What it can't do

Now that you know how it works, its weak spots make sense. Here are the three that catch people out. (Forgetting isn't on the list. You saw in lesson 6 that it's the interface, not a flaw.)

It has a cutoff. It learned from text up to a date. For Claude Haiku 5.5, Anthropic gives a reliable knowledge cutoff of June 2026 (checked October 2026). Anything newer, it doesn't know unless you paste it in or a tool looks it up. This sandbox has no tools.

It can make things up. Remember the lossy zip from lesson 4. When the details are fuzzy, the odds go flat. But the next token still comes out, and it reads just as fluently as a sure thing. Here are illustrative odds for a book that doesn't exist.

Each pick was close to a coin flip, but the sentence sounds sure of itself. That's a hallucination.

Claude is trained to push back. Looking inside an earlier Claude, Anthropic's researchers found that “I don't know” is its default, and recognising something switches that off. It goes wrong when a name feels familiar but it knows nothing else about it.

It slips on maths and counting. It does sums “in its head” with shortcuts it learned from text, not with a calculator. The same study caught it adding two numbers by working out a rough total and the exact last digit in parallel, then combining them. Ask how it did the sum, though, and it describes carrying the 1, the method from school. The fix is to hand it a calculator: a tool it can call, which is what the Give your model tools course is about.

So check what matters. Ask it to say when it's unsure. Paste in the source and ask it to quote from it. Check names, numbers, quotes and links yourself. Use a calculator for maths. And if an answer changes between runs, treat it as a guess.

Your turn. Ask about something obscure or made up (a half-familiar book works well), and give Claude permission to say it doesn't know.