Day 6: When AI Gets It Wrong -- And What To Do

The Concept

Knowing what AI gets wrong is not pessimism. It is professionalism.

Careful AI users are not people who distrust the tool. They are people who know how to use it without becoming careless. They can enjoy the speed, creativity, and structure AI provides while still noticing when an answer needs checking.

That balance matters because AI's mistakes are often polished. A weak paragraph written by a person may look clumsy. A weak answer from AI may look confident, complete, and beautifully formatted. The danger is not that AI is always wrong. The danger is that wrong answers can look unusually usable.

Today's goal: build a simple verification reflex you can apply before relying on AI output.

You do not need to become a professional fact-checker. You need a practical habit: identify what matters, check the parts that could cause trouble, and treat polished language as a draft rather than proof.

The three common failure patterns

AI can fail in many ways, but three patterns matter most for everyday users.

1. Hallucination

A hallucination is when AI generates something false while presenting it with the same fluency it uses for something true.

It might invent a statistic, describe a source that does not exist, misstate a policy, give the wrong date, or blend true and false details together. The answer may not include any obvious warning sign. It may sound completely normal.

Plain English: A hallucination is a confident-sounding mistake.

This is why you should be careful with exact claims. If the output includes a number, a quote, a source, a legal rule, a medical statement, a policy detail, or a current fact, pause before acting on it.

2. Outdated information

AI may describe the world as it existed during training, not as it exists today.

Products change. Prices change. People change roles. Laws and policies change. Tools add or remove features. Even when the model is strong, it may not know the latest version of something unless the tool has live browsing or you provide current source material.

Why this matters: A good explanation can still be out of date.

For current information, use AI to prepare your search, compare sources you provide, or explain a document you already have. Do not assume it knows the latest state of the world.

3. Biased or incomplete framing

AI learns from human-created material, and human-created material contains blind spots.

This can show up in subtle ways. The model may assume a default audience, overlook a group of people, frame a situation from a dominant perspective, or make recommendations that sound neutral but are shaped by patterns in its training data.

This matters most when AI is helping you write about people, make judgments, evaluate candidates, summarise feedback, or make decisions that affect others.

Use this rule: When people are involved, check not only whether the answer is accurate, but whether the framing is fair.

What actually needs checking?

Not every AI output deserves a full investigation. That would make the tool exhausting.

The better question is: does this matter enough to verify?

Low-risk outputs usually need light review:

• brainstorming ideas,
• drafting a casual message,
• reorganising your own notes,
• making a checklist,
• simplifying a concept for your own understanding.

Higher-risk outputs need verification:

• facts you will publish,
• numbers you will use,
• advice that affects money, health, law, employment, or safety,
• current information,
• claims about people or organisations,
• anything a customer, colleague, or client may rely on.

The practical standard: Verify the parts that could create consequences if they were wrong.

That keeps verification realistic. You are not checking every sentence. You are checking the claims that matter.

The verification workflow

Use a two-step workflow.

Step 1: Ask AI what to check

Before leaving the chat, ask:

What parts of this answer should I verify before I rely on it?

This does not catch everything, but it usually surfaces the obvious risk areas. It also trains you to notice the categories of information that need extra care.

You can make the question more specific:

• "Which claims in this answer are factual?"
• "Which parts might be outdated?"
• "Which numbers should I recalculate?"
• "Which parts are assumptions rather than verified facts?"
• "What source would I use to confirm this?"

Step 2: Check one independent source

Once AI flags the parts worth checking, leave the chat and verify the claim elsewhere.

Good sources include:

• an official website,
• the original document,
• a reputable publication,
• a current product page,
• a calculator or spreadsheet,
• a qualified professional,
• or a source your organisation trusts.

The source depends on the claim. A product feature should be checked against product documentation. A policy should be checked against the policy document. A number should be recalculated. A legal or medical issue should be checked with a qualified source.

Important: Asking the same AI conversation again is not independent verification. It may repeat the same mistake with more confidence.

A simple example

Imagine AI drafts a short explanation of a new workplace policy.

The writing may be useful. But before sending it, you would check:

• Did it describe the policy accurately?
• Did it invent details that are not in the document?
• Did it soften or exaggerate any obligation?
• Did it include dates, numbers, or eligibility rules?
• Did it leave out an important exception?

You do not need to distrust the draft. You need to compare the parts that matter against the actual policy.

That is mature AI use: use the speed, then verify the stakes.

Use this today

Take one AI response you have received this week. It can be from this course or from another task.

Run today's verification prompt against it. Then do three things:

1. Identify the claims AI says are worth checking.
2. Choose one claim that actually matters.
3. Verify it using an independent source.

The point is not to catch AI being wrong. The point is to build the reflex. Once you practise this a few times, you will start spotting verification moments automatically.

Remember this

If you remember nothing else from Day 6, remember these three ideas:

• Polished does not mean verified.
• Check the claims that would create consequences if wrong.
• Ask AI what to verify, then confirm important claims independently.

This habit lets you use AI confidently without being careless. It keeps the speed while protecting your judgment.