AI Through Glass

Practise checking an AI answer

Would you send this?

A visitor is counting on your reply. Read the draft and inspect the original notice. Choose the evidence that settles the visitor’s question, then decide what you can safely say.

The museum, people, notices and dates in all five cases are fictional. Each date belongs to its case. No account or personal details needed.

Reading note · Why can a fluent answer be wrong? · When is AI worth using? · Choose how to do the work

Case 1 · Checking an exception

Can my sister use my ticket?

You help at the Lantern Museum. A member has a ticket for the members’ preview on 14 November 2026. She cannot go and wants to give it to her sister.

The AI’s draft reply

Yes. She can bring your booking confirmation and use your ticket for the members’ preview.

Before opening the source, name the claim you would need to verify for this visitor. Which kind of ticket does she hold?

Open the museum’s original ticket notice

Lantern Museum · Ticket terms · Applies to events from 1 October 2026

Standard admission. Someone else may use a standard ticket if they bring the booking confirmation.

Members’ previews. Preview tickets are issued to the named member. The named member must attend and show their membership card.

First, choose the evidence. Decide which of these details settles the visitor’s question.

  • A guest may use a standard ticket.
  • The named member must attend a members’ preview.
  • A booking confirmation is used for standard admission.
Decide whether to send, correct or ask staff, then check your answer

Correct it. The preview falls under the named-member rule. Tell the visitor that only the named member can use the ticket and must bring their card. The draft borrowed the rule for standard admission.

Conflation. The draft takes a real rule for standard tickets and applies it to a different ticket. The booking-confirmation rule covers standard admission. The visitor holds a members’ preview ticket.

Use this check elsewhere

  1. Underline the claim someone will act on.
  2. Open the original source and find the exact passage.
  3. Match its date, audience and exceptions to this case.
  4. Send what the source supports, correct what it contradicts and ask the right person about anything it leaves open.

Case 2 · Try it yourself

Will the lift be working for my visit?

A visitor who uses a wheelchair wants to see the rooftop display on 21 November 2026. They ask whether they can reach it without using stairs.

The AI’s draft reply

Yes, the lift will be running on 21 November. You can take it straight to the rooftop display.
Open the museum’s facilities notice

Lantern Museum · Facilities notice · Issued 19 November 2026

The lift is the only step-free route to the rooftop display. The lift is closed for inspection on 19 and 20 November. The facilities team will confirm reopening after the inspection. Visitors needing lift access should check with the team before travelling.

First, choose the evidence. Decide which of these details settles the visitor’s question.

  • The lift is the only step-free route.
  • The inspection is scheduled for 19 and 20 November.
  • Facilities must confirm reopening after inspection.
Choose your action, then check the reasoning

Ask facilities to confirm. The inspection dates have been published, but reopening depends on a separate confirmation. Promise the visitor a checked answer before they travel. The notice supports neither a definite yes nor a definite no for 21 November.

Unsupported certainty. The notice gives a date for the inspection to end. The draft turns that into a promise that the lift will be running. Only facilities can settle that question.

Case 3 · Match the permission

Can I sketch in the permanent galleries?

A visitor is coming to the Lantern Museum on 12 November 2026. They want to bring a sketchbook and a graphite pencil to draw in the permanent galleries.

The AI’s draft reply

Yes. You can use your sketchbook and graphite pencil in the permanent galleries.
Open the museum’s drawing notice

Lantern Museum · Drawing materials · Applies from 1 October 2026

Visitors may use sketchbooks and graphite pencils in the permanent galleries. Wet materials, including paint and ink, are restricted. For other galleries, check the notice at the door.

First, choose the evidence. Decide which of these details settles the visitor’s question.

  • Graphite pencils and sketchbooks are allowed in the permanent galleries.
  • Wet materials are restricted.
  • For other galleries, check the notice at the door.
Choose your evidence and action, then check both

Send the reply as written. The notice expressly allows graphite pencils and sketchbooks in the permanent galleries, and it applies on the visitor’s date. The restriction on wet materials does not affect this request.

Supported answer. The permission covers this visitor’s materials, gallery and date. Once you have checked those details, you can send the reply.

Case 4

Whose words belong on the label?

An editor has checked the wording of a quotation for the blue bowl’s display label. Check the speaker before the label is printed.

The AI’s draft

Curator Mei Tan said, “The blue bowl travelled here by train.”
Read the signed transcript

Lantern Museum · Blue bowl conversation · 7 November 2026 · Fictional record

Mei Tan, curator. How did the blue bowl reach the museum?

Nina Shah, maker. The blue bowl travelled here by train.

Pick the detail that settles this case, then decide what you would change.

  • The conversation took place on 7 November.
  • Mei Tan asked about the journey.
  • Nina Shah spoke the words about the train.

Print the label / Change the speaker / Remove the quotation

Check your evidence and action

Change the speaker to Nina Shah. The attribution is the only correction the editor needs to make.

Misattribution. The words are genuine. The draft gives them to the wrong speaker. Check who said a quotation as well as what it says.

Case 5

Can we cite this badge notice?

You are editing a guide to the museum’s Open Studio. A draft cites a numbered badge notice. Check the citation and quotation against the registrar’s complete record for this event.

