Ask what the system has actually received and demonstrated
Review training and inference, generated content, narrow evaluation, confabulation and evidence for current facts.
Five matching choices describe this section’s finite scenarios. They do not establish expertise with every model or dataset.
Separate using context from changing parameters
A model can use a paragraph supplied during a task without that paragraph necessarily retraining its parameters. The practice item fixes the parameters for the session, making that distinction explicit. Storage, retention, later training and adaptation depend on the system and its settings; none should be assumed from the response alone. In a fictional exercise, describe which text was supplied now and which model properties were stated to remain fixed before explaining what the response can establish.
Identify the described output rather than a tool label
Drafting a new invitation produces content, while assigning an image to a predefined category describes classification. A system can support several kinds of task, so these descriptions do not divide all products into mutually exclusive boxes. Begin with the requested output: new text, a label, a number or an action record. Then identify which evidence would show that output meets its purpose. A familiar product name does not determine whether every task performed through it is generation, retrieval, calculation or classification.
Keep evaluation conditions attached to the claim
An evaluation on bright photographs does not establish performance on dark photographs merely because the same model is involved. Relevant input differences can change the task. A new invented exercise might compare clean scans with blurred scans: list the conditions evaluated and those still untested. This identifies a gap rather than declaring success or failure in the missing condition. Broad claims need suitable evidence for the intended setting, and a large count of examples from one narrow condition does not automatically cover every other condition.
Treat confidence as presentation rather than proof
A generated explanation can contain false facts, invented references or reasoning that does not support its conclusion. The website’s examples are invented to illustrate that possibility; they are not recordings from a named model. Check the actual claim against evidence rather than judging by tone, length or polish. Finding one error does not establish that every output is wrong, while finding one correct paragraph does not establish universal reliability. Keep the claim and the scope of the check separate.
Check where a current fact could have come from
If a tool is described as having no live search or newly supplied source, a confident answer about today’s changed schedule does not prove access to that change. Ask which current authoritative evidence is available and whether it supports the specific time or rule. Different systems update, retrieve and adapt in different ways. This guide does not assume that every tool is offline or that every tool browses automatically; the scenario’s declared capabilities matter to the conclusion.
Make the evidence boundary visible in a new scenario
Write a fictional task with a short source paragraph, a stated tool capability and a proposed claim. Mark what is directly provided, what is inferred and what remains unknown. You can do this on paper without running an AI model. The aim is to explain the boundary rather than invent a hidden training history or a performance percentage. This exercise asks for no private document, model account or paid tool, and the online quiz does not execute a model on your behalf.
Use the count as a pointer to a specific distinction
A count out of five reflects matching answers on these supplied scenarios, not a professional level or a population ranking. Every count receives the complete explanations and all four guides. For a useful retake, change the source or the declared capability and explain how that changes the supported conclusion. Repeating the same fixed bank after reading its keys can improve recall without demonstrating skill across unfamiliar systems. This independent practice does not award a UNESCO, OECD or NIST credential.