Assam HS Skill Subject Artificial Intelligence common mistakes
Assam HS Skill Subject Artificial Intelligence · Understand the traps and how to avoid them.
Writing `x = input()` and then comparing the result numerically (e.g. `if x > 10`) without converting it, so the program crashes with a TypeError or compares strings alphabetically.
Candidates learn input() early and forget that it always returns a string; the code looks correct when read aloud, and the error only appears at runtime, which they rarely test with real values.
What to do instead
Always wrap input in the target type: x = int(input()) or x = float(input()), and practise writing a two-line test that prints type(x) before the comparison. In the exam, mentally trace one sample input through the program before answering.
Mixing up supervised and unsupervised learning in scenario questions, labelling a clustering task (e.g. grouping customers with no labels) as classification, or calling a labelled spam-detection task unsupervised.
The definitions are memorised as slogans ('supervised has labels'), but the exam gives realistic scenarios where the presence of a target column is not stated explicitly, so candidates pattern-match on keywords like 'group' or 'predict'.
What to do instead
Apply a fixed two-question test to every scenario: (1) Is there a known target/output column in the data? (2) Is the output a category or a number? Yes+category = classification, yes+number = regression, no = clustering/unsupervised. Write this decision rule on rough paper before starting the section.
Choosing accuracy as the evaluation metric for an imbalanced dataset and concluding the model is good (e.g. 95% accuracy on fraud data where 95% of rows are non-fraud).
Accuracy is the first metric taught and the easiest to compute, and candidates do not habitually check the class distribution before judging a model.
What to do instead
Before evaluating, state the class balance. If one class dominates, discuss precision, recall, F1 or a confusion matrix instead. In answers, explicitly say 'accuracy is misleading here because...' and give the baseline (predicting the majority class) for comparison.
Describing a perceptron as if it were a full neural network or claiming a single perceptron can solve XOR, or stating that adding layers alone (without a non-linear activation) makes a network more expressive.
Textbook diagrams of a perceptron and a multi-layer network look similar, and the role of activation functions is often memorised as 'it adds non-linearity' without understanding why that matters.
What to do instead
Learn the three-part story: a perceptron computes a weighted sum plus bias and passes it through a step function, so it only separates linearly separable data (XOR fails). Stacking layers with only linear activations collapses to one linear function; non-linear activations (ReLU, sigmoid) are what allow complex boundaries. Be ready to sketch the XOR failure in one sentence.
Confusing AI, machine learning and deep learning as synonyms, or defining AI only as 'robots' and ML only as 'computers that think', producing vague definitions in short-answer questions.
Everyday media uses the terms interchangeably, and candidates repeat popular phrasing instead of the nested-set relationship the syllabus expects.
What to do instead
Memorise the nesting: AI is the broad goal of machines performing intelligent tasks; ML is a subset where systems learn patterns from data instead of explicit rules; DL is a subset of ML using multi-layer neural networks. For each term, prepare one concrete example (rule-based chatbot = AI, spam filter = ML, image classifier = DL) and use that example in the answer.
Treating data cleaning as optional and jumping straight to modelling in the project-lifecycle question, or listing 'collect data' and 'train model' with nothing in between.
Project questions feel like storytelling, so candidates write a short narrative and skip the unglamorous steps they have not personally performed.
What to do instead
Use a fixed lifecycle skeleton and fill each stage with one action and one artefact: problem framing (success metric), data collection (source and size), cleaning (missing values, duplicates, outliers, encoding), EDA/visualisation (one chart and what it showed), splitting (train/test), model building, evaluation (metric and result), documentation and presentation. Practise writing this skeleton from memory in under five minutes.
Answering ethics questions with a single sentence like 'AI should be fair and protect privacy' without naming a mechanism, a stakeholder or a trade-off.
Ethics feels like common sense, so candidates assume general goodwill statements earn full marks and do not prepare concrete vocabulary.
What to do instead
Prepare a small toolkit and deploy at least two items per answer: biased training data, proxy variables, informed consent, data minimisation, anonymisation, transparency/explainability, human-in-the-loop, audit, accountability. Structure the answer as harm → cause → mitigation → who is affected. For job displacement, distinguish task automation from job loss and mention reskilling.
Writing Python loops that are off by one (e.g. `for i in range(len(lst))` combined with `lst[i+1]`, or `range(1, n)` when n items are needed), or using `=` instead of `==` in conditions.
Candidates trace code by intention rather than by literal execution, and small index errors are invisible unless a concrete list is stepped through.
What to do instead
On rough paper, write a tiny concrete input (e.g. [3, 5, 7]), then maintain a table of variable values after each iteration and check the loop's first and last iteration explicitly. Check every condition for == versus = before finalising.
Describing an NLP pipeline as 'the computer understands the sentence', skipping preprocessing steps, or claiming a chatbot is trained from scratch on the school's data when it is rule-based or API-based.
Candidates repeat marketing language about AI understanding language and have not internalised tokenisation, stop-word removal, stemming/lemmatisation or the difference between retrieval and generation.
What to do instead
Learn the ordered pipeline: raw text → cleaning (lowercase, punctuation) → tokenisation → stop-word removal → stemming/lemmatisation → vectorisation (bag-of-words/TF-IDF) → model. For chatbots, classify as rule-based (pattern matching), retrieval-based (best match from a store) or generative, and state which one your project used and why.
Reading pandas or NumPy code as if it were plain Python: assuming `df[df['age'] > 30]` modifies the dataframe, or that `df.dropna()` changes df in place, or misreading `axis=0` versus `axis=1`.
Candidates transfer list and loop intuition to vectorised library calls and do not know which operations return a new object versus mutate.
What to do instead
Memorise three rules: boolean indexing returns a filtered copy unless assigned back; most pandas methods return a new object unless inplace=True is given; axis=0 operates down rows (per column) and axis=1 across columns (per row). Practise predicting the output of five short snippets with a small dataframe before the exam.