AI Language Tools Are Showing Up in the Most Stable Jobs, Not the Shakiest Ones

survey · Scandinavian Journal of Work Environment & Health · la publicación, 3 jun 2026 · gratis

Los empleos más precarios tienen menos tareas que un modelo de lenguaje podría hacer

El estudio encontró esa relación en Canadá, entre 2021 y 2024, y no puede decir qué pasará con su puesto ni cómo lo usará su empleador.

Versión breve · la versión detallada sigue, unos 6 min

Pregunte al chat de weeklyAI

Pregúnteme por este estudio: a quiénes se estudió, qué encontró y qué no dice.

Las conversaciones se guardan mientras exista weeklyAI, para mejorar la publicación. Se responde en el idioma en que usted escribe.

El estudio, de un vistazo
Quiénes: Ocupaciones canadienses, no trabajadores individuales
Cuántos: 512 ocupaciones
Dónde: Canadá, sin los territorios del norte
Cuándo: 2021 a 2024
Tipo de estudio: análisis de lo que hace la gente (encuesta nacional de fuerza laboral)
Quién lo hizo: Institute for Work & Health, Universidad de Toronto, The Dais y Public Health Ontario
El límite que importa: No dice cómo afecta la tecnología a las personas dentro de cada ocupación.

Exposición a modelos de lenguaje según nivel de precariedad de la ocupación

Precariedad baja0.386puntaje de 0 a 1

Precariedad media0.258puntaje de 0 a 1

Precariedad alta0.26puntaje de 0 a 1

Precariedad muy alta0.205puntaje de 0 a 1

Puntaje promedio de exposición a modelos de lenguaje en cada nivel de precariedad ocupacional, ajustado por género, educación, edad, provincia e industria; mide qué parte de las tareas de una ocupación podría ahorrar tiempo con estas herramientas, no el uso real ni el efecto sobre las personas.

Exposición a modelos de lenguaje: ocupaciones menos precarias frente a las más precarias

Ocupaciones con precariedad bajafrente aOcupaciones con precariedad media

Las de precariedad baja tuvieron exposición significativamente mayor (0,386 frente a 0,258)

Ocupaciones con precariedad bajafrente aOcupaciones con precariedad alta

Las de precariedad baja tuvieron exposición significativamente mayor (0,386 frente a 0,260)

Ocupaciones con precariedad bajafrente aOcupaciones con precariedad muy alta

Las de precariedad baja tuvieron exposición significativamente mayor (0,386 frente a 0,205)

Las ocupaciones menos precarias fueron las más expuestas a los modelos de lenguaje, con un puntaje promedio de 0,386.
Lectura de weeklyAI

Ver · 0:57

Si usted trabaja con horarios que cambian de una semana a otra, con un sueldo bajo, con un contrato temporal o con menos horas de las que quiere, su ocupación es lo que los investigadores llaman precaria.

Un grupo de investigadores de Canadá quiso saber si esas ocupaciones están expuestas a los modelos de lenguaje, las herramientas de inteligencia artificial que redactan, resumen y clasifican textos. Usaron la encuesta nacional de fuerza laboral de Canadá, con unos 100 000 trabajadores por mes, entre 2021 y 2024. El análisis fue por ocupación, no por persona: juntaron 512 ocupaciones.

La respuesta fue clara. Las ocupaciones menos precarias fueron las más expuestas a los modelos de lenguaje, con un puntaje promedio de 0,386. Las de precariedad media, alta y muy alta tuvieron 0,258, 0,260 y 0,205. El puntaje va de 0 a 1 y mide qué parte de las tareas de una ocupación podría ahorrar tiempo con estas herramientas.

Lo mismo pasó al mirar cada condición por separado: los empleos con más contratos temporales, con más horas irregulares y con más trabajo a tiempo parcial involuntario fueron los menos expuestos. La única excepción fue el sueldo bajo, que no mostró diferencia al comparar ocupaciones parecidas en lo demás.

El estudio no dice que los modelos de lenguaje vayan a reemplazar a los trabajadores. El artículo describe un estudio en el que, en la mayoría de los casos, la adopción de sistemas de inteligencia artificial no ha desplazado de forma significativa a los trabajadores y pueden usarse para complementarlos.

Ahora, lo que el estudio no puede decir. No puede decir si sus tareas concretas van a cambiar ni cómo su empleador va a usar estas herramientas. La medida es un cálculo sobre tareas descritas en una base de datos de Estados Unidos, cruzada con ocupaciones canadienses. No puede saber cuánto se usan de verdad esas herramientas en cada ocupación. Tampoco mide contrataciones, despidos, sueldos ni salud. Es una foto de un momento, en Canadá, sin los territorios del norte.

El estudio tampoco dice que la herramienta lo vaya a perjudicar ni a beneficiar. Eso no se midió.

Mire en su trabajo si empiezan a aparecer herramientas que redactan o resumen solas. Pregunte para qué tareas las piensan usar y qué pasa con sus horas.

