of them tracked the same students over time.
Adoption works converged on perceived benefit, peers and training. Eight of the 10 used surveys or mixed methods, and only one, the lecturer workshops, measured the result of a training. Only one adoption study was set in schools, in Ecuadorian secondary education, and the remaining nine involved universities.
The six reviews gave the widest view. A systematic review of 25 papers covered applications, perceptions and institutional challenges across eight countries.
Across axes, three findings cut through the regional corpus. Tools were adopted before they were evaluated, since surveys outnumbered comparative designs in the game, risk and adoption axes. Gains were reported mostly for completion, motivation and perceived usefulness, and in few studies for unaided learning. Risks were described by the same populations that used the tools, which gives their self-reports weight but no external check.
Regional evidence casts generative AI mainly as tutor and teaching assistant, and much less as game designer. Tutor works were more numerous (16 against 14) and more often experimental, whereas the game axis rested on teacher surveys and bundled interventions. The designer role appears in Latin American classrooms as a teacher task, such as preparing playful activities, interactive materials or exercises,
Opportunities cluster in saved time and in feedback, risks cluster in integrity, dependence and unequal access, and adoption follows perceived benefit.
Concordant evidence on tutoring comes from outside the region. In a randomized trial, students learned significantly more in less time with a custom AI tutor than in an active learning class,
Concordant findings on feedback and dependence are also available. A review of generative chat feedback reported better text structure and cohesion,
Mediation emerges as a shared condition. A review on secondary education concluded that the impact depended more on interaction quality, pedagogical purpose and teacher mediation than on frequency of use,
On games, concordance is partial. Gamification meta-analyses report positive effects,(10,11) and the Ecuadorian trials of gamified AI interventions also reported gains.(34,35) Generative tools were judged ready for use as exercise generators
Several results diverge. In Chile, instructional videos outperformed Copilot in early programming, whereas a custom tutor built on pedagogical best practices outperformed active learning outside the region.(18,69) A plausible explanation is design: a general-purpose assistant answers, and a scaffolded tutor structures learning. Chatbot meta-analyses disagree on duration: one found short interventions stronger and attributed this to novelty,
Avatars diverge too. A meta-analysis recommended human-like avatars for chatbots,
Short-term performance and unaided learning also diverge. The Mexican group that used ChatGPT scored higher on practical tasks but lower on an unaided final exam,
Integrity findings diverge in what predicts misconduct. In Ecuador, broader knowledge of AI tools was associated with less plagiarism,
Where numbers exist, the opportunities are specific. Of the Ecuadorian trainees, 83% reported shorter preparation time;
Four mechanisms are plausible, and each has partial support. Reduced mental effort can lower germane cognitive load without raising skill, as in the Ecuadorian storytelling trial.
Accuracy remains a design constraint. Hallucination is a documented property of language models,(4) and the Ecuadorian clinical case test showed higher scores for ChatGPT than for students while the authors warned of inaccuracies.
Interaction design bears on the usability concerns of the journal. A meta-analysis found that text-based interactions showed the largest chatbot effects,
Language and context matter for tutors. Spanish prompts with explicit context improved response quality,
Equity conditions every finding. A review linked the gender digital divide to education across Latin American countries,(9) gender patterns of tool use appeared in Uruguay,
Game-design evidence from outside the region defines what Latin American trials should measure. Playability, narrative originality, bias and content homogenization are the stated challenges,
Earlier regional reviews covered AI in higher education,
Strengths are limited but real. Queries ran in three languages, the regional coverage is declared by country, and each statement rests on the abstract of a verifiable record with a stable identifier. Counts of designs follow the wording of abstracts, so no study was upgraded to a randomized or controlled design that its abstract did not state.
Limitations are concrete. Only abstracts were read, so design details and effect sizes were not verified in full texts. Double coding of full texts with a formal risk-of-bias tool, such as RoB 2 for randomized trials, would remove this limit.
Timing limits the evidence. Of the 56 core works, 47 were published in 2025, eight in 2024 and one in 2023, so most observations followed the arrival of ChatGPT by a single academic cycle. Cohort follow-up over a full school year, with the same instruments at entry and exit, would show whether novelty effects fade.
Coverage is also uneven. Ecuador and Peru contributed 31 of the 50 primary studies, Argentina and Bolivia appeared only inside one five-country survey, and SciELO, Redalyc and LILACS were not searched directly. Direct searches in those indexes, with the records found reported, would correct the imbalance.
Most game works bundled gamification and AI, only three reported a control group, one used classical algorithms, and 13 of the 14 appeared in journals not matched to the Scopus catalogue used for this review. Some effect sizes were very large in small groups, such as d = 2.042 with 60 students and d = 2.05 with 60 lecturers. Factorial designs that vary the generative component and the game elements separately would isolate each effect, and prespecified outcomes would temper the small-sample estimates.
Screening and coding with an AI assistant also await independent human verification of the matrix. A second coder who rebuilds the counts of Table 1 from the abstracts, and who resolves disagreements by reading the full texts, would settle whether the axis assignments are reproducible.
Practical implications follow. Tutors should be scaffolded and tested against active teaching.
For games, generative drafts of activities should pass a human playtest for playability, originality and bias.
Regional studies place generative AI as a tutor that speeds task completion and feedback, with learning gains that depend on design and that were absent in some randomized comparisons. As a game designer, it appears chiefly in teacher preparation of playful activities, and none of the retrieved works compared it with conventional game design. Plagiarism, dependence and unequal access were the recurring risks, and they call for scaffolded designs, training in verification and trials that separate the generative component from gamification.
No financing
None
Drafting – original draft: Analía Ortigoza, Lucía Gerez.
Writing–review and editing: Analía Ortigoza, Lucía Gerez.