EXIN BCS Gen AI Award common mistakes
EXIN BCS Gen AI Award · Understand the traps and how to avoid them.
Treating a large language model as a factual database and answering 'what is the capital of X' style questions as if the model retrieves verified facts, rather than recognising it predicts likely tokens and can hallucinate.
Candidates conflate fluency with accuracy; the model's confident tone feels like retrieved knowledge, so they forget that generation is probabilistic and unverified.
What to do instead
Reproduce the mechanism in your own words: next-token prediction from training data, no live lookup, no guarantee of truth. For any scenario question, ask 'is a verification step present?' and if not, mark the output as unreliable regardless of how plausible it reads.
Answering questions about what generative AI 'cannot' do by denying capabilities it actually has, such as summarising, translating, code generation, or image synthesis from text.
Candidates overcorrect after learning about limitations and swing to blanket scepticism, or they only know one modality (text) and assume that is the whole field.
What to do instead
Keep a two-column list: demonstrated capabilities (summarisation, translation, drafting, classification, code assistance, image/audio generation) versus structural limits (no guaranteed facts, no true understanding, context window bounds, no persistent memory by default). Revise the capability column explicitly before the exam.
Ignoring the context window and assuming the model 'remembers' everything from earlier in a long conversation or a large pasted document.
Chat interfaces feel persistent, so candidates assume memory is unlimited; the token limit is an implementation detail rarely visible in demos.
What to do instead
Learn that context is a finite token budget covering prompt plus prior turns plus output, and that content outside it is dropped. In scenario questions, look for signs of truncation or lost earlier instructions and attribute errors to context limits rather than model 'laziness'.
Writing prompts that state a goal vaguely ('write something good about our product') and then blaming the model when output is generic or off-target.
Candidates treat the tool like a search engine or a human colleague who shares their context, so they omit role, audience, format, and constraints.
What to do instead
Use a checklist for every prompt-design question: task, role/persona, audience, context and source material, output format and length, tone, constraints and exclusions, plus examples where useful. If a scenario prompt lacks these, the correct answer is usually to add specificity rather than to change the model.
Assuming one prompt should produce a perfect result, and not recognising iteration, refinement, and follow-up correction as part of good practice.
Single-shot demos create the impression that prompting is one-and-done; candidates also fear that admitting iteration looks unprofessional.
What to do instead
Treat prompting as an iterative loop: draft, inspect output, give targeted feedback ('shorter', 'add a counter-argument', 'use formal register'), and repeat. On the exam, prefer answers that include review and refinement over answers that promise a perfect first attempt.
Confusing prompt techniques, for example describing few-shot prompting as 'giving the model more time to think' or chain-of-thought as 'sending the prompt several times'.
The terminology is new and overlapping, and candidates memorise labels without the underlying mechanism.
What to do instead
Anchor each technique to its mechanism: zero-shot (instruction only), few-shot (worked examples in the prompt), chain-of-thought (elicit intermediate reasoning steps), role prompting (persona), and structured output (specify schema). Write one sentence and one example for each from memory.
Answering ethics questions by saying 'the model is unbiased because it is a machine', or conversely claiming any biased output proves deliberate designer intent.
Candidates collapse the distinction between training data, model behaviour, and human accountability, and they lack a mental model of where bias enters.
What to do instead
Trace the pipeline: historical data reflects past bias, training amplifies or dampens it, deployment context determines harm, humans remain accountable. Choose answers that name a concrete mitigation (diverse data, testing, human review, documented limitations) rather than denying the problem exists.
Pasting confidential customer data, source code, or personal information into a public generative AI tool without considering data protection and confidentiality obligations.
Convenience and the chat-style interface make the tool feel private; candidates underestimate that inputs may be retained, logged, or used for improvement depending on the service and settings.
What to do instead
Apply a default rule: personal data, trade secrets, and client-confidential material stay out of consumer tools unless an approved enterprise arrangement exists. On the exam, identify the data classification first, then the tool's terms, then the lawful basis before choosing an answer.
Treating generated output as automatically free of third-party rights and safe to publish, ignoring copyright, licensing, and the need for human verification of AI-produced content.
The 'the machine made it' framing feels like it transfers or removes responsibility, and candidates do not know that output can reproduce protected material or contain fabricated citations.
What to do instead
Remember that the deploying organisation remains responsible for what it publishes. Check for similarity to protected works, verify every factual claim and citation, respect licence terms of the tool and of any training-content policy, and keep a human sign-off step in any workflow answer.
Proposing generative AI for a business problem without linking it to a measurable outcome, cost, or risk, and choosing 'use AI everywhere' style answers over targeted, evaluated use cases.
Hype-driven framing dominates public discussion, so candidates default to enthusiasm; they also skip the evaluation and pilot stage because it feels like exam-irrelevant project management.
What to do instead
Structure every value question as problem, baseline metric, AI-enabled change, evaluation method, and governance. Favour answers that start with a narrow, high-volume, low-risk use case, define success measures, and include human oversight and a rollback path over answers that promise transformation without evidence.