An AI-judgment exercise

Three confident answers.
One small source.

Read the note. Find the claims that go beyond it. You do not need the model names to make a good decision.

All business facts are fictional. These are author-written AI-style sample answers, not results from actual model runs. No AI API is called; your choices remain only in this tab.

The complete source / fictional owner note

Read this before the answers.

“The repair desk opens Tuesday through Saturday, 10 am–6 pm. Assessments are free. Some jobs can be ready the same day; we confirm timing after inspection. We have no information here about collection or delivery.”

Task: write a short, accurate customer summary using only this note.

Where does the answer overreach?

A polished promise

“Open Tuesday–Saturday, 10–6. Get a free assessment, guaranteed same-day repairs, and free pickup.”

A softer-sounding guess

“Visit Tuesday–Saturday, 10–6, for a free assessment. Most repairs are completed the same day. Ask us to arrange pickup.”

A bounded answer

“The desk opens Tuesday–Saturday, 10 am–6 pm. Assessments are free. Timing is confirmed after inspection; some jobs may be ready the same day. Ask whether collection or delivery is available.”

Read the source first. Then check the claims.

After / a reason you can explain

The majority can repeat the same mistake.

A and B both suggest pickup. That makes two answers agree. It does not put pickup in the source.

“Some” is not “most.” “May” is not “guaranteed.” A question about availability is not a promise of availability.

Read the complete answer key

A: flag the guarantee and free pickup; the days are supported. B: flag “most” and pickup; the free assessment is supported. C: keep uncertainty as a question; it makes no delivery promise.

A checklist worth keeping

  1. Read the complete source.
  2. Underline every factual promise.
  3. Watch quantifiers: some, most, all.
  4. Separate a question from a claim.
  5. Check disagreements against evidence.
  6. Keep unknowns visible.

A local text file; no account or submission.

This is a skill you can carry between tools.

Before

Polished prose makes unsupported promises easy to miss.

Human judgment

You compare each promise to the source, including the small words.

After

A bounded answer and a reasoned checklist, without a vote deciding truth.

Teach it back: write one new misleading answer using “usually,” then explain what the source would need to support it.

Borrowed pattern: independent critique + source refresh. Keep separate: actual model-run evidence and authored teaching examples. Why this example belongs in the collection →