Free AIF I practice test: 30 questions with answers

A hallucination is output that reads as plausible but is not supported by the input or by reliable knowledge, and the quote here has no source in the transcript. The summary was not produced by a web search, a memory of another customer, or a fixed template, so the fix is to ground the summary in the transcript and check every quote against it.

Models read and produce text as tokens, which are chunks such as whole words or word fragments, and context limits and usage are measured in tokens. The other options describe access credentials, saved chat history, and a provider's rating of users, none of which defines a token.

Generated text is sampled from a probability distribution over possible next tokens, so repeated runs can differ, and settings such as temperature control how much variation is allowed. The other options describe mechanisms that do not explain normal run-to-run variation: the model does not switch databases per run, change its weights after each answer, or mix in a random log of past answers.

A model answers from patterns learned during training plus whatever context is supplied, so it can lack recent events and can reflect gaps or biases in its data. Fine-tuning usually adjusts behaviour rather than erasing everything learned, and even simple questions can produce factual errors.

Training adjusts a model's parameters by learning patterns from data, while inference runs the trained model to produce outputs for new inputs such as a user's prompt. The other options swap the two roles or merge them into one step, which confuses the lifecycle stages.

A model's built-in knowledge reflects its training data and cannot show today's stock, and fine-tuning on last year's spreadsheets would freeze outdated figures. Supplying live records through retrieval or a tool grounds each answer in current data. Raising the temperature only adds randomness, which makes the answers less reliable.

Each model has a context window, a limit on how many tokens it can consider at once, so text beyond that limit may not be used in the answer. Legal phrasing is not the blocker, chat tools do not switch off questions by section number, and the training date has no bearing on reading a document supplied in the prompt.

Calculations with fixed rules need exact, repeatable results, which deterministic code provides, while language models produce variable text and can make arithmetic errors. Using a model for the calculation, raising its temperature, or averaging two runs does not make the result reliable.

Specific context, audience, and format give the model what it needs to produce output that fits the product, which is the core of effective prompting. Repeating a vague request or shortening the answer does not add the missing information, and a new chat with the same vague request will tend to return the same generic result.

Role prompting asks the model to respond from a stated perspective, which shapes tone, vocabulary, and focus. The other options describe output length limits, splitting files, or a token budget, and none of these is what the term means.

Iteration works best when feedback is specific, so the model can see what to fix and what to keep. Discarding the draft throws away useful work, asking for more length does not guarantee the missing sections, and accepting it leaves a known gap unfixed.

Versioned records let the team compare and roll back changes, and testing on representative inputs shows whether a revision helps across cases rather than one. Changing several things at once hides which change caused an effect, one output is only anecdotal, and deleting old versions removes the history needed to compare.

Stating the structure, such as one row per item with named columns, makes the output predictable and easy to import. A more creative style, a shorter source, or step-by-step reasoning does not control the format of the answer.

Vague goals produce vague rewrites, so stating the audience, tone, constraints, and what must be preserved gives the model criteria it can follow. A synonym list, more randomness, or deleting the source copy does not define what better means.

Labelled examples show where the boundaries between categories lie, which steers the model toward consistent decisions; this is a form of few-shot prompting. Creative justification, a longer billing list that drops two categories, and a higher temperature all add variation or distort the task.

Explicit instructions about what to do when information is missing stop the model from inventing owners, and marking items as unassigned keeps the record accurate. Requiring an owner for every item, asking for more items, or allowing longer answers all encourage the same invention.

Acceptance criteria must match the real requirements, and only a comparison against those sources and the team's product knowledge can confirm that. Fluent wording, a finished-looking layout, and the first result on screen are not evidence that the content is correct.

Grounding connects an answer to reliable supplied material, such as retrieved documents or data, so its claims can be traced and checked. Short answers, a zero temperature, or training on its own text do not tie a response to any source.

Verification means checking claims against authoritative sources, having experts review high-impact output, and comparing results with known-good examples. A model confirming its own answer can repeat the same error, and citations can be fabricated or misattributed, so neither counts as proof.

The confident summary contained a wrong figure, which is a known failure mode, so every number must be checked against the tracking tool before it is reported. The tracker is the source of record and should not be replaced by model output, and a model does not gain lasting knowledge from one conversation.

Assured wording describes how the text is written, not whether the facts behind it were checked, so the claim still needs evidence. A confident tone does not make a claim more likely to be right than a hesitant one, and sharing it without review skips the check.

Nine of ten samples is too small to show how the tool handles real invoice formats, so the failure should be understood and the test widened first. Declaring success, hiding the failure, or releasing to every customer and waiting for complaints all skip proper evaluation.

A representative set exposes how the feature handles typical and unusual inputs, which is the evidence needed before scaling. Hand-picked examples cannot cover every case, production monitoring is still needed, and testing does not train the model.

Organisations and the people who approve AI output remain accountable for what they send, whatever tool drafted it. The vendor, nobody, and the customer are all wrong because accountability does not transfer to a tool or to the recipient.

Human in the loop means a person stays part of the process at the points where AI output could cause harm or drive decisions, with authority to approve, change, or stop it. Fully autonomous operation, feedback collection alone, and physical supervision rules are not what the term means.

Sensitive business and personal data should only go to tools the organisation has approved for that classification, so checking the data-handling policy comes first. Roadmaps can be sensitive, removing a company name does not anonymise personal data, and a provider's safeguards do not remove the team's own obligations.

A traceable review record and a named accountable owner for each decision keep responsibility with people. Handing logs or sign-off to the tool, or treating AI output as neutral, removes the human accountability that the process depends on.

Transparency means being open about where AI was used and what its limits are, so the people relying on the work can judge it. Hiding AI use, publishing every raw exchange publicly, or sharing internal model settings is not the right kind of openness and can expose confidential material.

A consistent pattern against one group suggests possible bias, so it should be investigated, documented, and decided by accountable people rather than ignored. Accepting the ranking, deleting the requests, or banning the tool forever each skips the review the situation needs.

Without a record, the team cannot show who did what, which parts were AI-assisted, or how the findings were checked, which weakens accountability. Vendor billing, model learning from internal records, and replacing human review are not reasons such records matter.