Part of the Level 0 core path · Lesson 3 of 5
What AI Can't Know (And How to Check)
After this, you'll be able to identify which parts of an AI response to trust and which to double-check.
Before you start
Before diving in, complete Context Is Your Superpower so you have the prompt-writing habit this lesson's verification exercises depend on.
The idea
Claude and ChatGPT learned from a huge amount of text up to a certain date, called the training cutoff. After that date, nothing. They do not browse the internet in real time unless the tool explicitly says it does, and even then only when you ask it to search. This means anything that happened recently, any statistic that changes over time, any specific name or number or citation, is something the AI is producing from memory, not from a live lookup.
| Before | After | |
|---|---|---|
| Habit | Guess from a loose request | Use the lesson move |
| Work move | Skip What AI Can't Know (And How to Check) | Apply What AI Can't Know (And How to Check) |
| Check | No clear proof | Pass the lesson check |
The after column is the lesson target.
The deeper issue is that the AI does not know when it is wrong. It produces fluent, confident-sounding answers whether it is right or making something up entirely. This is called hallucination, and it is the main risk for anyone new to these tools. The output always sounds plausible. That is exactly what makes it dangerous to trust without checking.
Here is the before and after: Someone asked Claude for the citation for a research paper on remote work productivity. Claude gave a real-sounding author name, journal, year, and volume number. The details were invented. The paper did not exist. The person only found out because they went to look it up.
The rule is simple: use AI to think through ideas, draft text, and work out plans. When you need a specific fact, statistic, date, name, or citation, verify it with a source you trust. Ask Claude what to search for if you are not sure where to start. That is exactly what it is good at.
Try it (10 min)
Watch out for
- Trusting AI on recent events. If something happened in the last year or two, verify it. The training cutoff is real.
- Trusting specific numbers, statistics, or citations without checking. These are the most common hallucination targets.
- Assuming confidence means accuracy. Claude is always confident. That is how it is designed. Confidence is not a signal of correctness.
- Skipping the 'how confident are you' follow-up. Asking Claude to rate its own certainty often surfaces real warnings it did not volunteer.
Paste this into Claude
Tell me [a specific fact, name, date, or number you can independently verify]. Then tell me: how confident are you in that answer, and what should I double-check?
What a good response looks like
The minimum wage in California as of my training data is $16.00 per hour for most workers, which took effect January 1, 2024. However, I want to flag a few things you should double-check before relying on this. First, my training has a cutoff date, so if there have been increases or ballot measures since then, I would not know. Second, some cities in California, like San Francisco and Los Angeles, set their own minimum wages above the state level. Third, fast food workers in California were moved to a separate $20 minimum under AB 1228, so the number depends on the industry. I'd recommend verifying at the California Department of Industrial Relations website (dir.ca.gov) before using this in any official document or decision. My confidence on the baseline $16 figure is moderate. My confidence on it being current today is low.
What good looks like
- You got an answer from AI on a specific fact
- You checked it against a source you trust
- You asked AI to flag its own uncertainty, and noticed whether it did
- You identified at least one case where the AI's answer sounded confident but needed verification
- You can describe in one sentence the difference between what AI is good for and what needs a source check
When this breaks
- Breaks when you treat the AI as a source of record rather than a thinking partner, because confident-sounding fluency masks fabricated facts. The output reads identically whether the answer is correct, partially right, or fully invented.
- Breaks when stakes are high and you skip the verification step. Citations, legal language, medical advice, and recent statistics are exactly where hallucinations cost the most, and exactly where users are most tempted to trust the smooth-sounding reply.
- Breaks when you assume self-rated confidence is reliable. The model can mark itself uncertain on something it has right and confident on something it invented. Self-rated confidence is a hint, not a verdict.
AI can help with this
After getting any factual claim from Claude, ask it: 'What is the chance this is wrong? What would you need to look up to be sure?' It will tell you where it is uncertain. You do not have to guess which parts to verify. Ask it to flag them.
You can now
Identify at least three categories in any AI response (recent events, specific numbers, named citations) that need outside verification before you would use the answer in a real document or decision.
Key takeaways
AI is a thinking partner, not a source of record. Use it to reason, draft, and plan. Verify the facts yourself. You are the editor.
- AI is a thinking partner, not a source of record. You stay the editor.
- Specific facts, statistics, citations, and recent events are the highest-risk targets for hallucination.
- Confidence in tone is not the same as accuracy. The output sounds equally smooth when it is wrong.
- Ask Claude to rate its own certainty. It often flags real uncertainty it would not volunteer.
- Use AI for reasoning and drafting. Use a trusted source for the final number, name, or quote.