Qué significa para usted

En su trabajo, fíjese si empiezan a aparecer herramientas que redactan o resumen solas, y pregunte para qué tareas las piensan usar y qué pasa con sus horas. El estudio canadiense no puede decirle si su puesto cambiará ni cómo lo usará su empleador, porque miró ocupaciones enteras y no midió contrataciones, sueldos ni salud. Tome el dato como lo que es: una relación observada en Canadá, no un anuncio sobre su futuro.

Jetha, A., Liao, Q., Smith, P. et al. (2026). Large language model exposure and precarious occupations: Unpacking relationships in the Canadian labor force. Scandinavian Journal of Work, Environment & Health. https://doi.org/10.5271/sjweh.4312

Quién pagó: El estudio fue financiado por la Subvención de Asociación del Consejo de Investigación en Ciencias Sociales y Humanidades de Canadá (#895-2025-1000) que apoyó la Asociación sobre IA y Calidad del Trabajo (PAIQ); los autores son empleados del Instituto para el Trabajo y la Salud, que recibe apoyo del Ministerio de Trabajo, Inmigración, Formación y Desarrollo de Habilidades de Ontario, pero los financiadores no tuvieron ningún papel en el análisis o las conclusiones.

No tome esto como consejo médico profesional.

Los hallazgos de otros estudios que aquí se mencionan los conocemos por este documento, que fue el que leímos; no abrimos cada uno de esos estudios.

Versión detalladaLos pasajes copiados del artículo, las ilustraciones y cada fuente con cuánto leímos de ella · 1239 palabras · unos 6 minLeerla →Cerrar

Las herramientas de lenguaje llegan primero a los empleos estables, según un estudio canadiense

Un análisis de Statistics Canada encontró que las ocupaciones con más horarios irregulares y más empleo a tiempo parcial involuntario son las menos expuestas a los modelos de lenguaje.

Así podría verse · ilustración generada por weeklyAI.watch, no es una fotografía

En Canadá, las ocupaciones con mejores condiciones laborales son también las que tienen más tareas que un modelo de lenguaje podría realizar, según un estudio publicado en 2026 por investigadores del Institute for Work & Health, la Universidad de Toronto, The Dais y Public Health Ontario. El trabajo usó cuatro años de la Encuesta sobre la Fuerza Laboral de Statistics Canada, de 2021 a 2024, y analizó 512 ocupaciones.

Los investigadores midieron dos cosas por separado. Primero, qué tan precaria es cada ocupación, usando cuatro señales que la encuesta recoge: empleo temporal, salario bajo, horarios que varían de semana a semana y trabajo a tiempo parcial que la persona no eligió. Segundo, qué proporción de las tareas de esa ocupación podría hacer un modelo de lenguaje con un ahorro grande de tiempo.

El hallazgo central: las ocupaciones con menos señales de precariedad tenían una exposición media a los modelos de lenguaje de 0,386, en una escala donde 1 significaría que todas las tareas de la ocupación podrían hacerse con estas herramientas. Las ocupaciones con precariedad media, alta y muy alta tenían valores de 0,258, 0,260 y 0,205 respectivamente. El promedio general entre todas las ocupaciones fue de 0,34, con un mínimo de 0 y un máximo de 0,84.

Al mirar cada señal por separado, el patrón se repitió en tres de las cuatro: las ocupaciones con más empleo temporal, más horarios irregulares y más trabajo a tiempo parcial involuntario mostraron menos exposición a los modelos de lenguaje que las ocupaciones con menos de esas señales. La excepción fue el salario bajo: en el análisis que ajusta por otras características de la ocupación, la diferencia desapareció.

Los propios autores advierten sobre los límites. La medida de exposición no dice cómo la tecnología afecta a las personas dentro de cada ocupación, ni mide el uso real de estas herramientas en los lugares de trabajo canadienses. Tampoco captura los avances recientes ni otras tecnologías como la robótica. Y como el análisis es a nivel de ocupación, no de trabajadores, sacar conclusiones sobre una persona concreta a partir de estos resultados sería un error.

Vale la pena entender qué es exactamente un modelo de lenguaje. Se trata de un tipo de inteligencia artificial que usa aprendizaje automático, en particular redes neuronales, para reconocer patrones, contexto y significado en el lenguaje¹. En los lugares de trabajo se usa cada vez más para tareas como sintetizar información, redactar documentos, atender consultas de clientes y resolver problemas cognitivos con reglas claras². Son justamente las tareas que tienden a concentrarse en ocupaciones con mejores condiciones laborales².

Lo que estas herramientas hacen hoy con menos facilidad es el trabajo que exige trato interpersonal intenso o adaptarse a situaciones físicas impredecibles³. Ahí está la pista de por qué los empleos con horarios irregulares aparecen menos expuestos: sus tareas se parecen menos a lo que un modelo de lenguaje puede hacer.

