What AI still can’t do well, and why it matters
AI keeps improving, but it still has weak spots, from recent news and exact sums to your situation and taking responsibility. Here’s what to watch for, and simple checks.
In 30 seconds
- Today’s weak spots include recent events, exact sums, consistency and knowing what it doesn’t know.
- It knows nothing about your situation unless you tell it, and it can’t take responsibility.
- Models improve fast, but keep people in charge of decisions about health, money, law and people.
Ask AI to draft a tricky email and it’s superb. Ask it to count the working days to a deadline and it may slip. Its skills are uneven, which researchers call the Jargon busterJagged frontier: Researchers’ name for AI’s uneven abilities: excellent at some tasks, surprisingly poor at others that look similar., and the weak spots aren’t always where you’d expect. Models improve quickly, so treat what follows as today’s weak spots, with habits that protect you whichever tool you use.
Today’s weak spots
- Recent eventsIts knowledge stops at a Jargon busterKnowledge cutoff: The date a model’s training data stops. It knows little or nothing about later events unless it can search the web.. Without web search, it may not know recent news, prices or rule changes.
- Exact sums and countsIt writes by predicting words rather than calculating, so long sums, dates and word counts can slip.
- Knowing what it doesn’t knowIt can’t reliably judge how much it knows, so it may answer confidently where it should say “I’m not sure”.
- ConsistencyAsk the same question twice and you may get two different answers, or two different verdicts.
- The physical worldIt has never tasted, touched or mended anything. It can explain a boiler fault but can’t see yours.
- Your situationIt doesn’t know your job, your workplace’s rules or what you agreed last week, unless you tell it.
Checks that protect you
| If it involves… | Try this |
|---|---|
| Anything recent | Ask it to search the web, and check the dates on its sources. |
| Sums, dates or counts | Ask it to use a calculator or run code if it can, and to show its working. Check the total. |
| An answer you’ll rely on | Ask again in a new chat, or ask for the strongest case against. If the answers differ, dig deeper. |
| Your circumstances | Give it the details that matter, and ask what else it needs to know. |
| A practical job | Use it to prepare, then trust your eyes, the manual or a qualified professional. |
Work this out using a calculator or code, not in your head, and show each step: [your calculation, for example 17.5% of £1,240 a month for 9 months]. Then give the final answer on its own line, and list any assumptions you made.
Try it yourself 3 minutes
- Pick a question that matters to you and has a clear answer, such as how a rule or a deadline works.
- Ask it in two new chats, or in two different assistants.
- Compare the answers. If they disagree, that’s your cue to check an official source.
Where people must stay in charge
AI can set out options and arguments, but it can’t weigh what matters most to you, and it can’t be held responsible. If a decision goes wrong, the person who made it answers for it, not the chatbot. Some AI companies say so too: Anthropic’s usage policy requires businesses using its AI for advice or decisions on health, law, money, jobs or housing to have a qualified professional review them first.
- Health: symptoms, medicines and doses. Use AI to prepare your questions, then ask a pharmacist, your GP or NHS 111.
- Money and the law: contracts, debts, tax and your rights. Check with an official source or a qualified adviser.
- Decisions about people: hiring, discipline, grades or benefits. A person should decide, and be able to explain why.
- Safety: electrics, gas, driving, or anything that could hurt someone.
- Anything in your name: you’re responsible for what you sign, send or submit.
When AI states something false with confidence, that’s a Jargon busterHallucination: When AI states something false as if it were true, because it predicts plausible words rather than checking facts.: here’s how to catch one. For weighing up big choices with AI’s help, see AI for decisions.
Check yourself
3 quick questions nothing is savedSources (1)
- Usage PolicyAnthropic, September 2025
Spotted a mistake? Tell us and an editor will check it.
