Unit 8 · Technology & Future · Lesson 37
Working with AI
Learn to speculate about AI capabilities using must, might, could and can't. You'll evaluate AI output so you can discuss AI at work.
Warm-up · §1
5 minGet talking
What was the first AI tool you actually used? How did it change one small thing for you?
On a scale 1–10, how much do you trust AI to: write an email, give medical info, book a flight?
Name one task you'd never use AI for. Why?
Grammar focus · §2
8–10 minModals of speculation — must / might / could / can't
Quick rule
Certain YES → must. Certain NO → can't. Possible → might / may / could. Add 'have + past participle' for past.
Pattern builder
Present certainty
must / can't + base verb
e.g. That must be a hallucination.
Present possibility
might / could + base verb
e.g. It might be a caching issue.
Past certainty
must / can't + have + V3
e.g. They must have retrained it.
Past possibility
might / could + have + V3
e.g. The prompt could have been unclear.
Compare
| Modal | How sure? | Example |
|---|---|---|
| must | Almost 100% yes | That must be the right answer. |
| can't | Almost 100% no | It can't be right — dates don't match. |
| might / may / could | About 50% | It might be outdated. |
Examples
- →
That answer must be wrong — no source, no logic.
- →
The tool might be down; try again.
- →
It could be a bug in the update.
- →
That can't be the correct date — the event was in 2019.
- →
They must have trained it on old data. (past)
- →
The model might have hallucinated the source. (past)
Common mistakes
It mustn't be right.
✓ It can't be right.
For 'certain no', use 'can't', not 'mustn't' (which means 'not allowed').
It must have been outdated data. (No — you're sure)
✓ It must have been outdated data.
For past deduction, use must + have + past participle.
It could be a bug maybe possibly.
✓ It could be a bug.
'Could' already carries possibility — no need to pile on modifiers.
Natural English
In tech talk, native speakers often shorten: 'must be a bug', 'can't be right', 'might be down'. Dropping the subject is common in fast, informal analysis.
Deep dive (full explanation)
When we're not sure but want to say how likely something is, we use speculation modals. 'Must' = we're almost certain (positive). 'Can't' = we're almost certain (negative). 'Might / may / could' = it's possible. Use them for present and past.
Question 1.'That date ___ be right — the event is next month.'
Question 2.'The tool is really slow — the servers ___ be overloaded.'
Question 3.'The answer looks wrong — the model ___ hallucinated a source.'
Question 4.'I'm not sure why it crashed — it ___ be a memory issue.'
Question 5.Which is WRONG?
Build the sentence → spot the natural chunks → say it aloud → reply like a real conversation.
1.Rebuild the sentence — then say it aloud.
2.Rebuild the sentence — then say it aloud.
Quick check 1.'That date ___ be right — the event is next month.'
Quick check 2.'The tool is really slow — the servers ___ be overloaded.'
Vocabulary · §3
5–7 minWords & phrases to own
AI tool
software using artificial intelligence to help with a task.
to prompt
to give an AI an instruction or question.
prompt
(noun) the instruction you type into an AI.
to hallucinate
(AI) to invent information that sounds real but isn't.
chatbot
an AI that talks with you in text or voice.
to automate
to make a task run without human input.
bias
unfair preference built into a system.
training data
the material an AI learns from.
to fact-check
to verify a claim is true.
output
what the AI produces after you prompt it.
workflow
the sequence of steps you use to complete work.
to double-check
to verify carefully a second time.
Activate it now
Use each word about AI you've actually tried.
- Make it yours#1
What's one task you've automated (or would like to)?
to automateworkflow - Your view#2
Should AI output always be fact-checked, or only sometimes?
outputto fact-check - Rank#3
Rank most dangerous AI errors: hallucinated fact, biased answer, wrong tone, missed context.
to hallucinatebias - Complete#4
Finish: 'The best prompt I ever wrote asked the AI to ___.'
promptAI tool - Discuss#5
Are chatbots better teachers than search engines?
chatbottraining data
Tap an item on the left, then tap its match on the right.
Pronunciation · §4
3–4 minContracted modals in fast speech
Speculation modals contract heavily: 'must've been' → MUS-tuv-been; 'could've been' → COOD-uv-been; 'might've' → MITE-uv. Written English keeps 'have'; spoken English swallows it.
- It MUS-tuv been a bug.
- It COOD-uv been outdated.
- It MITE-uv hallucinated.
- It CAN-tuv been correct.
- They MUS-tuv retrained it.
Reading · §5
8–10 minThe Assistant That Sometimes Lies
Listen to the passage
Tap play to listen. Replay as many times as you need.