Este resultado no llega solo. El artículo describe un estudio anterior en Estados Unidos que encontró que apenas por debajo del 20% de las ocupaciones tenían la mitad de sus tareas expuestas a estos modelos, y que la exposición era mayor en las ocupaciones mejor pagadas⁴. El artículo describe un estudio sobre países latinoamericanos que encontró el mismo patrón: más exposición entre trabajadores con más educación e ingresos⁵. Y el artículo describe un análisis de trabajadores canadienses entre 2022 y 2025, cuando estas herramientas se multiplicaron, que mostró que las ocupaciones más expuestas a la inteligencia artificial tendían a tener jornadas completas y empleo permanente⁶.

El contexto importa porque la precariedad laboral tiene consecuencias medibles. El artículo describe un análisis de 231,307 reclamos por accidentes laborales en Ontario, Canadá, a lo largo de cuatro años, que encontró que más empleo temporal, salarios bajos, horarios irregulares o trabajo a tiempo parcial involuntario se asociaban con mayor riesgo de lesión en el trabajo⁷. El artículo describe que ese mismo estudio encontró que las ocupaciones con dos o más de esas señales tenían casi el triple de riesgo de lesión o enfermedad⁸. Y el artículo describe investigaciones que documentan que en los países de ingresos altos crece el número de trabajadores en ocupaciones precarias⁹.

Otros trabajos que leímos añaden una pieza que este estudio no examina. Una revisión sistemática que solo pudimos leer en su resumen —no pudimos acceder al texto completo— encontró que, entre 2022 y 2024, las ofertas de empleo para puestos iniciales e intermedios en desarrollo de software y creación de contenido cayeron entre un 14% y un 41% en economías de ingresos altos, aunque los autores advierten que esa cifra no es un promedio combinado sino el rango observado entre estudios distintos¹⁰¹¹. Esa misma revisión encontró indicios de que las economías que dependen de exportar servicios cognitivos enfrentan una disrupción desproporcionada¹².

Un estudio sobre Turquía, que también leímos solo en su resumen, encontró que los efectos sobre el empleo son fuertemente asimétricos por edad: desde 2023, el crecimiento del empleo se debilitó sobre todo para trabajadores al inicio de su carrera en ocupaciones muy expuestas, mientras que no se observaron pérdidas comparables para mayores de treinta años¹³. Ese patrón sugiere, según los autores, que la tecnología opera principalmente por la vía de las contrataciones, desplazando a los menos experimentados mientras complementa a los más experimentados¹⁴.

Hay una corriente que pide cautela con las predicciones catastróficas. Un trabajo que contrasta los pronósticos de pérdida masiva de empleos con las proyecciones de la Oficina de Estadísticas Laborales de Estados Unidos encontró que la evidencia respalda cambios graduales, no una ruptura estructural¹⁵¹⁶. Ese mismo trabajo señala que las afirmaciones sobre la desaparición de casi la mitad de los empleos estadounidenses recibieron poca evaluación crítica¹⁷.

Así lo leemos nosotros. Si las herramientas de lenguaje presionan primero a los empleos mejor pagados y más estables, el conflicto no aparecerá donde solemos buscarlo. No será en los trabajos de horarios irregulares, sino en las oficinas y los empleos profesionales que hasta ahora se sentían a salvo. Ahí es donde esperamos ver negociaciones, resistencias y preguntas sobre qué tareas se automatizan y para qué. Si nos equivocamos, lo sabremos si los empleos precarios empiezan a mostrar la misma exposición que los estables, o si la presión se siente igual en ambos lados.

El estudio ajustó por la proporción de mujeres en cada ocupación, pero no informó ningún resultado sobre la exposición de las mujeres a estas herramientas.

¿Qué puede hacer usted con esto? Mire su propio trabajo y separe las tareas que son pasos repetibles —redactar un resumen, llenar un formato, responder consultas estándar— de las que exigen criterio, trato con gente o adaptarse a lo imprevisto. Esas segundas son las que hoy le dan menos riesgo. Pregunte en su trabajo si hay formación pagada para usar estas herramientas; si no la hay, búsquela por su cuenta o entre compañeros. Y esté atento a si los aumentos de productividad que se atribuyen a la tecnología se reflejan o no en su pago y en sus condiciones.

La pregunta que puede llevar a la próxima reunión sobre cambios en su trabajo no es si la herramienta puede hacer su tarea. Es qué decidió su empleador con lo que la herramienta le permitió ahorrar.

