ICDL AI Essentials common mistakes
ICDL AI Essentials · Understand the traps and how to avoid them.
Labels any automated or rule-based software as AI - e.g. calling a fixed-script chatbot, a calculator or a simple if/then workflow 'artificial intelligence'.
Vendors brand almost every product 'AI', and exam scenarios deliberately use ordinary office tools, so candidates pattern-match on the marketing label instead of the underlying mechanism.
What to do instead
Apply a definitional test before answering: does the system learn patterns from data and generalise to inputs it was not explicitly programmed for, or does it follow fixed rules written by a developer? Fixed-rule automation is not AI; learning-based systems are.
States that deep learning is broader than machine learning, or treats generative AI as a separate field unrelated to the other two.
The terms overlap and are used interchangeably in the media, so candidates memorise a flat list of buzzwords instead of the nested structure.
What to do instead
Memorise the nesting: AI is the broad field, machine learning is a subset that learns from data, deep learning is a subset of machine learning using multi-layer neural networks, and generative models (large language models, image diffusion models) sit inside deep learning. Practise placing a given example at the correct level.
Writes that an AI 'understands', 'knows', 'wants' or 'decides to' do something, attributing human intentions to the system.
Fluent conversational output and human-like phrasing in chat interfaces invite anthropomorphism, and everyday language rewards it.
What to do instead
Rephrase capability statements mechanistically: the model recognises statistical patterns, predicts probable outputs, optimises an objective function. Reserve words like understanding and intention for humans, and note that current systems are narrow - they do not possess general intelligence.
Identifies only chatbots and obvious tech products as AI, missing recommendation feeds, spam filters, route planning, fraud detection, speech recognition, translation and predictive maintenance.
The most widespread AI is embedded and invisible, so candidates recall what they can see on screen rather than what runs in the background.
What to do instead
Revise a category checklist - recommendation, filtering and ranking, prediction and forecasting, recognition (vision and speech), navigation and optimisation, generative content - and map each exam scenario to one of these categories.
Assumes AI recommendations are neutral, objective and purely helpful, without naming the commercial objective behind them.
Personalisation feels like a convenience service, so the engagement metrics and advertising incentives driving the ranking are invisible to the user.
What to do instead
For any recommender scenario, state what is being optimised (watch time, clicks, purchases) and the side effects: filter bubbles, reinforcement of existing preferences, amplification of extreme or low-quality content, and reduced exposure to opposing views.
Sends a single vague line such as 'write about sustainability', receives generic output, and concludes the tool is incapable.
Candidates carry over search-engine habits, expecting the tool to infer intent from a few keywords rather than from explicit instructions.
What to do instead
Use a structured prompt: task, context and audience, desired format and length, tone, constraints and an example of good output. Then iterate - review the result, point out what is wrong, and refine rather than abandoning the attempt.
Accepts generated facts, statistics, quotations, citations and references without checking them against a primary source.
Output is fluent, well formatted and confidently phrased, which readers unconsciously read as evidence of accuracy; fabricated details look indistinguishable from correct ones.
What to do instead
Treat every output as an unchecked draft. Verify each factual claim, number, quotation and reference in an authoritative source, and remove or flag anything that cannot be confirmed - particularly citations, which are a frequent hallucination.
Pastes client names, personal data, unpublished financials or confidential documents into a free public AI tool to get help summarising or rewriting them.
The chat box feels like a private notebook, and free tiers often reserve the right to use submitted content for model improvement or human review.
What to do instead
Check the tool's data-retention and training policy first, use an approved enterprise or private instance where available, and redact or anonymise names and identifiers before prompting. If policy does not permit it, do not paste it at all.
Argues that an AI system cannot be biased because it is 'just maths' and has no opinions.
Mathematical operations feel objective, so candidates overlook that the data, labels, features and optimisation target all encode human choices and historical inequalities.
What to do instead
Explain the pathway: bias enters through unrepresentative training data, biased labelling, poorly chosen features and objectives that favour one group's outcomes. Require bias testing, evaluation across subgroups and human review before deployment.
Answers that the AI is responsible when a harmful or wrong decision is made, or that nothing can be done because the system decided.
Automation bias makes automated outputs feel authoritative, and diffuse ownership inside organisations makes it easy to attribute the outcome to the tool.
What to do instead
State that human accountability does not transfer to the system: a named owner remains responsible. High-impact decisions need human-in-the-loop review, documented checks and an escalation route, plus a way to challenge an automated outcome.