Ask a modern chatbot who won the 1974 World Cup and you'll get a confident, correct answer. Ask it the name of your neighbour's cat and it might just invent one — a plausible name, delivered with the same confidence. This is what engineers politely call a hallucination. The problem isn't that the tool lies. It's that it doesn't know when it's lying. The model was trained to sound helpful, and 'I don't know' rarely sounds helpful. So the output arrives smooth, fluent and — sometimes — completely fabricated. Good users adjust. They treat AI output like a first draft written by a very well-read but unreliable intern. They fact-check anything specific: names, dates, quotes, statistics. They ask the same question three ways. They notice when the tool's tone changes — over-confident answers can hide the weakest reasoning. AI won't stop hallucinating soon. What will change is us. The workflows that survive will be the ones which combine machine speed with human doubt.
Question 1.Why do chatbots hallucinate?
Question 2.What does 'good users' do?
Question 3.What warning sign is mentioned?
Question 4.The author predicts the future will require…
Q1.Chatbots know when they're hallucinating.
Q2.The writer suggests treating AI output like a first draft.
Q3.Confident answers are always reliable.
Listening · §6
8–10 minTwo colleagues discussing an AI report (about 1:30)
Listening audio
Tap play to listen. Replay as many times as you need.
Show transcript
Nadia:Look at this — the AI's summary says the meeting was in March.
Lee:That can't be right. The meeting was last week.
Nadia:So it must have pulled the wrong date from somewhere.
Lee:Or hallucinated it entirely. What was your prompt?
Nadia:'Summarise the attached notes.' Nothing fancy.
Lee:It might have confused two files. It's been glitchy all morning.
Nadia:Should I re-run it?
Lee:Re-run with the date in the prompt this time. And fact-check the names before you send it.
Nadia:Right. It's a good first draft. It just can't be trusted as final.
Lee:Exactly. Machine speed, human doubt.
Question 1.What's wrong with the AI summary?
Question 2.What does Lee suggest might have happened?
Question 3.What must Nadia do next?
Question 4.What's their overall attitude to the tool?
Exam skills · §7
5 minStrategy focus — Speculating in Speaking Part 2/3
Task
Examiners love speculation prompts: 'Why do you think this happened?' Weak answers use only 'maybe'. Strong answers vary modals.
Strategy
Stack three levels of certainty: 'It must be…' → 'Or it could be…' → 'It can't be…'. This shows range in ten seconds.
Example
The tool must have been trained on older data — otherwise the date would be right. It could also be a caching issue. What it can't be is intentional, because there's no reason for the system to lie about a date.
Practice · §8
8–10 minFill in the blank
Question 1.That answer ___ be correct — the numbers don't add up.
Question 2.The tool is so slow — the servers ___ be overloaded.
Question 3.The model ___ hallucinated the source.
Question 4.Always ___ AI output before sending.
Question 5.The AI's ___ made no sense.
Question 6.You need a clearer ___ to get a better answer.
Q1.Fill: 'It m___ have been a bug.' (certain past positive)
Q2.Fill: 'It c___ be a caching issue.' (possibility)
Q3.'To invent info that isn't true' = to h___.
Writing · §9
5 minPut it in writing
Your task
Write a 130–170 word short report on one AI tool you've used at work or study. Use at least THREE speculation modals and FIVE items of AI vocabulary.
Your answer
Show model answer
Over the past three months I've been using an AI writing assistant to draft first versions of internal emails. The workflow is simple: I prompt it with a rough note, it produces an output, I edit. The results have been genuinely useful, but not without problems. Roughly one in five emails contains a small hallucination — usually a fabricated name or a date that can't be right. The tool must have been trained on messy data, or it might have confused two similar files. Either way, I always fact-check specific claims before sending. On tone, it's surprisingly good. The bias I expected — overly formal, slightly robotic — is mostly gone. Occasionally the output could be described as bland, but that's easily fixed. Overall, I'd recommend it as a first-draft assistant, not a final one. Machine speed, human doubt: that combination must be the right way to use these tools for now.
Speaking · §10
10–15 minMake it a real conversation
ROLEPLAY — Explaining an AI mistake. Your teacher/partner plays a colleague and says: 'The AI report I sent has a wrong date — what happened?' Answer in 60–90 seconds using THREE speculation modals and FIVE AI vocabulary items.
Useful phrases
- • It must have ___
- • It can't be ___
- • It might have confused ___
- • The training data ___
- • You'll need to fact-check ___
- • As a first draft it's fine, but ___
- ColleagueThe AI report I sent has the wrong date — how did that happen?
- You_______________
- ColleagueCan I trust the rest?
- You_______________
- ColleagueSo should we stop using it?
- You_______________
Optional · Teacher-led
Teacher Activities
Optional group extensions. ~18 min total
Homework · §11
Take-homeTake it home
Write a 150-word review of an AI tool with three speculation modals.
Record a 60-second voice note explaining one AI error you've seen and why it happened.
Write true sentences with: prompt, hallucinate, fact-check, output, bias, workflow.
Watch a 3-minute AI news clip and note two speculation modals used.
Recap · §12
2–3 minWhat you've learned
- You can use must / might / could / can't for present and past speculation.
- You can talk about AI tools with precise vocabulary.
- You can explain and evaluate AI errors.
- You're ready for Lesson 38 — 'Smart Cities'.