De dónde sale cada dato de contexto, y cuánto leímos de cada documento

Jetha, A., Liao, Q., Smith, P. et al. (2026). Large language model exposure and precarious occupations: Unpacking relationships in the Canadian labor force. Scandinavian Journal of Work, Environment & Health. https://doi.org/10.5271/sjweh.4312 - el artículo del que trata esta nota — el artículo completo — el pasaje: "In this study, we focus specifically on LLM, a specific form of AI, which utilize machine learning techniques, especially deep learning and neural networks, to recognize patterns, context, and meaning in language."
Jetha, A., Liao, Q., Smith, P. et al. (2026). Large language model exposure and precarious occupations: Unpacking relationships in the Canadian labor force. Scandinavian Journal of Work, Environment & Health. https://doi.org/10.5271/sjweh.4312 - el artículo del que trata esta nota — el artículo completo — el pasaje: "LLM are increasingly being introduced in workplaces to perform job tasks such as language and information synthesis, documentation, communication, and rule-based cognitive problem-solving that are most often clustered in occupations with higher quality employment conditions."
Jetha, A., Liao, Q., Smith, P. et al. (2026). Large language model exposure and precarious occupations: Unpacking relationships in the Canadian labor force. Scandinavian Journal of Work, Environment & Health. https://doi.org/10.5271/sjweh.4312 - el artículo del que trata esta nota — el artículo completo — el pasaje: "LLM may currently be less likely to be used to perform job tasks requiring intensive interpersonal or physically unpredictable skills ( 20 )."
Jetha, A., Liao, Q., Smith, P. et al. (2026). Large language model exposure and precarious occupations: Unpacking relationships in the Canadian labor force. Scandinavian Journal of Work, Environment & Health. https://doi.org/10.5271/sjweh.4312 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Eloundou and colleagues ( 18 ) recently found that just under 20% of occupations in the US consisted of job tasks in which 50% are exposed to LLM. Occupational exposure to LLM was highest among occupations that were characterized by higher wages (one indicator of whether an occupation is precarious or not) ( 18 )."
Jetha, A., Liao, Q., Smith, P. et al. (2026). Large language model exposure and precarious occupations: Unpacking relationships in the Canadian labor force. Scandinavian Journal of Work, Environment & Health. https://doi.org/10.5271/sjweh.4312 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Similarly, a study of LLM exposure in Latin American countries found that occupational LLM exposure was highest among workers with greater educational attainment and income ( 22 )."
Jetha, A., Liao, Q., Smith, P. et al. (2026). Large language model exposure and precarious occupations: Unpacking relationships in the Canadian labor force. Scandinavian Journal of Work, Environment & Health. https://doi.org/10.5271/sjweh.4312 - el artículo del que trata esta nota — el artículo completo — el pasaje: "An analysis of Canadian workers using data from 2022–2025 in which there was a rapid proliferation of generative AI tools, showed that occupations with high exposure to AI and where AI would be most complementary to workers were also more likely to be characterized by full-time work hours and permanent work arrangements ( 23 )."
Jetha, A., Liao, Q., Smith, P. et al. (2026). Large language model exposure and precarious occupations: Unpacking relationships in the Canadian labor force. Scandinavian Journal of Work, Environment & Health. https://doi.org/10.5271/sjweh.4312 - el artículo del que trata esta nota — el artículo completo — el pasaje: "A recent analysis of 231 307 lost-time compensation claims in Ontario, Canada, over a four-year period, found that greater exposure to temporary employment, low wages, irregular work hours or involuntary part-time work was associated with a greater workplace injury risk."
Jetha, A., Liao, Q., Smith, P. et al. (2026). Large language model exposure and precarious occupations: Unpacking relationships in the Canadian labor force. Scandinavian Journal of Work, Environment & Health. https://doi.org/10.5271/sjweh.4312 - el artículo del que trata esta nota — el artículo completo — el pasaje: "The study also found that occupations with a high level of exposure to ≥2 dimensions of precarity were associated with a nearly threefold risk of injury or illness when compared to workers in occupations facing less ( 3 ) precarity."
Jetha, A., Liao, Q., Smith, P. et al. (2026). Large language model exposure and precarious occupations: Unpacking relationships in the Canadian labor force. Scandinavian Journal of Work, Environment & Health. https://doi.org/10.5271/sjweh.4312 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Studies conducted in high income countries document that a growing number of workers are employed in occupations which are characterized as precarious ( 5 – 8 )."
Dehouche N. (2026). Creation, validation, obsolescence: observed evidence of AI-driven labor market displacement, 2020–2025. Frontiers in Human Dynamics. https://doi.org/10.3389/fhumd.2026.1815037 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "a 14–41% reduction in postings for entry- and mid-level software development and content-creation roles in high-income economies between 2022 and 2024 (range across individual studies: −14% to −41%; median: −23%)"
Dehouche N. (2026). Creation, validation, obsolescence: observed evidence of AI-driven labor market displacement, 2020–2025. Frontiers in Human Dynamics. https://doi.org/10.3389/fhumd.2026.1815037 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "these figures are not pooled estimates but represent the span observed across non-overlapping study designs and geographies, and should be interpreted as illustrative of the order of magnitude of the effect rather than as a meta-analytic point estimate."
Dehouche N. (2026). Creation, validation, obsolescence: observed evidence of AI-driven labor market displacement, 2020–2025. Frontiers in Human Dynamics. https://doi.org/10.3389/fhumd.2026.1815037 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "preliminary but material evidence that developing economies reliant on cognitive services outsourcing face disproportionate disruption through both direct exposure and indirect demand-erosion channels"
Gürci̇han B, Beyhan B, Akçomak İS. (2026). Labour Market Implications of Generative AI in an Emerging Economy: The Case of Türkiye. Science, Technology and Society. https://doi.org/10.1177/09717218261480259 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Employment effects are strongly age-asymmetric: since 2023, job growth has weakened mainly for early-career workers in highly exposed occupations, while no comparable losses are observed for workers over thirty."
Gürci̇han B, Beyhan B, Akçomak İS. (2026). Labour Market Implications of Generative AI in an Emerging Economy: The Case of Türkiye. Science, Technology and Society. https://doi.org/10.1177/09717218261480259 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "These patterns survive controls for sectoral reallocation, suggesting that generative AI operates primarily through hiring margins, displacing inexperienced workers while complementing experienced ones."
Handel MJ. (2026). Reassessing the Impact of Artificial Intelligence on Employment: Evidence Against the Mass Job Loss Hypothesis. LABOUR. https://doi.org/10.1111/labr.70018 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "The evidence strongly supports BLS projections of gradual change."
Handel MJ. (2026). Reassessing the Impact of Artificial Intelligence on Employment: Evidence Against the Mass Job Loss Hypothesis. LABOUR. https://doi.org/10.1111/labr.70018 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "There is little evidence of mass displacement or any kind of technology‐driven structural break with previous employment trends, arguing for caution in making predictions of the future impacts of robotics and AI on jobs."
Handel MJ. (2026). Reassessing the Impact of Artificial Intelligence on Employment: Evidence Against the Mass Job Loss Hypothesis. LABOUR. https://doi.org/10.1111/labr.70018 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Frey and Osborne's influential study concluded AI and robots might eliminate nearly half of all US jobs between 2010 and 2030. However, these and related claims have received little critical evaluation."

