Teaching Tech in the AI Era

What I’m learning across communities of practice and the university classroom

Yanina Bellini Saibene

2026-09-17

Where This Talk Comes From

1. A bit of framing about how GenAI shows up in education
2. How widespread this already is, some regional numbers
3. A Shared Vocabulary, borrowing a few terms so we can talk precisely
4. Classroom experience, what I see and what I do
5. What this means if your format is a one- or two-day workshop

A bit of framing

Arrival vs. Adoption Technologies

• Adoption technologies enter the educational scene after a process of planning.

• Arrival technologies, by contrast, find their way into the classroom and disrupt our activities. Their adoption is not planned.

GenAI is an arrival tecnology

What GenAI Makes Us Rethink

1. Curricular dimension: Rethinking what we should teach and what learners should learn

2. Pedagogical dimension: Rethinking how we teach and how learners learn

• Policy and assessment dimension: deciding what guidelines and rules to establish, particularly around assessing competencies

1 and 2 directly concern us at The Carpentries

Where I Teach

Space 1 — A Long Curriculum, Taught in Short Training Format

rOpenSci Champions Program

• A year-long mentorship program: R users become R package developers
• Short, focused training sessions embedded inside a long-format curriculum
• Participants are already invested in open source, they apply to be part of the program. They have chosen to be here.

The program has a small cohort of mentees paired with mentors over roughly a year, working toward actually shipping an R package, with periodic synchronous sessions covering specific skills (testing, documentation, git workflows, etc.). So the “teaching” moments are short, but they sit inside a much longer arc of skill-building and community integration. This matters because the curriculum is long, and we have room to build habits, not just deliver information once.

Space 2 — Many Short-Format Communities

The Carpentries · Institute for Research Software · ENANDES+

• Short-format training for scientists. This is not a captive audience.
• If you are instructor/trainer for The Carpentries this is likely your space too
• One or two days, sometimes a single session — very little room for a semester-long arc

Space 3 — The University Classroom

Undergraduate and graduate education - Austral, UCES, UCA

• New students encountering programming for the first time.

• Semester-long, mandatory, graded

• Professional continuing education — e.g., agriculture professionals in diploma programs

• Semester or year long, (graded)

How Widespread Is This, Already?

This Is Already the Norm, Not the Exception

Numbers for Latin America (2026)

Students

• 92% have used AI in their learning. Latin America, above the global rate
• 67% use it daily or weekly
• 65% use it as search engine

Faculty

• 79% have used AI in their teaching, also above the global rate
• 94% expect to use it in future practice

These numbers have increase from the previous survey

Source: Digital Education Council, Tecnológico de Monterrey, & Institute for the Future of Education (2026). Encuesta sobre la IA en la Educación Superior en América Latina 2026. digitaleducationcouncil.com

The survey has 30,000 responses, including 22,941 from students and 7,319 from faculty, across 29 higher education institutions in Latin America.

ChatGPT dominates tool choice at 88% of students, Google Gemini a distant second at 50%.

And It Isn’t Uncritical Adoption

65% of students worry AI use makes learning too shallow and discourages critical thinking

76% of faculty worry about students becoming excessively dependent on AI

This is the tension I live in every semester. The rest of this talk are some experience on how that tension looks like in my teaching.

Source: Digital Education Council, Tecnológico de Monterrey, & Institute for the Future of Education (2026). Encuesta sobre la IA en la Educación Superior en América Latina 2026. digitaleducationcouncil.com

Adoption being high does NOT mean people are comfortable with it. Students themselves are the ones flagging the shallow-learning risk, not just faculty.

A Shared Vocabulary

• The comprehension–completion trade-off: AI can finish a task without the student developing the skill the task was meant to build. Task completion goes up; independent skill often doesn’t follow.

• Pseudo-apprenticeship: when learners appear to practice computational skills, but AI performs the core problem-solving, coding, or debugging work, limiting opportunities to develop those skills themselves.

• Prior knowledge as moderator: The size and direction of AI benefit depend on the learner’s prior knowledge.

Arun Kumar, Thakerng Wongsirichot, Aziz Nanthaamornphong, Generative AI in computing education: A systematic review and a framework for responsible integration,Computers and Education: Artificial Intelligence,Volume 11,2026, 100677, ISSN 2666-920X,https://doi.org/10.1016/j.caeai.2026.100677.

Four Ways (Begginers) Students Use GenAI to Learn to Code

• Help-seeking displacement: ask the AI instead of a peer, an instructor, the documentation

• Passive acceptance: take the first suggestion, don’t read it

• Verification avoidance: skip testing the output; assume it’s right

• Prompt dependency: can’t start a task without AI support

Instructors report they experience some of this ways in The Carpentries workshops.

