What AI Can't Do (Yet)
Knowing the limits makes you more effective. Here's where AI still fails.
It can't reliably do math
This surprises people. LLMs are trained on language, not mathematics. When asked to calculate, they're predicting what answer looks like — not computing it.
For simple arithmetic, they're usually right. For anything complex — multi-step calculations, exact percentages, financial projections — verify every number independently. Use a calculator or ask the AI to write code that performs the calculation.
It doesn't know what happened recently
Most LLMs have a knowledge cutoff — a date after which their training data stops. GPT-4o's cutoff is April 2024. Claude's is early 2025. Anything that happened after that date, they don't know.
For current information — news, prices, trending products, recent events — use Perplexity or enable web browsing if your tool supports it.
It doesn't remember between sessions
By default, every conversation starts fresh. The model has no memory of what you discussed yesterday, what your business is, or what preferences you've shared before.
Workarounds: use the "memory" features that ChatGPT and Claude offer, or start every session with a brief context block — your name, your business, what you're working on.
It can't take actions without tools
A base language model can only output text. It can't click buttons, send emails, browse websites, or book meetings on its own.
Agents extend this — they give AI tools to interact with the world. But a standard ChatGPT or Claude conversation is read-and-write only.
It can't read your mind
Vague input produces vague output. If you ask "write me a good email," you'll get a generic email. If you ask "write a 3-sentence email to a customer who hasn't paid their invoice in 30 days — firm but professional, no threats," you'll get exactly that.
AI is only as good as the context you give it. The more specific you are, the better the output.
It isn't always right about itself
AI models can be confidently wrong about their own capabilities, knowledge cutoffs, and limitations. They don't always know what they don't know. Treat their self-descriptions with the same skepticism you apply to everything else they say.
Use AI for generation, ideation, and drafting. Verify math. Use current sources for current facts. Give it context — it can't guess what you haven't told it.