The tasks AI is genuinely bad at
Maths, live facts, and anything that needs to be exactly right.
Most "what AI can do" content is generous about AI's capabilities. This is the opposite. A clear, honest list of tasks AI is genuinely bad at right now, and likely to stay bad at.
Not "needs careful prompting." Not "improves with the right tool." Like, really bad. As in, you should usually do these things yourself or get someone else to help.
The mechanical limits
Two of these are simply how the system is built. Anything current is out of reach: models work from a training cutoff, and even the ones with web search bolted on are slower and less reliable than you checking a source yourself. Prices, regulations, this week's news, what a company has done recently. Look those up, then let AI do the summarising afterwards.
Arithmetic is the other. Language models predict text, they don't compute. Word counts come back 10 to 30% out. Multi-step sums drift. Comparisons across a long table quietly go wrong. AI can describe a calculation perfectly well. It just shouldn't be the thing performing it.
Genuinely bad at
Real-time information and anything after the training cutoff
Counting and exact arithmetic beyond simple sums
Strategic judgement in your specific context
Anything needing lived experience, humour included
Judging the quality of its own output
Work where being slightly wrong is worse than being obviously wrong
Do this instead
Use a search engine or a person, then let AI summarise
Use a calculator or spreadsheet. AI can describe the maths, not do it
Use AI to list the options and stress-test them, then decide yourself
Draft the scaffolding with AI, write the lived parts yourself
Get a human to review anything that matters
Build the underlying knowledge first, then use AI to amplify it
The most common failure isn't that AI is wrong. It's that the person using it can't tell.
The limits that don't move
The rest aren't engineering problems waiting on a better model. Strategic judgement in your context needs tacit knowledge of your industry, your team and your market. AI can lay out the considerations and stress-test your thinking, but the call on whether to take this particular client is not available to it.
Lived experience is the deepest one. It has read about everything and been through nothing, which shows up wherever credibility depends on having actually been there: the unspoken concern in a difficult conversation, why a particular team is dysfunctional, humour that lands rather than pattern-matching to funny.
And it cannot judge its own output. Ask whether a draft is good and it will say yes, warmly. Ask for the weaknesses and you'll get a list, which may or may not be the real ones.
The one that catches people
For most tasks an obviously wrong answer is harmless. You spot it and throw it away. The dangerous case is the subtly wrong one: legal wording, financial figures, contract language, safety information, code going to production. You accept it, and it ships.
Which is why the last item on the list is the one people skip, work you don't yet understand. AI amplifies what you know. Where you have no way to judge the output, it hasn't saved you anything. It has made the not-knowing harder to see.
Three things to do this week
01
Use a real source for anything current
Prices, regulations, this week's news. Use a search engine or a person, and save AI for the thinking around it.
02
Do the arithmetic somewhere else
Counting, tracking and exact figures go wrong quietly, and they go wrong with total confidence.
03
Build the understanding before you delegate the task
Use AI to amplify what you know, not to stand in for what you don't. If you can't judge it, it hasn't saved you anything.
Conclusion
01
Live information, counting and exact maths, and strategic judgement in your context.
02
Also lived experience, judging its own output, and anything you don't yet understand.
03
On these, AI isn't slow. It's the wrong tool entirely.
Go to next