B9 · AI, Society and the Future

How to Read an AI Product Claim

Reading an AI product claim means asking what a number such as '95% accurate' actually measures before believing it. Three key questions are: accurate at which task, tested on which examples, and who is affected, and how, when it is wrong.

10 min Beginner Jump to the activity

What you need: Nothing. This page starts from the beginning, with no programming and no formulas.

Definition

An ad says: 'CropCheck AI: 95% accurate disease detection!' Should a farmer trust it with this season's crop?

Reading an AI product claim means asking what a number such as '95% accurate' actually measures before believing it. Three key questions are: accurate at which task, tested on which examples, and who is affected, and how, when it is wrong.

In simple words

If a shop said 'our umbrellas are 95% waterproof', you would ask: in light rain or a storm? Tested by whom? What happens to me if it leaks? An AI claim deserves the same questions.

Why it matters

AI products are often sold with impressive numbers. A number can be true and still misleading if it was measured on easy examples, on different people or places, or on a task that isn't the one you care about. Asking the right questions protects your money, your time and sometimes your safety.

Key terms

Claim
A statement a seller makes about a product, such as how accurate it is.
Accuracy
The share of test examples the model got right.
Test examples
The cases used to measure a model; they should be like the cases where it will really be used.
Cost of a mistake
What happens, and to whom, when the model is wrong.
Independent testing
Checking done by someone who does not sell the product.

Explained in detail

Question 1: accurate at what task?

'95% accurate' only means something when you know the task. Does CropCheck detect every disease, or only three common ones? Does it work for tomatoes and also for cotton? Does 'accurate' mean that it named the right disease, or just that it said 'diseased' or 'healthy' correctly? A tool can be excellent at an easy task and poor at the one you need.

Question 2: tested on which examples?

A model is only as good as its tests. If CropCheck was tested on clear photos taken in a laboratory, it may do worse on blurry phone photos taken in a field at noon. If the test photos came from one region, it may struggle with crop varieties or diseases from another. Ask how many examples were used, where they came from and whether they look like yours. Also ask who did the testing: the company itself or an independent group?

Question 3: what happens when it is wrong?

Every model makes mistakes; what matters is their cost and who pays it. If CropCheck misses a disease, a farmer may lose part of the harvest. If it raises a false alarm, the farmer may spend money on pesticide that was not needed. These two kinds of mistake have different costs, and a single accuracy number hides which kind happens more often.

A trap: when 95% means nothing

Imagine that 95 of every 100 leaves tested were healthy. A tool that simply says 'healthy' every time, without even looking, would be 95% accurate and completely useless, because it would never find a disease. That is why a good claim also reports how many real problems the tool catches. In India, the Advertising Standards Council of India (ASCI) handles complaints about misleading advertisements, so a claim that cannot be backed up can be reported.

These questions work for any number in an advert, not only for AI: 'twice as fast', '10,000 happy customers', 'recommended by experts'. Each time, ask what exactly was measured, on whom, by whom and compared with what. A claim that survives these questions is worth considering. A claim that cannot answer them is just a slogan, however impressive the number looks.

How it works

  1. Find the exact claim and the number in it.
  2. Ask: accurate at which task, exactly?
  3. Ask: tested on which examples, how many, from where and by whom?
  4. Ask: what happens, and to whom, when it is wrong?
  5. Look for how many real problems it catches, not just one overall number.

Worked example

Claim: 'CropCheck AI: 95% accurate disease detection!' Question 1, the task: the small print says it was tested on tomato leaves for three diseases only, so it says nothing about the farmer's cotton. Question 2, the examples: the 2,000 test photos were taken in a laboratory under bright lights, while the farmer's photos will be taken in a field on a phone. Question 3, the cost of a mistake: a missed disease could cost part of the harvest, and a false alarm wastes money on spraying. Conclusion: the 95% may be true for the laboratory test, but the farmer should ask for results on field photos of her own crop before relying on it, and keep checking with an agricultural expert.

Try it yourself

Reading the ad

Follow a farmer asking the three questions.

Helpful question or not?

Which questions help you judge the claim?

  • What colour is the app's logo?

  • Is the ad's music catchy?

  • Who tested it: the company or someone independent?

  • What happens to me if it misses a disease?

  • Which diseases and crops was it tested on?

  • Were the test photos like the ones I will take?

A new case: '99% accurate attendance by face'

Question the claim on paper.

  1. Write the three questions you would ask the company.
  2. Who could be affected if the app marks a learner absent by mistake, and how?
  3. Rewrite the advert's claim so that it is honest and specific.

Think about it

  • Which missing detail in the claim worries you most?

Common mistakes

  • Trusting a percentage without asking what task and examples it was measured on.
  • Forgetting that a tool can score high just by always giving the most common answer.
  • Ignoring who pays the price when the tool is wrong.

Check yourself

Question 1

1. Which question best tests the claim '95% accurate'?

Question 2

2. 95 of 100 leaves are healthy. A tool always says 'healthy'. How accurate is it?

Question 3

3. Why do the test examples matter?

Question 4

4. New case: a homework app claims 'answers checked by AI, 98% correct'. What should you ask first?

Summary

  • Ask three questions of any AI claim: which task, which test examples and what happens when it is wrong.
  • A high accuracy can hide a useless tool, so ask how many real problems it catches.
  • Prefer independent tests on examples like yours, and report claims that cannot be backed up.

Finished reading?

Sources: Advertising Standards Council of India