No tome esto como consejo médico profesional.

Los hallazgos de otros estudios que aquí se mencionan los conocemos por este documento, que fue el que leímos; no abrimos cada uno de esos estudios.

survey · Scandinavian Journal of Work Environment & Health · the paper, 3 Jun 2026 · free

AI Language Tools Are Showing Up in the Most Stable Jobs, Not the Shakiest Ones

Canadian researchers looked at 512 occupations and found the tools were linked to the steadiest work. The study cannot say what happens to any one worker.

Short version · the longer version follows, about 6 min

Ask the weeklyAI chat

Ask me about this study: who was studied, what it found, and what it does not say.

Conversations are saved for as long as weeklyAI exists, to improve the publication. Answers come in the language you write in.

The study at a glance
Who: Canadian occupations, not individual workers
How many: 512 occupations
Where: Canada, outside the Territories
When: 2021 to 2024
Kind of study: analysis of what people did
Who did it: Institute for Work & Health, Toronto, and University of Toronto
The limit that matters: It looked at whole occupations, not individual workers, and cannot say what happens to any one worker.

How much of an occupation's work a language model could do, by how shaky the job is

Least shaky jobs0.386score from 0 to 1

Medium shakiness0.258score from 0 to 1

High shakiness0.26score from 0 to 1

Most shaky jobs0.205score from 0 to 1

Average share of an occupation's tasks a language model could do, after accounting for who works in those jobs; the study looked at whole occupations, not individual workers.

Shaky work versus steady work, and how much of it a language model could take on

Occupations with the least shaky conditionsagainstOccupations with the most shaky conditions

The steadiest jobs scored 0.386; the shakiest scored 0.205

Jobs with hours that change week to weekagainstJobs with steady hours

Exposure to language models was significantly lower in the jobs with the most irregular hours

Jobs with temporary contractsagainstJobs with permanent contracts

Exposure to language models was significantly lower in the most temporary jobs

Jobs with involuntary part-time workagainstJobs with the hours workers wanted

Exposure to language models was significantly lower in the most involuntary part-time jobs

Occupations with the least shaky conditions had the highest exposure to these tools.
weeklyAI's reading

Watch · 0:54

If your hours change from week to week, or you work part-time because you cannot get full-time work, the tasks you do are less likely to be the kind that AI language tools can take over. That is what a team of Canadian researchers found.

They used Canada's Labour Force Survey, a national survey of about 100,000 workers each month, collected from 2021 through 2024. The team sorted 512 occupations by four things that make work shaky: temporary contracts, low pay, hours that change week to week, and part-time work when full-time was wanted. They compared those occupations with how many of their tasks could be done by a large language model, the software behind chatbots.

Occupations with the least shaky conditions had the highest exposure to these tools. After the researchers accounted for who works in those jobs, those occupations scored 0.386 on a scale from 0 to 1. Occupations with the most shaky conditions scored 0.205. The middle groups fell between. Low pay was the one shaky condition that did not hold up once the researchers accounted for who works in those jobs.

The article describes a study in which the last big wave of automation hit repetitive, lower-skilled work hardest. Language tools do best with writing, summarizing and routine problem-solving, which cluster in better-paid, steadier jobs.