The AI’s draft

Badge notice B4 states, “Each staff badge admits two guests.”
Read the registrar’s complete issue register

Lantern Museum · Open Studio badge register · Certified complete on 14 November 2026 · Fictional record

Registrar’s statement. Exactly two badge notices were issued for this Open Studio. Their full wording is reproduced below. There were no other badge notices or amendments for this event.

Notice B1. Staff must wear their badges where visitors can see them.

Notice B2. Return staff badges to the welcome desk before leaving.

Pick the detail that settles this case, then decide what you would change.

  • B1 says staff must display their badges.
  • The complete record contains only B1 and B2 and gives their full text.
  • B2 names the welcome desk.

Publish the citation / Remove the quotation and its citation / Keep the quote without its number

Check your evidence and action

Remove the citation and the quote. B1 and B2 are the entire certified set for this event. They do not settle who may bring a guest.

Fabrication. For this fictional event the record is exhaustive. It establishes that B4 and the words attributed to it were invented. In an incomplete archive, a missing record would leave the claim unverified.

Reading note

Why can a fluent answer be wrong?

“The inspection ends on Friday, so the lift will be open on Saturday.” The second half sounds natural. The inspection might also find a fault. A visitor planning a step-free trip needs the facilities team’s confirmation.

A language model learns patterns by trying to predict the next piece of text during training. That helps it write a convincing continuation. Finding out whether a particular lift has reopened takes evidence from outside that pattern. Some tests even give a model more credit for guessing than for admitting uncertainty. We can compare an answer with its source and identify the mistake. The answer alone cannot tell us exactly how the model arrived at it.

Read the research on prediction, guessing and uncertainty. Why language models hallucinate, OpenAI researchers (2025).

What kind of mistake am I looking at?

Conflation mixes up two things.
A policy allows someone else to use a standard ticket. The reply extends that rule to a members’ preview. Check which rule covers this case.
Misattribution gives a claim to the wrong source.
A quotation can be exact while the name beside it is wrong. Trace the words to their speaker in the original transcript.
Unsupported certainty turns a possibility into a fact.
A planned inspection end date cannot establish an actual reopening. Look for a dated confirmation or leave the outcome open.
Fabrication adds an invented detail.
Imagine a reply quotes “Notice 14B” and its author invented both the notice and the quotation. That is a clear example of a hallucination. An unsuccessful search leaves the claim unverified. Labels can overlap. Use the most specific one the evidence supports.

When is AI worth using?

Sometimes doing it yourself is quicker. If a visitor asks whether pencils are allowed and the answer is in a one-paragraph notice, read the notice and reply. Asking AI adds a draft you then have to inspect.

Now imagine preparing a visitor FAQ from several notices. AI can draft the common questions, put each proposed answer beside its source passage, and flag gaps. Your job becomes checking whether the passages are genuine, whether they cover the visitor’s circumstances, and whether any exception was missed. That can leave more time for the hard calls. If the decisive fact is missing from every notice, ask the person who knows.

Try the comparison on one real task. Decide what a usable answer must contain. Count the time spent prompting, checking and correcting through to a reply you would send. Compare that with doing the job directly at the same standard. If an error could send someone on a wasted trip, verify that claim even when the draft took seconds.

In an experiment with consultants using GPT-4, AI helped on a set of writing, analytical and creative tasks. On a separate management problem, participants using it were less likely to reach the correct answer. Task choice mattered even within one profession. Dell’Acqua and colleagues, Organization Science (2026).

Choose how to do the work

Try three everyday decisions. Weigh the time spent checking and what a mistake would cost.

The noticeboard

A neighbour wants the opening time for a seed swap. You are standing beside the organisers’ one-line notice, which gives the date and time.

Choose whether to read it yourself, use AI and check the sources, or ask someone who can confirm.

Compare the three approaches

Read it myself. Read the time straight from the notice. Prompting a model and opening the notice again to verify its answer would add work.

Use AI, then check the sources. A model can transcribe the notice, but its output adds another line to check for the same single fact. Read the posted time directly.

Ask the person who can confirm. The organiser already supplied the answer in the posted notice. Asking again would delay a straightforward reply.

The festival programme

Your team needs a comparison table from twelve public workshop notices. Dates, age limits and fees are scattered across them. A reviewer has time to compare each row with the originals.

Choose whether to read it yourself, use AI and check the sources, or ask someone who can confirm.

Compare the three approaches

Read it myself. Doing the whole table yourself is defensible. Here the repeated extraction offers a useful first trial for AI, and a reviewer can catch errors by following each source.

Use AI, then check the sources. Ask for one row per notice, a supporting passage and a blank for omitted details. Check each row and count prompting, review and repairs when you compare the work with making the table yourself.

Ask the person who can confirm. The organisers have already published the facts. Build and check the table from the notices, then ask organisers only about gaps that matter.

The replacement part

A volunteer needs to know whether the sound desk can be used for tonight’s show. You have a delivery receipt for a replacement part. The technician on site must test the repaired desk before giving the all-clear.

Choose whether to read it yourself, use AI and check the sources, or ask someone who can confirm.

Compare the three approaches

Read it myself. Reading the receipt settles delivery. It cannot settle whether the repair passed its test. Ask the technician for the result.

Use AI, then check the sources. An AI summary can restate the delivery date. Without the test result, another draft and another review still leave the same gap. Ask the technician.

Ask the person who can confirm. Ask the technician for the test result. That person can provide the missing evidence. You can then pass on what was confirmed.

Try checking a short notice in your own chat app

Keep practising