Read / talk / write / Five stages

The AI Archive

Recover five files tracing how computers moved from instructions to generated answers.

Work with a partner. Read the three files, talk about what the school should do, then answer three short questions. Keep your answers for the school debate.

All rooms
Start here

What to do

Read the three files and talk with your partner. Then write three short answers. Brief notes are fine.

AI words in this glossary
Algorithm
A set of steps a computer follows to calculate something or make a recommendation.
IF/THEN rule
A written instruction: if a certain situation occurs, then take a stated action.
Learning from examples
Finding patterns in sample data instead of writing a separate rule for every case.
Generative AI
A system that makes new text or images from learned patterns; its claims still need checking.
AI winter
A period when enthusiasm and funding fell because AI had not met high expectations.
Bias
A pattern of unfair results, even when nobody meant to treat people unfairly.
Safeguard
A specific action that reduces a risk, such as a check by a named person.
Appeal
A chance to challenge a decision and have the evidence checked again.
Pair workYour answers save in this browser. Print a copy whenever you need one.
Investigation progress0 of 5 stages complete

Can ORACLE be trusted with school deadlines?

Test a supplier's claim, compare two kinds of trial and decide whether deadline advice is a suitable school use.

Three files

File A
Supplier's announcement | fictional trial

Six volunteers exchanged messages with ORACLE without knowing whether they were speaking to a person or a computer. Four thought they were speaking to a person. This trial was inspired by the imitation game proposed by Alan Turing in 1950. Volunteers called the replies friendly and convincing. The supplier says immediate answers, even when the school office is closed, show that ORACLE is ready for everyday school questions.

File B
School information check | 12 questions

Staff checked 12 replies against current school information. Ten were correct; two were wrong. Both wrong answers matched an older information sheet. ORACLE gave no warning that its information might be out of date. Staff have not tested differently worded questions.

QuestionORACLE's replyCurrent information
When is the history project due?FridayMoved to Wednesday
Can I return my trip form on Monday?YesForms due Friday

File C
Pupil-support proposal

Pupils sometimes miss announcements or cannot reach the office at lunchtime. Staff also occasionally give outdated information. Immediate answers could help, but a wrong date could mean missing a trip or submitting work late. The school could restrict questions, try a limited pilot, or wait for more testing. Staff cannot check every reply instantly.

Should pupils use ORACLE to find out homework and trip deadlines?

Now imagine: A teacher changes the homework deadline this morning. ORACLE still has the old date. A pupil asks it when the homework is due.

Optional multiple-choice questions
What did the six-volunteer trial record?
What did the separate information check find?

Answer all three questions to continue. Brief notes are fine.

Need a nudge?

Separate judging who is speaking from checking whether an answer matches current school information.

The rule machine

A school help desk follows written IF/THEN rules. Test what happens when a pupil asks something the designers did not predict.

Three files

File A
Help desk rule card

The fictional school tool follows instructions its designers wrote: IF the message says 'forgot password', THEN send reset steps. IF it says 'two deadlines clash', THEN send it to a tutor. Otherwise, send the standard help page. A pupil writes, 'I can do the homework on paper, but I cannot use the online system at home this week.' There is no rule for an offline way to hand in work. The tool sends the standard page, which only explains how to log in.

File B
Technician's test log

The team tried 12 password-reset messages; all 12 got the reset steps. Four messages about offline work all got the standard page. After a technician added an offline-work rule, all four test messages were sent to a tutor. The test shows a useful strength: the same written rule gives the same response each time. It also shows that the tool did not invent a helpful new rule by itself.

File C
Tutor's note to governors

The tutor says routine resets take staff away from other support. Pupils often describe the same problem in different words, and the office has limited lunchtime cover. One proposal puts an 'Ask a person' button behind the pupil login. Another allows a message without logging in but needs someone to monitor it. Neither route has been tested with locked-out pupils.

Which questions should the school help desk answer?

Now imagine: A pupil cannot log in. The help desk does not recognise their message. Its 'Ask a person' button is on a page that needs a login.

Optional multiple-choice questions
What happened before an offline-work rule was added?
What changed in the technician's repeat test?

Answer all three questions to continue. Brief notes are fine.

Need a nudge?

Compare the 12 familiar requests with the four requests that the original rules did not cover.

The abandoned laboratory

A team makes a bold promise after a small demonstration. Decide how far the evidence supports it.

Three files

File A
Atlas launch announcement | fictional project

The Atlas team says its computer will translate school letters into many languages reliably within two years. At a public demonstration, it gets 19 of 20 short sentences right. The sentences were picked in advance and used topics the team had practised. A school governor calls the display impressive and asks when families could use the tool. The team has not yet tried it on unfamiliar letters.

File B
Independent test | six months later

Reviewers give Atlas 100 sentences taken from letters it has not seen. It translates 31 accurately. It especially struggles with unusual phrasing and messages about appointments. The reviewers point to limited computing power and too few varied examples to train or test with. They do not say translation by computer is impossible. They say that the demonstration did not show it was ready for real school letters.

File C
Funding committee minute

The promised two-year deadline passes without a dependable service, and the committee cuts this project's funding. Other research continues. Falling confidence and investment after over-promising is sometimes called an 'AI winter'. The school still wants to reach families in different languages, but qualified translation checking costs money. It must decide whether to fund a narrow pilot, demand another test, or spend the budget elsewhere.

Should the school use Atlas to translate letters to families?

Now imagine: Atlas translates an appointment letter but gives the wrong time. The staff member checking it cannot read the translated language.

Optional multiple-choice questions
Which results were reported in the two tests?
What did the funding committee do after the deadline passed?

Answer all three questions to continue. Brief notes are fine.

Need a nudge?

Compare the narrow demonstration with the independent review and what happened afterwards.

