Definition
If AI can beat world champions at chess, why can it not reliably count the fingers in a photo?
The benefits of AI are speed, scale, consistency and the ability to find patterns in more data than a person could ever read. Its limitations are that it depends on its training data, lacks common sense and real understanding, can be confidently wrong, can repeat unfairness in the data, and often cannot explain its answer.
In simple words
AI is like a student who has read a million books but has never stepped outside. Ask about anything in the books and the answer is fast and often excellent. Ask about something the books never covered and the student may invent an answer and say it with a straight face.
Why it matters
People make two opposite mistakes with AI: trusting it too much and trusting it too little. Knowing exactly what it is good and bad at lets you use it where it helps and check it where it is weak.
Key terms
- Benefit
- A useful result of a technology.
- Limitation
- Something a technology cannot do, or does poorly.
- Over-trust
- Relying on a machine more than its real accuracy deserves.
- Automation bias
- The habit of accepting a machine's answer even when there are signs it is wrong.
- Calibrated trust
- Trusting a system exactly as much as its tested reliability justifies.
- Brittleness
- Failing suddenly when a situation differs slightly from what the system was trained on.
Explained in detail
Benefits you can point to
The benefits of AI are not only promises. A blind person points a phone at a street and hears a description of what is ahead. A health worker in a small town screens patients for eye disease that would otherwise be found too late. A flood warning reaches a village days earlier than before. Scientists receive predicted shapes for more than two hundred million proteins, work that would have taken laboratories centuries, and use them to study diseases. A shopkeeper who never learned English reads a supplier's email in Marathi. In each case AI extended what people could do.
Limitations you can point to
The limitations are just as concrete. In 2016 a company released a chatbot that learned from the messages people sent it; within a day users had taught it to post abuse, and it was switched off. In 2017 researchers showed that a few stickers placed on a stop sign could make a vision model read it as a speed-limit sign, although any person would still see a stop sign. In 2023 lawyers in the United States were fined after submitting a court document that cited cases a chatbot had invented. None of these systems was broken. Each behaved exactly as its design and data allowed.
Where each limitation comes from
| Limitation | Its source | What follows |
|---|---|---|
| Fails on unfamiliar cases | It learned only from its training data | Test it on the cases it will really meet |
| No common sense | It has patterns, not experience of the world | Keep a person able to say "that makes no sense" |
| Confident errors | It produces the most likely answer, with no built-in check of truth | Verify anything that matters |
| Unfair results | Its data left people out or recorded past unfairness | Check results for each group |
| Cannot explain itself | Its reasoning is spread across millions of numbers | Avoid it where a reason is legally or morally required |
Costs that are easy to miss
- Energy and water: large models and data centres use a great deal of both.
- Fading skills: people who always use a calculator, a map or a writing tool may lose the ability to manage without it.
- Privacy: useful AI usually needs data about people.
- Concentration: a few large companies control the most powerful systems.
A four-question check
Before relying on an AI system for a task, ask four questions. How bad would a mistake be? Could the mistake be undone? Can I check the answer myself or with another source? Is a person responsible for the final decision? If a mistake would be minor, reversible and easy to spot, use the tool freely. If it would be serious, permanent or hard to detect, use the tool only as an assistant to a person who checks its work.
How it works
| Benefit | Example |
|---|---|
| Speed | Checks a payment for fraud in a fraction of a second |
| Scale | Translates millions of pages a day |
| Consistency | Applies the same check to the first item and the millionth |
| Pattern finding | Spots early signs of disease in scans |
| Always available | Answers a question at 3 a.m. |
| Limitation | What it looks like |
|---|---|
| Depends on its data | Fails on cases unlike anything it was trained on |
| No common sense | Suggests a route through a river because the map shows a line |
| Confidently wrong | Invents a fact or a source and states it smoothly |
| Can be unfair | Works worse for groups missing from its data |
| Hard to explain | Gives an answer but not a clear reason |
| Uses resources | Large models need a lot of electricity and water |
Two kinds of wrong trust
Over-trust means accepting what the machine says without thinking, even when your own eyes disagree. Drivers have followed navigation into lakes. Under-trust means refusing a tool that would genuinely help, such as ignoring a flood alert because "computers are always wrong". The skill is calibrated trust: more trust where the system has been tested and the stakes are low, more checking where it is untested or the stakes are high.
Try it yourself
The same AI, two very different days
A model was trained to recognise crop diseases from photos of leaves taken in daylight.
Benefit or limitation?
Decide what each situation shows about AI.
A voice assistant understands city accents but not village accents
A spam filter works all night without a break
A model notices a tiny pattern in weather data that people had missed
A model reads 10,000 X-rays overnight
A loan model rejects an application and cannot say why
A chatbot gives the name of a book that does not exist
Common mistakes
- Thinking AI is either brilliant or useless. It is strong at some things and weak at others.
- Thinking a confident tone means a correct answer.
- Using AI for a high-stakes decision without any human check.
Check yourself
Summary
- AI's strengths are speed, scale, consistency and finding patterns.
- Its limits come from its data: no common sense, confident errors, unfairness and weak explanations.
- Aim for calibrated trust: rely on it where tested, check it where the stakes are high.
Finished reading?
Sources: Elements of AI (University of Helsinki) · AI4K12: Five Big Ideas in AI