Where AI goes wrong, and how to check it
- Video
- 3 min
- Reading
- 13 min
- Exercise
- 12 min
- Total
- 28 min
Educational content, not legal advice.
In this lesson you will
- Name the four common ways AI output goes wrong and recognize each one in a work example.
- Check AI output with two questions: where each fact comes from, and who is affected if it is wrong.
- Match the depth of your check to the stakes, from a quick read to a second person.
- Use prompting habits that make errors less likely, without treating them as a substitute for checking.
Video lesson
Read the transcript
Welcome back to AI Literacy at Work. Lesson two: where AI goes wrong, and how to check what it gives you.
AI tools go wrong in four common ways. They make things up. They repeat bias from the data they learned from. They give outdated information. And they misread what you gave them, like a summary that drops the one sentence that mattered. None of these comes with a warning label. The wrong answer sounds as calm and sure as the right one.
Start with made up facts. A chat assistant can invent a court case, a quote, a statistic, or a policy that sounds real. It can give a web link that leads nowhere. This happens because the tool predicts likely text. A fake source looks just like a real one. So the rule is simple. Every fact that matters needs a source you can open, and you open it. If the tool gives no source, find one yourself, or take the fact out.
Next, bias. AI learns from past data, and past data carries past unfairness. A tool that ranks applicants or tenants can score some groups lower for reasons that have nothing to do with the job, or with paying rent on time. It can also treat people unfairly through language, for example a tool that reads accented speech less well. When AI output touches people, look at who it might leave out. You do not need to audit the tool. You need to notice, and tell its owner.
So how do you check? Use two questions. First, the source: where does each fact come from, and does that source really say it? Second, the stakes: who is affected if this is wrong? A summary for your own notes needs a quick read. A notice to two hundred tenants, or a decision about one person, needs a careful check, and sometimes a second person. Then you edit, correct, or ask.
There is a legal side too. Where PIPEDA applies, personal information has to be kept as accurate, complete and up to date as its purpose needs. PIPEDA Sch.1 cl. 4.9, 4.6 So an AI summary that gets a tenant’s history wrong is not just a typo when it goes into their file. At our fictional property manager, Quillfield, an AI meeting summary once put a complaint under the wrong unit. A manager caught it before it went into the file.
Recap: AI can be wrong, biased or out of date, and it will not tell you. Check the source, weigh the stakes, and you stay the author. In the next lesson: what never goes into an AI tool.
Four ways AI goes wrong
Lesson 1 showed that a chat assistant predicts likely words. That one fact explains most AI mistakes. They fall into four groups.
| What goes wrong | What it looks like at work |
|---|---|
| Made up facts | An invented policy, quote, figure, court case or web link that sounds real |
| Bias | Worse results for some groups of people, learned from past data |
| Outdated information | An old rule, price or contact described as current |
| Misreading | A summary that drops the key sentence, or mixes up two people or two units |
None of these comes with a warning. The wrong answer is written in the same calm, confident style as the right one.
Made up facts
People sometimes call these "hallucinations". The tool is not lying and it is not broken. It is doing what it was built to do: produce likely text. A made up source looks exactly like a real one because the tool learned what sources look like.
Typical examples at work:
- A policy that sounds like your company’s but is not in any of your documents.
- A "quote" from a law or a contract that does not appear in the text.
- A statistic with a precise figure ("63% of tenants prefer...") and no real study behind it.
- A web link or reference number that leads nowhere, or to a page on a different topic.
- A name or job title that is close to right but wrong.
The fix: every fact that matters needs a source you can open, and you open it. If the tool gives a link, click it and find the sentence. If it gives no source, find one yourself, or remove the fact.
Bias
AI learns from past data, and past data carries past unfairness. If a tool learned from years of decisions in which some groups were treated worse, it can repeat that pattern, quietly and at scale.
Where bias shows up at work:
- Ranking and scoring people: job applicants, rental applicants, customers for collections. A score can lean on things that stand in for age, origin or disability, such as a postal code, a gap in work history, or a name.
- Language: a tool may understand some accents, dialects or writing styles less well, and then treat those people’s requests as less urgent or less clear.
- Images and words: generated pictures and text can repeat stereotypes about who does which job.
You are not expected to audit a model. You are expected to notice. When AI output touches people, ask: who might this leave out, or treat worse? If something looks off, raise it with the tool’s owner.
Outdated information
Most models learned from data up to a certain date. Unless the tool searches the web or your documents, it may describe an old rule as current, quote last year’s price, or name someone who has moved on. It might not tell you that its knowledge stops at a date.
Be most careful with anything that changes: rules and laws, rates and fees, contacts, product details and recent events.
Misreading what you gave it
Even when you give the tool the right document, it can get it wrong:
- A summary leaves out an exception, a date or the word "not".
- Two people, two units or two invoices get mixed up.
- Numbers are rounded, added up wrongly, or copied from the wrong row.
- A translation changes the tone, or turns a request into a demand.
Long documents raise the risk. So does asking the tool to do several things at once.
Why people miss AI mistakes
The biggest risk is not the tool. It is the person who stops checking. When a tool is right most of the time, people relax and start to approve its output without reading it. This is sometimes called automation bias. It is strongest when you are busy, when the output looks polished, and when the tool has been right many times before.
The cure is a habit, not a mood: decide in advance which outputs get which check.
How to check: source and stakes
Use two questions every time.
