“Break the AI: Investigating Complex Image Classification Errors”

Main Idea

Building on the previous scenario, where students trained an AI model and explored why it can make incorrect predictions, this activity deepens their understanding of how Artificial Intelligence interprets visual information. Using cultural heritage images (such as statues, pottery and temples) as a working context, students further investigate how AI models classify images based on patterns in the training data.

Through testing the model with new and unexpected images, students observe how AI systems can misinterpret objects when they rely only on visual similarities. For example, an AI model trained with museum sculpture images may classify a store mannequin as an ancient statue because both share a human-like shape and smooth surfaces. Likewise, a rock formation may be labeled as a temple because vertical stone patterns resemble architectural columns.

These examples help students understand that AI systems do not truly interpret meaning or historical context; they identify patterns in data. By analysing these limitations, students develop a more critical understanding of how AI works and why human supervision and careful dataset design are essential when using AI tools in real-world contexts.

Creator Aspasia Michanou
Subject Digital competence / ICT, History
Length 90′  minutes
Pedagogical Approach Inquiry-based learning, collaborative learning, participatory experimentation
Competences AI use cases, Critical thinking, Responsible AI use, Tool validation, Visualization of Datasets
Grades Students in 8th -11th grade
Software & Materials Google Teachable Machine (Model trained from scenario 1), heritage images (open-licensed by Europeana and Google Arts and Culture), additional images (rocks, everyday objects, mannequins etc.), 6 laptops/tabs, projector, worksheet.
Evaluation Tools

Entry ticket, exit ticket, observation checklist, and a short rubric assessing:

  • interpretation of AI errors
  • quality of group explanations
  • understanding of visual patterns vs meaning
   

Learning Activities (LA) – 90 minutes total

Click to enlarge

LA1 (10 min) Entry ticket

Teacher asks: “Are all AI mistakes equally easy to explain?”

Students choose (Yes/No/Not sure).

Teacher explains that today they will investigate harder AI mistakes, where the answer is not obvious at first sight.

LA2 (10–15 min) Revisiting the model

Students reopen the model from the previous lesson and briefly recall:

  • what categories they used (statues, pottery, temples, other objects)
  • what kinds of mistakes they already observed

Teacher explains that the class will now test AI with more challenging images.

LA3 (20 min) AI Challenge cases

(Worksheet: Steps 1+2)

Students form 6 groups. Each group tests the model with selected challenge images.

Core cases include:

  • mannequin → statue?
  • rock formation → temple?

Students record the model’s predictions.

LA4 (15–20 min) Group Investigation

Students discuss each confusing case.

Teacher guides the discussion with questions such as:

  • What visual features does the mannequin share with a statue?
  • What visual features does the rock share with a temple?
  • Is the AI understanding the object or only recognizing visual patterns?

Expected ideas:

  • mannequin → human body shape, upright posture, smooth surface
  • rock formation → vertical stone lines, symmetry, column-like structure

Students fill in Step 3 of their Worksheet.

LA5 (20 min) Whole-class discussion: simple vs complex errors

Groups present one case each. Teacher helps students distinguish between:

a) Simple errors
The model fails because it has not seen enough examples.

b) Complex errors
The model identifies a misleading visual similarity and treats it as meaningful.

Teacher explains:

  • AI does not understand “ancient,” “museum object,” or “temple” as humans do
  • it detects patterns, not historical or cultural meaning
  • this is why human interpretation remains necessary.

LA6 (10 min) Exit ticket: Reflection

Students answer Step 4 of their Worksheet.