💻 computer science

Data Annotations as Pedagogical Hints: From Subjective Labels to Critical Thinking

This study demonstrates that incorporating manual data annotation tasks into machine learning courses effectively shifts students from viewing data labels as objective facts to understanding them as subjective interpretations, thereby fostering critical thinking about bias and model behavior, though future iterations must better frame interpretive disagreement as a pedagogical feature and mitigate emotional discomfort from sensitive content.

Original authors: Ralf Raumanns, Theresa Elstner, Louis Ferger-Andrews, Louise M. Carlsen, Martin Potthast, Gerard Schouten, Josien P. W. Pluim, Veronika Cheplygina

Published 2026-07-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ralf Raumanns, Theresa Elstner, Louis Ferger-Andrews, Louise M. Carlsen, Martin Potthast, Gerard Schouten, Josien P. W. Pluim, Veronika Cheplygina

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Invisible Hand Behind the Robot's Brain

Imagine you are teaching a robot to recognize different types of fruit. You show it a thousand pictures of apples and oranges, and you tell the robot, "This is an apple," or "This is an orange." In the world of machine learning, this process of labeling pictures is called data annotation. For a long time, students learning how to build these robots were handed pre-labeled datasets, like a textbook with all the answers already filled in. They learned how to train the robot, but they never saw the messy, human work of deciding what the pictures actually were.

The big question this research tackles is: What happens when the "truth" isn't actually black and white? In many real-world situations, two smart people can look at the same image and honestly disagree on what they see. If a robot learns from data where humans argue, does the robot become confused? Or does it learn something deeper? This paper explores a bold idea: instead of just giving students pre-labeled data, what if we made them do the labeling themselves? By forcing students to become the "human brain" behind the robot, the researchers wanted to see if this would help them understand that AI isn't just math—it's also a reflection of human opinion, bias, and the messy reality of how we see the world.

The Experiment: Labeling Hair on Skin Lesions

The researchers, a team from universities in the Netherlands, Germany, and Denmark, decided to test this idea in two different classrooms. They didn't use simple pictures like cats or dogs, because those are usually easy to agree on. Instead, they gave students hundreds of medical images showing skin lesions (spots on the skin) and asked them to do a very specific, tricky task: count the amount of hair covering the spot.

The students had to decide for each image: Is there no hair, some hair, or a lot of hair?

This sounds simple, but it's a trap for the human brain. Unlike a cat that is clearly a cat, hair on a skin lesion is vague. Is a few stray hairs "some" or "a lot"? Is a patch of thin hair "some"? One student might say "some," while their partner looks at the exact same picture and says "a lot." There is no single "correct" answer written in the stars. The researchers wanted to see what would happen when 43 students (30 from Fontys University and 13 from the IT University of Copenhagen) got stuck in this gray area of disagreement.

What They Found: The "Aha!" Moment and the "Fix It" Reflex

The results were a mix of great learning and a funny human reaction.

The Good News: The Lightbulb Went On
Before the task, many students thought data was just facts waiting to be collected. After doing the labeling, their understanding changed dramatically.

Subjectivity: They realized that their own personal opinion actually shaped the data. They saw that two smart people could look at the same thing and see different things.
Bias: They understood that if they all agreed to call the hair "a lot," the robot would learn that "a lot" is the truth. But if they disagreed, the robot would get confused. They learned that the robot's mistakes often start with the humans who labeled the data, not just the code itself.
Effectiveness: The students rated this messy, hands-on activity as much better than a boring lecture for understanding how bias works. They felt more motivated to learn about AI because they had "felt" the problem themselves.

The Twist: The "Fix It" Reflex
Here is where it gets interesting. Even though the students learned that disagreement is normal and part of the process, their first instinct when asked how to improve the task was to stop the disagreement.

When the researchers asked, "How can we make this better?" the most common answer was, "Give us clearer rules so we all agree!" or "Show us examples so we don't get confused."

The researchers call this a "pedagogical tension." The students had discovered that the world is messy, but they still wanted a clean, perfect answer. They treated the disagreement as a "bug" (a mistake to be fixed) rather than a "feature" (a natural part of how humans see things). They wanted to eliminate the very thing that made them learn.

The Bumps in the Road

It wasn't all smooth sailing. The researchers noted a few specific hurdles:

The "Gross" Factor: A lot of students felt uncomfortable looking at the medical images of skin lesions. It was a bit gross, and for some, that discomfort distracted them from the learning.
The Boredom: Labeling 100 images one by one was repetitive. Students called it "boring" and "slow." They wanted to spend more time discussing why they disagreed and less time clicking buttons.
The Tool Trouble: Some students struggled with the software used for labeling, wishing for simpler tools or pre-made scripts so they could focus on the thinking part.

The Big Takeaway

The paper suggests that making students do the labeling is a powerful way to teach them that AI is not objective. It shows them that the data feeding the robot is built by humans, with all our biases and disagreements.