Sources: Kazemitabaar, M., Hou, X., Henley, A., Ericson, B. J., Weintrop, D., & Grossman, T. (2024). How novices use LLM-based code generators to solve CS1 coding tasks in a self-paced learning environment. Koli Calling 2024. https://doi.org/10.1145/3631802.3631806

Hou, I., Man, O., Hamilton, B., Muthukrishnan, R., Johnykutty, S., Zadeh, L., & MacNeil, S. (2025). “All roads lead to ChatGPT”: How programming students and instructors perceive the impact of GenAI. ITiCSE 2025, 79–85. https://doi.org/10.1145/3724363.3729024

Experience 1: Short format training

• Exercise: in breakouts rooms modify a live-coded Google Earth Engine script to use different imagery, accept a polygon (instead of a set of coordinates) and calculate NDVI
• I live-code in JavaScript. The exercise never specifies a language.
• Students ask an LLM for help for the “accept a polygon” part, it defaults to Python
• They paste that back to the AI a second time to translate it into JavaScript

I live-code the GEE script in JavaScript because that’s the GEE code editor’s native language; the exercise itself never says “in JavaScript”; students ask their AI tool of choice for help, which defaults to Python because that’s statistically the dominant language in AI training data and in agricultural/data-science contexts generally; students realize the mismatch, and, rather than adapting on their own, they paste the Python solution back into the AI a second time asking for a JavaScript translation.

This is help-seeking displacement and a mild form of prompt dependency happening simultaneously, and it reveals something about default assumptions baked into these tools that even the instructor hadn’t anticipated needing to specify. It’s also a reminder that “which language” is the kind of tacit knowledge an experienced live-coder doesn’t think to state explicitly, but that an AI tool will confidently resolve one way regardless of what’s actually expected.

Experience 2: R Package Developers and Software Peer-Review

First: build foundational competence

• Collaborative software development. R package development. Software review. Community management & outreach.

Then: Introduce AI in context

• AI in software review: good practices for submitting AI-assisted software. rOpenSci policies and recomendations.

Finally: make AI an explicit topic

• Understanding AI and limitations. Using AI for computational work.

Graduated and scaffolded AI access: AI support is introduced as competence develops, not before.

Sources: Padgham, M., Ross, N., Salmon, M., Bellini Saibene, Y., Lepore, M., Riederer, E., Helske, J., & Rodriguez-Sanchez, F. (2026, February 26). rOpenSci’s AI Policy for Software Peer Review. rOpenSci blog. https://doi.org/10.59350/5tavw-mke71

Experience 3: Computer Programming

Our initial guidance

For code

• Think about the problem, before start coding
• Understand the code before running it
• Use AI for small, single-function tasks
• Use it to debug, not to originate

We also make AI’s limitations visible

We examine examples of AI-generated errors, particularly in data analysis, where students need to learn to question, verify, and validate computational results.

And yet…

Some students use AI as a crutch rather than a scaffold. They use it as a shortcut—offloading the very work they need to learn.

Experience 3: Ask Them Their Goals First

Activity before any content: “What are YOUR goals for this course?”

• Every student writes 2–5 goals for the course
• No right or wrong answers: the point is knowing what they expect to learn
• Submitted anonymously, before the syllabus is even fully presented

Exercise becomes the anchor for an explicit conversation about AI and how the use can undermine or help their goals.

The shift

Students are coming to ask questions because the AI solution feels wrong, even though it works.

• They disclose AI use,
• and they’re suspicious of what it handed them.

What This Means

Format Changes What’s Possible

• Semester course: room for graduated access, scaffold withdrawal.
• Long community of practice (rOpenSci Champions): room for norms and culture to form over a year
• Short workshop (1–2 days): almost none of that. The intervention has to fit in the short-term workshop.

The literature’s strongest recommended interventions: graduated access, scaffold withdrawal, Self-Regulated Learning scaffolding, assume time you may not have.

Almost everything in the literature that “works” — progressively reducing AI access as competence grows, structured multi-week SRL interventions, semester-long redesigned assessment — assumes a time horizon that a two-day Carpentries workshop simply does not have.

For Instructors Like Us

What’s Realistic in a Day or Two

• Build a verification habit in a single exercise: ask “would you have known if this were wrong?”

• Use the learners motivation to set the tone of the workshop.

Experiment with assessments such as:

• Explain this code
• Find the bug
• Predict the output before running it
• Write tests for this function
• Critique two AI-generated solutions
• Improve an AI-generated solution
• Explain why the AI’s answer is wrong
• Modify the code when the requirements change
• Debug unfamiliar code
• Use AI, but document and evaluate how you used it

Back to You

• What are you seeing in your own workshops ?
• Where’s the equivalent of my “office hours” moment in a format that doesn’t have office hours?
• What have you tried as formative assessment ?

Thank you :-)