This does not mean your tasks are safe, or that they will change. The study looked at whole occupations, not individual workers, and it used a measure built from US job descriptions. It cannot show what your employer will do, or what will happen to hiring, pay or hours.

Watch your workplace for AI tools being introduced. Ask what tasks they are meant to handle, and whether your hours or duties will shift.

What this means for you

What this means for you is narrow, because the study looked at whole occupations, not at people, and it cannot say what your employer will do next. Watch whether language tools arrive in your workplace, and ask which tasks they are meant to handle. Nobody measured pay or hours here, so do not conclude that your job is safe or that it is next.

Who paid: The study was funded by the Social Sciences and Humanities Research Council of Canada Partnership Grant (#895-2025-1000) supporting the Partnership on AI and Quality of work (PAIQ); authors are employees of the Institute for Work & Health, which is supported by the Ontario Ministry of Labour, Immigration, Training and Skills Development, but the funders had no role in the analysis or conclusions.

Do not take this as professional medical advice.

The findings of other studies mentioned here are known to us through this document, which is the one we read; we did not open each of those studies.

The longer versionThe passages copied from the paper, the pictures, and every source with how much of it we read · 1161 words · about 6 minRead it →Close

Steady jobs have the most tasks a chatbot can do, Canadian survey finds

Canada's national labour survey, covering 512 occupations over four years, found the steadiest jobs had the most tasks a chatbot could do—and the most precarious had the least.

How it could look · illustration generated by weeklyAI.watch, not a photograph

Occupations with the least precarious working conditions are the ones most exposed to large language models, according to a study of Canadian occupations published in the Scandinavian Journal of Work, Environment & Health. The team, led by Arif Jetha of the Institute for Work & Health in Toronto, pooled four years of Statistics Canada's Labour Force Survey, 2021 to 2024, roughly 100 000 workers in each monthly cycle, and worked at the level of occupations.

The measure of exposure came from an earlier American project, where a team of human labelers went through the tasks one by one and judged where a language model alone or with a simple interface would save half the time. Applied to Canadian occupations, the average exposure score was 0.34 on a scale from 0 to 1, with the highest occupation reaching 0.84. On the combined precarity index, occupations with the least precarity scored 0.386 (95% CI 0.356–0.417), against 0.258 for medium, 0.260 for high and 0.205 for very high precarity.

Precariousness here means four things measured separately: temporary rather than permanent work, low wages, hours that vary from week to week, and part-time work when more hours were wanted. For temporary employment, irregular hours and involuntary part-time work, exposure was significantly lower in the most precarious occupations. Low wages were different: the gap appeared in the raw numbers but vanished once education, age, gender, province and industry were taken into account.

The finding is an association, not a cause. The authors say their measure cannot tell us how workers inside an occupation are affected, cannot capture the newest versions of these tools, and says nothing about other technology such as robotics. It is a snapshot, and it covers Canada outside the Territories.

The stakes attached to precarious work are not small. The article describes a study of 231 307 lost-time compensation claims in Ontario over four years that found that more temporary employment, low wages, irregular hours or involuntary part-time work went with greater workplace injury risk¹. The article describes a study in which occupations loaded on two or more of those dimensions carried nearly three times the risk of injury or illness compared with occupations facing less².

A language model is software trained with machine learning, deep learning and neural networks to recognise patterns, context and meaning in language³. At work it is put to synthesising information, writing documentation, handling communication and rule-based problem-solving—tasks the authors describe as most often clustered in occupations with higher-quality employment conditions⁴. The article describes a study in which tasks needing heavy interpersonal contact or physically unpredictable work are, for now, less likely to be handed to it⁵.

Two earlier studies point the same way. The article describes a study in the United States in which just under 20% of occupations consisted of job tasks in which half are exposed to a language model, and exposure was highest in better-paid occupations⁶. The article describes a study across Latin American countries in which exposure was highest among workers with more education and higher income⁷. The article describes a separate Canadian analysis using 2022–2025 data that found that occupations where AI would complement workers most were also more likely to offer full-time hours and permanent arrangements⁸.

Not every study of this technology agrees on what is happening. One review of observed labour market data since 2020 reported a 14–41% reduction in postings for entry- and mid-level software development and content-creation roles in high-income economies between 2022 and 2024, with the range spanning different study designs and countries rather than a single pooled figure, according to the summary of that study; we could read only the summary, the full paper is behind a subscription⁹¹⁰. The same summary reported a 15–22% wage premium for workers showing AI-augmentation capabilities¹¹, and found infrastructure, security and quality-assurance roles expanding while developer roles contracted¹².

A study of Türkiye using social security records found 17% of total employment in the most AI-exposed fifth of occupations, and job growth weakening since 2023 mainly for early-career workers in those occupations, with no comparable losses for workers over thirty—patterns that held after accounting for shifts between industries, according to its summary; we could read only the summary¹³¹⁴¹⁵. A reassessment of the famous prediction that AI and robots might eliminate nearly half of all US jobs between 2010 and 2030 found the evidence instead supporting gradual change, with little sign of mass displacement or a break with previous employment trends, according to its summary; we could read only the summary¹⁶¹⁷¹⁸.