However, the researchers warn that this method needs careful handling. If you just throw students into a messy task, they might just get frustrated and want to "fix" the mess instead of learning from it. The key is to guide them to realize that disagreement is the lesson, not the error.

In short, the paper proves that getting your hands dirty with data helps you understand the robot's brain better. But it also shows that we have to teach students to love the mess, because in the real world, there is no single "correct" label—only different human perspectives.

Technical Summary: Data Annotations as Pedagogical Hints

Problem Statement
Machine learning education frequently relies on pre-labeled datasets, rendering the inherent subjectivity of human annotation invisible to students. This pedagogical approach risks cultivating a "false sense of objectivity," where learners assume data labels represent a single ground truth rather than the product of human judgment. Consequently, students may fail to appreciate the interpretive diversity and contestability of algorithmic outputs, potentially leading to an overly trusting and harmful perspective on AI bias and fairness. While research has shifted toward viewing annotator disagreement as meaningful signal rather than noise (the "disagreement paradigm"), educational practices largely remain rooted in the "agreement paradigm," treating disagreement as error to be minimized.

Methodology
The authors conducted a case study across two institutions: Fontys University of Applied Sciences (Netherlands, bachelor's level) and IT University of Copenhagen (Denmark, master's level). The intervention involved integrating a manual data annotation task into existing data science courses.

Task Design: Students annotated skin lesion images for "hair coverage" using a 3-point ordinal scale (none, some, a lot). This task was chosen for its inherent subjectivity; unlike objective classification (e.g., cat vs. dog), judging hair density requires interpreting visual features like thickness and density, leading to natural disagreement.
Procedure: The intervention consisted of two parts:
1. Individual Assignment: Students explored an existing ISIC archive dataset via Jupyter notebooks to understand annotation variability and agreement metrics.
2. Group Assignment: Groups of students used Label Studio to annotate 100–200 images. They were required to document decisions and compare labels to analyze inter-annotator disagreement.
Data Collection: A survey was administered to 43 students post-activity. It included:
- Retrospective pre-/post-assessments of familiarity with four core concepts (subjectivity, agreement measures, bias, data quality).
- Likert-scale items measuring five constructs: Annotation Ambiguity, Data Quality, Bias and Fairness, Implementation Barriers, and Pedagogical Effectiveness.
- Open-ended questions for qualitative thematic analysis.
Analysis: Quantitative data were analyzed using descriptive statistics and Fleiss' $\kappa$ to measure actual inter-annotator agreement. Qualitative data underwent thematic analysis (Braun & Clarke framework) to identify patterns in student reflections.

Key Contributions

Educational Intervention: A practical framework using skin lesion annotation to teach responsible AI concepts, specifically the relationship between human judgment and model behavior.
Pedagogical Insights: Empirical evidence from two institutions demonstrating that manual annotation serves as an effective "pedagogical hint," shifting student understanding from a traditional "single ground truth" mindset to a "subjectivity-aware" mindset.
Design Recommendations: Practical takeaways for designing annotation-based assignments that connect human judgment to model behavior, emphasizing the need to frame disagreement as a learning opportunity rather than a defect.

Results

Familiarity Gains: Self-reported familiarity with subjectivity, bias, and data quality increased substantially. No students remained at the lowest familiarity levels post-intervention.
Perception of Subjectivity: Most students recognized that personal interpretation affects annotations and that creating training data is challenging. They linked visual features (e.g., hair) to potential model misclassifications.
Pedagogical Effectiveness: Students rated the annotation activity as more effective than traditional lectures for understanding bias. The activity increased motivation to learn AI concepts.
The "Consensus Illusion": A critical finding emerged where students reported high levels of group consensus in surveys, despite objective Fleiss' $\kappa$ scores indicating only moderate agreement (mean $\kappa \approx 0.62$ ). This suggests students perceive convergence as success, failing to recognize that remaining disagreement reflects domain complexity.
Student Reaction to Disagreement: While students acknowledged subjectivity, their primary suggestion for improvement was to "make annotations more consistent" through clearer guidelines or calibration. This indicates they view disagreement as a problem to be solved (a bug) rather than an inherent feature of the task.
Barriers: The main drawbacks reported were emotional discomfort from viewing medical images and the repetitiveness of the labeling process.

Significance and Claims
The paper argues that manual data annotation is a vital pedagogical tool for demystifying the "black box" of data preparation. It demonstrates that engaging students in the annotation process helps them internalize that human judgment shapes model behavior and that disagreement reflects domain complexity, not merely noise.

However, the authors maintain a modest stance regarding the immediate internalization of these concepts. They note a "productive pedagogical tension": while students gain awareness of subjectivity, they often instinctively seek to eliminate it through better guidelines. The paper concludes that the struggle with inconsistency is the lesson itself. Therefore, the significance lies not just in the activity's ability to generate data, but in its capacity to expose the gap between the "agreement paradigm" (common in education) and the "disagreement paradigm" (common in modern research). The authors suggest that future iterations must explicitly frame disagreement as a feature, not a bug, and mitigate emotional and technical barriers to ensure the learning objectives are met.