1. Source: where does each fact come from, and does that source really say it?
- Open every link. Find the actual sentence.
- Check names, dates, amounts and unit numbers against your own system.
- Compare a summary with the original for anything that changes the meaning: exceptions, dates, amounts, "not".
- If the tool cannot point to a source, treat the fact as unconfirmed.
2. Stakes: who is affected if this is wrong?
| Stakes | Example | Check |
|---|---|---|
| Low | Notes for yourself, ideas, a first draft only you will see | A quick read |
| Medium | An email to one customer or tenant, an internal report | A careful read, every fact checked |
| High | A notice to many people, anything about money, safety or someone’s rights, a decision about a person | A careful check plus a second person, or no AI at all |
Then do one of three things: edit the output, correct it at the source, or ask someone who knows.
Checking is good practice, not a legal formula. But where PIPEDA applies, personal information has to be kept as accurate, complete and up to date as its purpose needs. PIPEDA Sch.1 cl. 4.9, 4.6 So a wrong AI summary that goes into a person’s file is more than a typo.
Writing better prompts reduces errors
You cannot prevent every mistake, but you can make them less likely.
- Give the tool the source text, and ask it to use only that text.
- Ask it to quote the sentence it relied on for each point.
- Ask it to say "I don’t know" when the text does not answer the question.
- Ask for one thing at a time.
- Ask it to list what it is unsure about. Then check those first.
These habits help. They do not replace checking.
A rent increase notice that almost went out wrong
Fictional organization. Quillfield Property Management Inc. is invented for this course. Any resemblance to a real company is a coincidence.
A property manager at Quillfield needed a notice for the 140 tenants of one building in Halifax, explaining the yearly rent review and the new laundry card system. She was short on time, so she asked the company’s approved chat assistant to draft it from her bullet points.
The draft was clear and friendly. It also contained three problems.
| What the draft said | What was true | Type of error |
|---|---|---|
| "Under provincial law, you get three months’ notice of this change." | Her bullet points said nothing about the law or the notice period. | Made up fact |
| "Laundry cards can be reloaded at the front desk or online." | Online reloading was planned for next year. | Outdated, taken from an old memo she pasted in |
| "The new rate applies to all units." | Her notes said the new rate applied to units on fixed term leases only. | Misreading: the exception was dropped |
How she found them. She used the source and stakes questions from this lesson.
- Source. She went through the draft line by line and asked "where did this come from?". Two statements had no source in her notes. One came from an old memo.
- Stakes. A notice about rent to 140 households is high stakes: it concerns money, it reaches many people and it is hard to take back. So she asked the leasing team lead to read the corrected draft before it went out.
What she changed. She removed the legal sentence entirely. Quillfield’s rule is that any statement about what the law says comes from the compliance manager, never from an AI draft. She deleted the online reloading line and restored the exception about fixed term leases. The second reader spotted one more thing: the date for the laundry change was a Sunday, when the front desk is closed.
What it cost. Twenty minutes of checking, against a notice that would have promised something Quillfield could not deliver and misstated the rent for some tenants. The draft still saved her an hour of writing.
Check a draft before it goes out
A maintenance coordinator asked a chat assistant to draft an email to a tenant whose dishwasher leaked into the unit below. She gave it the work order. Here is the draft:
> "Dear tenant, thank you for reporting the leak on March 3. Our records show this is the second leak from your unit this year. As your lease states, tenants are responsible for all water damage caused by appliances. A technician will visit on Friday between 9 and 11. Studies show that 70% of appliance leaks are caused by misuse."
The work order says: leak reported March 4, first report from this unit, technician booked Thursday 1 to 3 pm. The coordinator does not know what the lease says about water damage.
- List each statement you would check, and what you would check it against.
- Which errors can you already see?
- What are the stakes, and who else should look at the email before it goes out?
- Rewrite the risky sentences, or say what you would remove.
Show the model answer
- Statements to check: the date of the report (work order), "second leak this year" (maintenance history), the lease statement on water damage (the tenant’s actual lease, or the property manager), the visit time (work order or schedule), and the 70% statistic (a real source, if one exists).
- Errors already visible: the date is wrong (March 4, not March 3), it is the first leak, not the second, and the visit is Thursday 1 to 3 pm, not Friday 9 to 11. The statistic has no source and looks made up.
- Stakes: medium to high. The email blames the tenant for damage, which can lead to a dispute about money. The property manager or whoever owns lease questions reads it before it goes out.
- Rewrite: "Thank you for reporting the leak on March 4. A technician will visit on Thursday between 1 and 3 pm." Remove the statement about who is responsible for the damage until the property manager has checked the lease. Remove the statistic entirely.
Why this is the strongest answer
Four of the five factual statements were wrong or unsupported, yet the email read well. That is typical: AI drafts fail on exactly the details that matter most (dates, times, history, responsibility and figures), while the tone is fine. The strongest answer checks every one of those details against a source, removes what cannot be supported, and brings in a second person because the email assigns blame for money. Notice that the fix does not throw the draft away. It keeps the useful structure and replaces the facts.
AI output checking card
A pocket card for checking AI output: the source questions, the stakes table that sets how deep to check, the details AI most often gets wrong, and prompts that reduce errors.
Check your understanding
Sources cited in this lesson
- PIPEDA Sch.1 cl. 4.9, 4.6 official text current to 2026-09-30, checked on 2026-10-24