WBCHSE Applied AI common mistakes
WBCHSE Applied AI · Understand the traps and how to avoid them.
Defining AI only as 'machines that think like humans' and then using AI, machine learning, and deep learning as interchangeable labels in short-answer questions.
Candidates memorise a slogan from introductions and never build a hierarchy, so they cannot place a given example such as spam filtering or a chatbot at the correct level.
What to do instead
Write a three-level definition table: AI as the broad goal, ML as systems that learn patterns from data, and DL as ML using multi-layer neural networks. For each level, add two examples and one non-example, then practise classifying new examples aloud.
Answering an ethics case by naming a single issue such as 'bias' or 'privacy' without identifying the affected stakeholder, the concrete harm, the cause, and a realistic mitigation.
Ethics questions are treated as moral opinion rather than structured analysis, so candidates produce labels instead of reasoning.
What to do instead
Use a four-column framework for every case: stakeholder, harm, cause, mitigation. For example, in a hiring model, the stakeholder is a rejected applicant, the harm is lost opportunity, the cause is biased training data, and the mitigation is auditing data and testing outcomes across groups.
Writing Python loops with an off-by-one range, such as using range(1, n) when the task requires values 1 through n inclusive, or mutating a list while iterating over it.
Natural-language instructions are inclusive while Python's range stop value is exclusive, and list mutation changes indices during iteration.
What to do instead
Trace the loop with n=1 and n=3 before running it, use range(1, n + 1) for inclusive counting, and iterate over a copy such as for item in items[:] when removing items. Add a one-line assertion or print of the first and last values.
Loading a CSV and immediately plotting or training on it without checking shape, data types, missing values, duplicates, or impossible values such as negative ages.
Candidates equate 'the file opened' with 'the data is ready' and skip the audit step under time pressure.
What to do instead
Run a fixed audit sequence first: shape, head, info, describe, missing-value counts, duplicate count, and range checks for key columns. Record one decision per problem column, such as drop, impute, or correct, before any chart or model.
Reporting only training accuracy for a classifier, especially on an imbalanced dataset, and concluding that the model performs well.
High accuracy looks convincing, and candidates often forget that a majority-class predictor can score well while detecting nothing useful.
What to do instead
Split data into training and test sets before fitting, use stratification for imbalanced classes, and report a confusion matrix with precision, recall, and F1. State the baseline accuracy from always predicting the majority class beside the model result.
Mixing up NLP and computer-vision tasks, for example describing image classification as sentiment analysis, or saying a vision model 'understands' a scene exactly as a human does.
Everyday media language blurs task names, inputs, and outputs, and candidates revise application examples without mapping each to its data type.
What to do instead
Build a task-output table: text input to tokenisation, POS tags, named entities, or sentiment; image input to classification, detection, or segmentation. For each row, write the input type, output format, and one limitation, then quiz yourself by covering the output column.
Giving domain answers such as 'AI is used in healthcare' or 'AI helps agriculture' without naming the input data, task, output, metric, and domain constraint.
Generic examples are easy to memorise and appear sufficient, but they do not demonstrate applied understanding of a use case.
What to do instead
Write each example as a use-case card: problem, data source, AI task, output, success metric, and one constraint such as privacy, connectivity, or cost. For healthcare, specify medical images, classification task, diagnosis support output, recall metric, and radiologist oversight.
Listing tools such as ChatGPT, TensorFlow, or Canva as if they are equivalent, without stating each tool's purpose, input, output, and limitation; or claiming a generative chatbot is a reliable factual database.
Tool names are familiar from advertising, but their categories and limits are rarely compared explicitly.
What to do instead
Create a tool-purpose table with columns for category, typical input, typical output, and one limitation. Separate model libraries, application products, and generative assistants, and note that generated text must be verified against authoritative sources.
Submitting a notebook whose cells were run out of order, so the displayed result cannot be reproduced by restarting the kernel and running all cells.
Interactive notebooks encourage piecemeal execution, and hidden state makes a result look correct even when earlier cells were changed or skipped.
What to do instead
Before submission, restart the kernel and run all cells from top to bottom. Use relative file paths, set random seeds where relevant, remove unused cells, and save a clean output. If a result changes, fix the dependency rather than rerunning selected cells.
Presenting a project as a code dump or an entrepreneurship pitch as 'AI will do everything', with no problem statement, data provenance, evaluation, limitations, users, or cost model.
Candidates focus on the artefact they built and omit the reasoning that makes it credible to a teacher, evaluator, or customer.
What to do instead
Use a one-page project canvas: problem, target user, data source and permission, method, evaluation metric and result, limitation, and next step. For an entrepreneurship answer, add a simple value proposition, cost driver, and adoption barrier, then check every claim against evidence in the project.