Do language models always answer the same way?

In this lesson, we explore a core characteristic of language models: their non-deterministic nature. When asked the same question, they can generate different responses. Why does this happen? Because they do not operate on fixed rules, but rather on probabilities built from vast amounts of data. We analyze how variables such as "temperature" influence the degree of creativity or predictability in responses, and how certain parameters can modify a model's behavior without requiring retraining.

We also examine the so-called system prompt, that invisible initial instruction that guides tone, boundaries, and interaction rules. What are the implications of configurations hidden from the user? To what extent are these systems truly "black boxes"? The goal is to reflect on transparency, auditability, and trust: if behavior can be adjusted without our knowledge, what regulatory and pedagogical challenges arise?

At the end of the lesson, you are invited to answer a few questions. 

👨‍🏫 Instructor: Marcos Gómez

Alignment: Balancing Accuracy and Responsibility

In this lesson, we work with the concept of alignment, understood as the process of fine-tuning artificial intelligence model responses so they respect human values, ethical standards, and specific goals. In the educational sphere, alignment aims to ensure safe and appropriate responses for the school context. An unaligned model might provide risky instructions when given an seemingly innocent prompt, whereas an aligned model prioritizes safety and adapts its output to the student's environment. One of the most widely used techniques for aligning models is Reinforcement Learning from Human Feedback (RLHF), where human evaluators rate different model responses, which are then used to teach the system which types of answers to prioritize.

We also analyze the challenges of alignment. The "correct" response depends on the user's context, age, and culture. A model can fail if it ignores cultural differences, automatically rejects sensitive topics without contextualizing them, or answers from a perspective centered on another country. Cultural alignment requires systems to understand local knowledge and references—such as traditions, holidays, or specific vocabulary—rather than relying solely on a limited global subset. Without research and culturally diverse data, models tend to make errors or oversimplifications. Thus, advancing alignment is not just a technical challenge, but a cultural and pedagogical one as well.

At the end of the lesson, you are invited to answer a few questions. 

👩‍🏫 Instructor: Sofia Martinelli

Algorithmic Sycophancy

In this lesson, we address key concepts related to AI design that allow us to reflect on how these systems operate and what kind of relationship we build with them. Algorithmic sycophancy, transparency, concealment of uncertainty, and anthropomorphism are some of the defining features of language models. Understanding them enables us to critically evaluate how we use these technologies and consider the socio-educational implications that arise when these features remain invisible.

We also reflect on how sycophancy can impact pedagogical relationships and trust-building. Models adopt an approachable tone, offer constant availability, and can foster feelings of closeness or trust during interactions, which may influence the teacher–student–AI dynamic. We analyze varying degrees of dependency—ranging from critical and reflective usage to a concerning substitution of human connections—and warn against using AI for emotional support without professional mediation. This lesson invites us to question not only what AI answers, but how it answers and what repercussions it may have on the autonomy and critical judgment of its users.

‼️ Important: In this lesson there is an assignment to complete before watching the lecture. It is important that you carry it out; while this preliminary activity is ungraded, completing it is a requirement to pass the course. At the end of the lesson, you will be asked to answer a few questions. These final answers are graded, and completing them is required to finish and earn credit for the course.

👩‍🏫 Instructor: Emilia Echeveste