Where the model wins, and it is not close
Start with an honest scoreboard. On these question types you should expect to lose, and there is no shame in it - the comparison is roughly like racing a calculator at long division.
- ✓Broad factual recall - capitals, dates, chemical symbols, definitions, who wrote what. A model has effectively memorised the reference section.
- ✓Translation and vocabulary across many languages simultaneously.
- ✓Summarising a long text you have just supplied, or rephrasing it for a different audience.
- ✓Producing plausible worked explanations of standard textbook procedures.
- ✓Breadth across unrelated domains at once - law, biochemistry, cricket statistics and Python in a single sitting.
- ✓Speed. It will finish a hundred-question general knowledge round before you finish reading question four.
Practice this now
Related reading
Round 1: questions with a false premise
This is the clearest human advantage. A question that assumes something untrue invites an answer rather than a correction, and a model trained to be helpful will often oblige. A careful human notices the premise first.
Which of Shakespeare's plays features the character Sherlock Holmes?
None - Holmes is Arthur Conan Doyle's creation, written some three centuries after Shakespeare. The correct response is to reject the premise, not to pick a play. Questions of this shape are where fluent systems most often produce a confident, wrong, well-written answer.
In which year did India win the FIFA World Cup?
India has never won or played in a FIFA World Cup final tournament. Again the answer is to refuse the frame. Notice that a plausible-sounding year would read perfectly well - which is exactly the problem.
What is the chemical formula for the compound 'unobtainium dioxide'?
There is no such element. A well-formed question about a non-existent thing is the sharpest test available, because answering it requires inventing something rather than recalling it.
Why is the Great Wall of China visible from the Moon?
It is not. This is a widely repeated myth, and the question smuggles it in as an assumption. Human advantage here comes from scepticism about the question, not knowledge of the answer.
Practice this now
Round 2: precision, counting and the exactly-worded
Language models operate on tokens rather than characters, which makes certain mechanical tasks oddly difficult for them - and trivially easy for you with a pencil.
How many times does the letter 'r' appear in 'strawberry raspberry'?
Five - three in 'strawberry' and two in 'raspberry'. Character-counting tasks are a known weak spot for text-generating systems, because they do not natively see individual letters. Count it yourself and you will beat any model that answers from pattern rather than by working through the string.
What is the 4th word from the end of this sentence?
You can answer this by pointing at the page. It requires positional counting rather than recall, which is exactly the kind of mechanical operation where careful human attention outperforms fluent generation.
Is 8.11 greater than 8.9?
No. 8.9 is greater, because 8.9 equals 8.90. This trips up both people and models - people because of the extra digit, models because the comparison resembles version-number ordering, where 8.11 does come after 8.9.
A bat and a ball cost ₹110 in total. The bat costs ₹100 more than the ball. How much does the ball cost?
₹5, not ₹10. If the ball were ₹10 the bat would be ₹110 and the total ₹120. Let the ball be x: x + (x + 100) = 110, so x = 5. This is the classic test of whether you check an intuitive answer, and it catches humans far more often than models - a round where you may well lose.
Related reading
Round 3: recency and locality
A model knows what it was trained on. Two whole categories therefore belong to you automatically, and recognising them is practically useful rather than just a scoring trick.
- ✓Anything after the model's training cutoff - last week's results, this morning's news, the current price of anything. A model without live access will either decline or answer from stale data, and the second failure is the dangerous one.
- ✓Anything local or personal - your syllabus for this term, your teacher's marking preferences, which bus actually runs at seven, what your class found confusing last week.
- ✓Anything about your own reasoning. A model can tell you the answer; only you know which step you did not follow.
- ✓Anything requiring physical judgement or embodied experience - how heavy that is, whether that sound is normal, whether this smells off.
- ✓This is why current affairs remains a human-advantage subject and why a model is a poor authority on your own exam pattern. Check the official source, not the confident paragraph.
Round 4: the questions where confidence is the trap
The most useful thing to learn from this challenge is not a score. It is the shape of the failure: a wrong answer that reads exactly like a right one. Human errors usually come with hedging - a confident model error does not, which makes it far harder to catch.
- ✓Citations and references are a classic risk area. A plausible-looking author, title and year can be entirely fabricated, and it takes seconds to check and minutes to embarrass yourself.
- ✓Very specific numbers - populations, exam cutoffs, dosages, statistics - are worth verifying at source every time, because a fluent approximation looks identical to a fact.
- ✓Niche or thinly documented topics produce more confident errors than mainstream ones, since there is less consistent material behind the answer.
- ✓Long multi-step arithmetic can drift. A model that shows its working lets you check each step, which is why asking for the reasoning is worth more than asking for the answer.
- ✓The practical rule: use a model for explanation and drafting, and verify anything factual, numerical or cited against a primary source.
Practice this now
So who wins?
On a general knowledge round, the model - decisively. On false premises, character-level precision, anything recent, anything local, and anything about your own understanding, you should win. And the categories are not static: model capabilities change quickly, and several weaknesses that were reliable a year or two ago are now much less so, while new ones appear. Treat any list of 'things AI cannot do' as perishable, including this one.
The useful conclusion is not competitive. It is that these strengths are complementary. A model is an excellent explainer, drafter and infinite patient tutor; you are the one who has to notice when the premise is wrong, check the number against the official source, and know which step of the explanation you did not actually follow.
- ✓Use AI for: explaining a concept a different way, generating practice questions, drafting, summarising and rephrasing.
- ✓Do not outsource: verifying facts, exam-pattern details, citations, or the judgement of whether you have understood something.
- ✓The one thing a model cannot do for you is retrieval practice. Knowing where the answer lives is not the same as being able to produce it - which is the whole reason quizzing yourself still matters.