A likely answer is not a checked answer
This lesson focuses on text assistants. Their language models learn patterns for generating text; fluent wording does not prove that a fact is correct. Training can encourage careful answers, and tools can supply useful evidence, but neither makes every answer reliable. A model can say “I do not know”, give a correct answer, or make a mistake. It does not always fill a gap with a confident guess.
Four patterns to watch for
- Hallucination: an invented or unsupported claim, such as a book title that does not exist or a detail missing from the source being summarized.
- Bias: an unfair pattern in how people or groups are represented or treated. Data, training and design choices, and how a system is used can all matter. For example, repeatedly assuming a club leader must be a man is a stereotype to question.
- Stale information: training alone does not supply continuous updates. An assistant may receive newer information through tools or documents, but the version and date still need checking.
- Misunderstood instructions: words such as “bank” can mean different things. Give context or ask a clarifying question. A confident answer to the wrong interpretation is still unhelpful.
Worked example: check a school notice
Imagine this fictional notice is the only source you have: “Robotics club meets on Wednesdays at 3 pm in Lab 2. The club leader will be announced later.” You ask an assistant when the club meets and who leads it. It replies: “Wednesday at 3 pm in Lab 2, led by Ms Patel.” Most of the answer matches the notice. The name does not.
Separate the supported claim from the guess
- List the claims: Wednesday, 3 pm, Lab 2, and Ms Patel as the leader. Check each one, rather than treating the whole sentence as correct or incorrect.
- The notice supports the day, time and room. It does not identify the leader. Ms Patel might happen to be the leader, but this notice cannot establish that.
- A supported reply is: “Wednesday at 3 pm in Lab 2. The notice does not name the leader; check the later announcement.” Keep the missing detail visible instead of filling it in.
Try it before looking at the answer
An assistant adds “Bring a laptop” to its summary of that notice. Can you repeat this as a club requirement? No: the notice says nothing about laptops. Check the official instructions or ask the organizer. A sensible-sounding extra is still an unsupported claim.
Retrieval helps, but the answer still needs checking
Retrieval-augmented generation, or RAG, gives a model relevant source material with the question. It can help an assistant answer from current documents. It can also retrieve an old notice, miss an important passage, or produce an answer that misreads the right source. Even a real citation is not proof that every claim is supported. Grounding in a source and being factually correct are related checks, not identical ones: a source itself can be wrong.
A short check before using an answer
- Open the source. Check that it exists, is relevant, and comes from someone who can establish the fact.
- Match each important claim to the evidence. Check names, dates, numbers, conditions and exceptions; do not stop at seeing a citation.
- Check the source’s date or version when the answer can change. Use another trusted source or ask the responsible person when a detail remains unclear.
- Rephrase to clarify what you meant, not to prove accuracy. A repeated answer can still be wrong; a changed answer can reflect new context or a correction.
Confidence is a reason to check, not a test result
You do not need to reject every AI answer. Use it to explain, organize or draft, and decide which claims need verification before you rely on them. If the available evidence does not answer the question, saying what is missing is more useful than producing a polished guess.
Sources and a next step
Check yourself · 3 quick questions
The notice gives the club’s day, time and room, but says the leader will be announced later. An AI names Ms Patel. What does the notice establish?
Show answer
The day, time and room; it does not establish the leader’s name — Check each claim. The meeting details are supported, while the name needs another source. Missing evidence does not by itself prove the name false.
Why can an AI answer show an unfair pattern?
Show answer
Data, training or design choices, and how the system is used can all contribute — Bias can have several contributors. A citation or a fluent answer does not show that people and groups are represented fairly.
An assistant repeats the same club requirement three times. What is the best accuracy check?
Show answer
Check the requirement against current official club instructions — Repetition does not verify a claim. Relevant, current evidence is a better check than wording, confidence or consistency alone.