Here is how we read it. Put the two pictures side by side and the shape is plain. The jobs where a machine can already draft the memo, summarise the case file or answer the routine message are, by and large, the jobs that were already steady, full-time and decently paid. The jobs with the unpredictable weeks are the ones where the work still runs through hands, voices and places. If that holds, a reader on a shifting schedule should expect little of their day to be quietly taken over by a chatbot in the near term—and should expect the pressure, if it comes, to arrive through the schedule, the software that assigns it, and who is offered training rather than through the tasks themselves. You would know we had it wrong if the irregular-hours and involuntary part-time occupations around you started being reorganised around these tools—if your employer began handing the writing, the summaries or the customer replies to software and reshaping your hours to match.

We also read other work that complicates the picture, and two things are worth carrying. The article notes that a worker's social position, such as age or gender, may be linked to working in occupations with different levels of AI exposure and precarity, and that some early research shows younger workers are more likely to start their careers in jobs highly exposed to generative AI. Neither of these changes what the Canadian survey found. Both tell you what to keep an eye on.

So the useful move is not to guess which job disappears. It is to ask a narrower, answerable question about your own workplace: are the new tools being used to take work off your hands, or to measure and divide it? You can ask whether training on them is offered to everyone or only to some, whether the software that sets your hours is the same software that evaluates you, and whether new tasks created alongside the tools come with time and pay attached. Those questions do not need anyone's permission, and the answers are visible where you work long before they show up in any survey.

What would you ask your employer about the software that now decides your week?