Learning from examples

Two school email filters handle the same messages differently. Work out what each method gains and misses.

Three files

File A
How the two filters work

Filter One blocks messages only when they contain phrases on a list written by staff. Filter Two was trained using messages that staff labelled 'spam' or 'not spam'. It looks for patterns in those examples. When someone sends spam using new wording, Filter One misses it. Filter Two catches some of it. Staff still have to decide whether the labels and examples they supplied were good enough.

File B
Trial results | 50 new messages

In a fictional test, Filter One blocks 12 of 20 spam messages and incorrectly blocks one of 30 genuine messages. Filter Two blocks 17 of 20 spam messages but also blocks five genuine messages, including three club newsletters. The team calls Filter Two 'more effective' because it catches more spam. The school office asks whether missing a genuine club message also matters.

File C
Training examples audit

The learning filter's examples contain many sales messages but few genuine newsletters. That could help explain blocked newsletters, but has not been proved to be the only cause. The team could add examples and test again. Keeping blocked messages for review would avoid permanent deletion, but the office currently checks that folder only at 4 pm. Some club messages need a reply earlier in the day.

Optional: read more about this topic

What do AI and machine learning mean?

Raspberry Pi Foundation, Experience AI | 3-minute focused read

Their explanations for ages 11 to 14 distinguish systems built from rules from systems that learn patterns from example data.

Investigate: How does that distinction help explain the two filters? What does it not tell you about Filter Two's reliability?

Read at Raspberry Pi Foundation, Experience AI (opens a new tab) ↗

If this site is blocked, use the short description and the files.

Which email filter should the school use?

Now imagine: A real club message gets blocked. Pupils need to reply by 2 pm, but staff only check blocked messages at 4 pm.

Optional multiple-choice questions
How was Filter Two built?
What did the trial report?

Answer all three questions to continue. Brief notes are fine.

Need a nudge?

Compare the source of each filter's behaviour and check which messages are missing from the examples.

The convincing answer

An AI-generated fact box sounds confident. Check its separate claims instead of accepting or rejecting the whole thing at once.

Three files

File A
ORACLE fact box | fictional town archive

A teacher asks ORACLE for a short introduction to a local-history project. It replies: 'The town archive opened in 1912 and lists 840 maps. A 2022 study by the Centre for Civic Memory proved that every map is accurate. Pupils can safely copy details from any map without checking them.' The answer sounds ready to use. It gives no link or page number for the claimed study.

File B
Archive catalogue search

The catalogue says the archive opened in 1912 and lists 840 maps. The number describes items in the collection, not how accurate their contents are. A search of the catalogue and its linked publications finds no study with the title or research centre that ORACLE named. That search is not proof that no such study exists anywhere, but the class has not found a source that supports the claim.

File C
Archivist's advice

The archivist explains that some maps advertised proposed roads rather than roads actually built. Those maps can still show what planners hoped to do. The teacher has one lesson available and wants pupils to begin comparing sources. One pupil wants a simple fact box; another wants to inspect original maps. The class has not yet checked the date or purpose of the map chosen for the first assignment.

Optional: read more about this topic

Checking AI-generated information

BBC Bitesize Guide to AI | 2-minute focused read: see the source-checking tips

The BBC advises checking sources rather than relying on a single piece of convincing-looking AI-generated content.

Investigate: Apply that advice to the unverified study in the archive. Which claim needs another source?

Read at BBC Bitesize Guide to AI (opens a new tab) ↗

If this site is blocked, use the short description and the files.

Should pupils use ORACLE's local-history fact box?

Now imagine: A pupil finds a map showing a road and writes that the road existed. The map actually shows a planned road that was never built.

Optional multiple-choice questions
Which details does the catalogue confirm?
What did the catalogue search establish about the claimed study?

Answer all three questions to continue. Brief notes are fine.

Need a nudge?

Separate the confirmed date and count from the unverified study and the archivist's warning about maps.

Pair case file

Your answers for the approval hearing.

All stages are shown, including unfinished answers. Return to a stage to edit or print this case file.

Try the optional companion escape mission → Continue to the other room →

Project ORACLE / The AI Archive

Year 9 paired investigation / CodeBash / Stages complete: 0 of 5

Stage 01 / Can ORACLE be trusted with school deadlines?

Should pupils use ORACLE to find out homework and trip deadlines?

What we found

What the school should do

New problem

A teacher changes the homework deadline this morning. ORACLE still has the old date. A pupil asks it when the homework is due.

Would our plan work?

Stage 02 / The rule machine

Which questions should the school help desk answer?

What we found

What the school should do

New problem

A pupil cannot log in. The help desk does not recognise their message. Its 'Ask a person' button is on a page that needs a login.

Would our plan work?

Stage 03 / The abandoned laboratory

Should the school use Atlas to translate letters to families?

What we found

What the school should do

New problem

Atlas translates an appointment letter but gives the wrong time. The staff member checking it cannot read the translated language.

Would our plan work?

Stage 04 / Learning from examples

Which email filter should the school use?

What we found

What the school should do

New problem

A real club message gets blocked. Pupils need to reply by 2 pm, but staff only check blocked messages at 4 pm.

Would our plan work?

Research note

Research lead: Raspberry Pi Foundation, Experience AI | https://www.raspberrypi.org/blog/explaining-ai-terms-young-people-educators

Stage 05 / The convincing answer

Should pupils use ORACLE's local-history fact box?

What we found

What the school should do

New problem

A pupil finds a map showing a road and writes that the road existed. The map actually shows a planned road that was never built.

Would our plan work?

Research note

Research lead: BBC Bitesize Guide to AI | https://www.bbc.co.uk/bitesize/articles/z3ms7v4

Project ORACLE / Case file for the school approval hearing.