What are AI tools actually bad at?
The failure modes that are structural rather than temporary, so you know where not to rely on them.
- Difficulty
- beginner
- Time
- 10 min
- Read
- 2 min
Short answer
Precise facts, arithmetic, anything recent, counting, and knowing what it does not know. It is strongest at language and structure, weakest wherever exactness matters.
Some limitations will improve and some are properties of how these systems work. Knowing which is which tells you where to keep checking rather than assuming the next version will have fixed it.
Step by step
- Expect weakness on precise facts.Dates, statistics, names, page numbers, citations. It produces plausible values rather than looking them up, and plausible is not the same as correct.
- Do not rely on it for arithmetic.It predicts text, so calculation is imitation rather than computation. Use a calculator or a spreadsheet — and note that some tools now run real code, which is more reliable when they do.
- Assume its knowledge has a cutoff.Unless it is actively searching, it does not know about recent events. It may answer anyway, confidently and out of date.
- Expect poor self-knowledge.It cannot reliably tell you what it does or does not know. Absence of hedging carries no information about accuracy.
- Watch for agreeableness.These systems tend to go along with the framing they are given. Push back, and they often reverse — which means agreement is weak evidence.
- Do not expect real reasoning about physical space or counting.Counting items, spatial relationships and precise logical constraints are common failure points, even in otherwise impressive output.
- Remember it has no stake in the outcome.It is not accountable for being wrong. For anything consequential, a human who is accountable should be in the loop.
- Note where it is genuinely strong.Explaining, drafting, summarising material you supply, restructuring, translating, brainstorming, and being a patient tutor. That is a lot.
Tips
- Match the checking to the stakes. Rewriting an email needs none; a legal or medical question needs a professional.
- Supplying the source material rather than asking it to recall shifts it towards its strengths and away from its weaknesses.
- Treat it as a capable, fast, tireless assistant who is sometimes confidently wrong and never says so.
Questions people ask
Will these problems be fixed in future versions?
Some improve steadily. But generating plausible text rather than retrieving verified facts is how the technology works, so verification remains sensible.
Can AI do maths?
Not reliably by itself, because it predicts text rather than calculating. Some tools now run actual code for calculations, which is much more dependable.