Where each piece of context comes from, and how much of it we read

Jetha, A., Liao, Q., Smith, P. et al. (2026). Large language model exposure and precarious occupations: Unpacking relationships in the Canadian labor force. Scandinavian Journal of Work, Environment & Health. https://doi.org/10.5271/sjweh.4312 - the article this story is about — the whole article — the passage: "A recent analysis of 231 307 lost-time compensation claims in Ontario, Canada, over a four-year period, found that greater exposure to temporary employment, low wages, irregular work hours or involuntary part-time work was associated with a greater workplace injury risk."
Jetha, A., Liao, Q., Smith, P. et al. (2026). Large language model exposure and precarious occupations: Unpacking relationships in the Canadian labor force. Scandinavian Journal of Work, Environment & Health. https://doi.org/10.5271/sjweh.4312 - the article this story is about — the whole article — the passage: "The study also found that occupations with a high level of exposure to ≥2 dimensions of precarity were associated with a nearly threefold risk of injury or illness when compared to workers in occupations facing less ( 3 ) precarity."
Jetha, A., Liao, Q., Smith, P. et al. (2026). Large language model exposure and precarious occupations: Unpacking relationships in the Canadian labor force. Scandinavian Journal of Work, Environment & Health. https://doi.org/10.5271/sjweh.4312 - the article this story is about — the whole article — the passage: "In this study, we focus specifically on LLM, a specific form of AI, which utilize machine learning techniques, especially deep learning and neural networks, to recognize patterns, context, and meaning in language."
Jetha, A., Liao, Q., Smith, P. et al. (2026). Large language model exposure and precarious occupations: Unpacking relationships in the Canadian labor force. Scandinavian Journal of Work, Environment & Health. https://doi.org/10.5271/sjweh.4312 - the article this story is about — the whole article — the passage: "LLM are increasingly being introduced in workplaces to perform job tasks such as language and information synthesis, documentation, communication, and rule-based cognitive problem-solving that are most often clustered in occupations with higher quality employment conditions."
Jetha, A., Liao, Q., Smith, P. et al. (2026). Large language model exposure and precarious occupations: Unpacking relationships in the Canadian labor force. Scandinavian Journal of Work, Environment & Health. https://doi.org/10.5271/sjweh.4312 - the article this story is about — the whole article — the passage: "LLM may currently be less likely to be used to perform job tasks requiring intensive interpersonal or physically unpredictable skills ( 20 )."
Jetha, A., Liao, Q., Smith, P. et al. (2026). Large language model exposure and precarious occupations: Unpacking relationships in the Canadian labor force. Scandinavian Journal of Work, Environment & Health. https://doi.org/10.5271/sjweh.4312 - the article this story is about — the whole article — the passage: "Eloundou and colleagues ( 18 ) recently found that just under 20% of occupations in the US consisted of job tasks in which 50% are exposed to LLM. Occupational exposure to LLM was highest among occupations that were characterized by higher wages (one indicator of whether an occupation is precarious or not) ( 18 )."
Jetha, A., Liao, Q., Smith, P. et al. (2026). Large language model exposure and precarious occupations: Unpacking relationships in the Canadian labor force. Scandinavian Journal of Work, Environment & Health. https://doi.org/10.5271/sjweh.4312 - the article this story is about — the whole article — the passage: "Similarly, a study of LLM exposure in Latin American countries found that occupational LLM exposure was highest among workers with greater educational attainment and income ( 22 )."
Jetha, A., Liao, Q., Smith, P. et al. (2026). Large language model exposure and precarious occupations: Unpacking relationships in the Canadian labor force. Scandinavian Journal of Work, Environment & Health. https://doi.org/10.5271/sjweh.4312 - the article this story is about — the whole article — the passage: "An analysis of Canadian workers using data from 2022–2025 in which there was a rapid proliferation of generative AI tools, showed that occupations with high exposure to AI and where AI would be most complementary to workers were also more likely to be characterized by full-time work hours and permanent work arrangements ( 23 )."
Dehouche N. (2026). Creation, validation, obsolescence: observed evidence of AI-driven labor market displacement, 2020–2025. Frontiers in Human Dynamics. https://doi.org/10.3389/fhumd.2026.1815037 — only the abstract - the full text could not be fetched — the passage: "a 14–41% reduction in postings for entry- and mid-level software development and content-creation roles in high-income economies between 2022 and 2024 (range across individual studies: −14% to −41%; median: −23%)"
Dehouche N. (2026). Creation, validation, obsolescence: observed evidence of AI-driven labor market displacement, 2020–2025. Frontiers in Human Dynamics. https://doi.org/10.3389/fhumd.2026.1815037 — only the abstract - the full text could not be fetched — the passage: "these figures are not pooled estimates but represent the span observed across non-overlapping study designs and geographies, and should be interpreted as illustrative of the order of magnitude of the effect rather than as a meta-analytic point estimate."
Dehouche N. (2026). Creation, validation, obsolescence: observed evidence of AI-driven labor market displacement, 2020–2025. Frontiers in Human Dynamics. https://doi.org/10.3389/fhumd.2026.1815037 — only the abstract - the full text could not be fetched — the passage: "a 15%–22% wage premium for workers demonstrating AI-augmentation capabilities"
Dehouche N. (2026). Creation, validation, obsolescence: observed evidence of AI-driven labor market displacement, 2020–2025. Frontiers in Human Dynamics. https://doi.org/10.3389/fhumd.2026.1815037 — only the abstract - the full text could not be fetched — the passage: "heterogeneous sectoral effects, with infrastructure, security, and quality-assurance roles expanding alongside developer role contraction"
Gürci̇han B, Beyhan B, Akçomak İS. (2026). Labour Market Implications of Generative AI in an Emerging Economy: The Case of Türkiye. Science, Technology and Society. https://doi.org/10.1177/09717218261480259 — only the abstract - the full text could not be fetched — the passage: "We find that the share of AI-exposed employment within the top 20% of the exposure distribution (fifth quintile) amounts to 17% of total employment. About 2.3 million registered jobs in Türkiye are highly exposed to generative AI."
Gürci̇han B, Beyhan B, Akçomak İS. (2026). Labour Market Implications of Generative AI in an Emerging Economy: The Case of Türkiye. Science, Technology and Society. https://doi.org/10.1177/09717218261480259 — only the abstract - the full text could not be fetched — the passage: "Employment effects are strongly age-asymmetric: since 2023, job growth has weakened mainly for early-career workers in highly exposed occupations, while no comparable losses are observed for workers over thirty."
Gürci̇han B, Beyhan B, Akçomak İS. (2026). Labour Market Implications of Generative AI in an Emerging Economy: The Case of Türkiye. Science, Technology and Society. https://doi.org/10.1177/09717218261480259 — only the abstract - the full text could not be fetched — the passage: "These patterns survive controls for sectoral reallocation, suggesting that generative AI operates primarily through hiring margins, displacing inexperienced workers while complementing experienced ones."
Handel MJ. (2026). Reassessing the Impact of Artificial Intelligence on Employment: Evidence Against the Mass Job Loss Hypothesis. LABOUR. https://doi.org/10.1111/labr.70018 — only the abstract - the full text could not be fetched — the passage: "Frey and Osborne's influential study concluded AI and robots might eliminate nearly half of all US jobs between 2010 and 2030. However, these and related claims have received little critical evaluation."
Handel MJ. (2026). Reassessing the Impact of Artificial Intelligence on Employment: Evidence Against the Mass Job Loss Hypothesis. LABOUR. https://doi.org/10.1111/labr.70018 — only the abstract - the full text could not be fetched — the passage: "The evidence strongly supports BLS projections of gradual change."
Handel MJ. (2026). Reassessing the Impact of Artificial Intelligence on Employment: Evidence Against the Mass Job Loss Hypothesis. LABOUR. https://doi.org/10.1111/labr.70018 — only the abstract - the full text could not be fetched — the passage: "There is little evidence of mass displacement or any kind of technology‐driven structural break with previous employment trends, arguing for caution in making predictions of the future impacts of robotics and AI on jobs."

Do not take this as professional medical advice.

The findings of other studies mentioned here are known to us through this document, which is the one we read; we did not open each of those studies.