What AI actually cannot do

Claude is useful enough that it is tempting to over-rely on it. It is also imperfect enough that over-relying on it will cost you. The most effective users of AI are not the most enthusiastic — they are the most accurate about where the limits are.

Here is a plain guide to what Claude genuinely cannot do.

It cannot look anything up in real time

Unless you are using a tool that has given Claude access to the internet, it cannot check current prices, today's news, stock quotes, whether a website is live, or anything that has happened after its training cutoff.

If you ask it a question that requires current information, it will either tell you it does not know or — more dangerously — give you a confident answer based on outdated information.

What to do: For anything time-sensitive, use Claude to help you think through the question, but verify the facts yourself from a live source.

It cannot remember your previous conversations

Each conversation with Claude starts fresh. It does not remember what you discussed last week, what preferences you have shared over time, or any context from a previous session.

What to do: Give Claude the relevant context at the start of each conversation. Many people keep a short "briefing document" they paste in when starting a new session.

It cannot reliably produce accurate numbers

Claude is not a calculator, and it is not a fact database. It can produce numbers that look plausible but are wrong — especially estimates, statistics, and calculations. This is not always obvious because the wrong numbers are often stated confidently.

What to do: For any numbers that matter, verify independently. Use Claude to structure the analysis, not to produce the final figures.

It can be wrong about recent events and real people

Claude knows a lot about the world up to its training cutoff — but less about recent events, and its knowledge of specific individuals (especially non-famous people) is limited. It can invent biographical details, confuse two people with similar names, or state outdated information about someone's current role.

What to do: Do not rely on Claude for factual claims about real people without verifying through a primary source.

It cannot read files, emails, or documents unless you paste them in

Unless you are using a version of Claude connected to your files, it cannot access your emails, Google Docs, Notion pages, or any other application. It only knows what you tell it in the conversation.

What to do: Paste the relevant text directly into your message. For longer documents, paste the section that matters most and describe the rest.

It cannot learn from your corrections (within a session)

If Claude gets something wrong and you correct it, it will adjust for the rest of that conversation — but not for future conversations. It does not update its underlying knowledge based on your feedback.

What to do: Expect to give context every session. Do not assume Claude has gotten "smarter" based on what you have told it before.

It can be confidently wrong

This is the most important thing to understand. When Claude does not know something, it does not always say so. It may produce a fluent, confident, plausible-sounding response that is simply incorrect. This is called hallucination.

It is more likely to happen when:

• The question involves specific factual details (dates, names, statistics)
• The topic is niche or recent
• The question is phrased in a way that suggests a specific answer is expected

What to do: The higher the stakes of the output, the more you need to verify it. Use Claude for drafting, thinking, and structure — not as the final authority on facts.

What this means practically

None of this means Claude is not useful. It means you should use it the way you would use a brilliant colleague who is new to the job: great at helping you think, drafting, structuring, and generating options — but you would not send their output directly to a client without reading it first.

The people who get the most out of AI are the ones who treat it as a capable first draft, not a finished product.

Further reading

• Models overview — what each Claude model is designed for and its limitations