PSLE Science Reality Lab Vol No.045 | “95% Accurate” — Accurate on Which Data?

PSLE-SCI-REALITY-0045

Wait, What? A system can be 95% accurate and still be poor at the job you care about.

A website says a new image system is “95% accurate” at identifying whether photographed objects belong in Recycling Bin A or Bin B. Ninety-five percent sounds almost complete. It is easy to turn the number into a general belief:

“It gets recycling right 95% of the time.”

But a scientific accuracy percentage has a denominator and a testing history. Ninety-five percent of which cases? Were those cases used to build the system? Were they genuinely new? Were they easy studio photographs or messy real-world images? Did the test include every kind of object the system will later encounter? Were the mistakes spread evenly, or concentrated in one important category?

The percentage is evidence. The surrounding test design tells you how far that evidence can travel.

This is not a lesson in computer programming. It is PSLE Science reasoning applied to a modern communication object. You already know the central habits: ask what was tested, compare conditions, check the sample, separate observation from conclusion, and do not extend a result beyond the evidence that produced it.

Quick Answer

When you see “95% accurate”, ask four things: 95% of what test cases, were those cases genuinely separate from the cases used to build or tune the system, do they represent the conditions of real use, and what kinds of mistakes make up the remaining 5%? A high test percentage is useful evidence only within the test’s boundaries. It is not a universal guarantee.

Reality Lab rule: Accuracy belongs to a defined test, not to a system floating free of conditions.

The Learner Job This Reality Lab Owns

This article owns one real-world evidence-transfer job: how a Primary 5/6 learner should evaluate an accuracy claim for a scientific classifier, sensor-assisted system or AI model. It does not teach machine-learning algorithms or generic AI literacy. It does not replace eduKate owners of sampling, fair comparison, percentages, evidence quality or model limits. It applies those skills to the public claim “X% accurate”.

The Original Case: The Recycling Camera

A fictional school builds a camera system to classify objects into two groups:

Group R: accepted in a particular recycling stream;
Group N: not accepted in that stream.

Students first give the system 1,000 labelled photographs so it can learn patterns. Those photographs include bottles, cans, boxes and several non-recyclable objects. Later, a poster announces:

“Our model is 95% accurate.”

Before being impressed, you ask: How was 95% measured?

Case A: Tested on the Same Photographs It Learned From

The system was built using 1,000 labelled photographs. Its accuracy is then calculated by asking it to classify those same 1,000 photographs. It gets 950 correct.

The arithmetic is real: 950 ÷ 1,000 = 95%.

But this is a weak test of how the system handles new photographs. A learner who memorised the answers to a practice worksheet might also score highly when given the same worksheet again. That does not tell us how well the learner—or the classifier—will handle unfamiliar cases.

The system’s performance on its construction data may show that it can represent those examples. It is not yet strong evidence of performance on future objects.

Case B: Tested on New Photographs From the Same Setup

The school sets aside 200 photographs that were not used while building the system. The images were taken with the same camera, same table, same lighting and same kinds of objects. The model classifies 190 correctly: again, 95%.

This is stronger evidence. The test cases are genuinely new to the model. It has to apply what it learned rather than simply matching the exact construction examples.

But the real-world claim still has a boundary: new photographs under similar conditions.

Case C: Tested in the Real School Bin Area

Now the system sees crumpled containers, dirty packaging, objects partly hidden by hands, bright sunlight, shadows and unusual angles. The real school environment differs from the neat photo table.

If accuracy falls to 78%, which number is “the accuracy”?

Both numbers describe real tests. The 95% result describes one test distribution. The 78% result describes another. If the intended use is the school bin area, the second test may be more relevant to that use.

This is exactly why NIST guidance on trustworthy AI says accuracy measurements should be paired with clearly defined, realistic test sets and test methods, including whether they represent expected conditions of use.

Training, Tuning and Testing: Three Jobs

You do not need to memorise technical vocabulary, but you should understand the evidence jobs.

Data job	Simple question
Build / train	Which examples help the system learn or set its internal rules?
Tune / select	Which examples help developers choose settings or decide which version looks best?
Test	Which genuinely held-back examples estimate performance after those choices are finished?

If the same test cases are checked again and again while developers keep changing the system to improve the score, those cases can slowly become part of the tuning process. The “test” stops being fully independent.

NIST’s Artificial Intelligence Technology Evaluation programme addresses this kind of problem by using sequestered test data in a controlled evaluation environment, reducing the risk that test data leak into model development.

New Data Are Not Enough if They Are the Wrong Kind of New

Suppose the model’s 200 held-back test photographs are all clear images of clean bottles on a white table. They are technically new. But the real job is classifying mixed objects outdoors.

The test is independent but not very representative of expected use.

This gives us two separate gates:

Separation gate: Were the test cases genuinely held out from building and tuning?
Representativeness gate: Do the test cases resemble the conditions, objects and variation the system will actually face?

A strong accuracy claim needs both.

One Percentage Can Hide Very Different Mistakes

Imagine a test with 100 objects. Ninety are ordinary bottles and cans. Ten are unusual black plastic items that matter because the local recycling system does not accept them.

The classifier gets all 90 ordinary cases correct and only 5 of the 10 unusual cases correct.

Total correct = 95 out of 100, so overall accuracy = 95%.

That headline hides a weakness: the system is only 50% correct on the unusual subgroup. If that subgroup is scientifically or practically important, “95% accurate” gives an incomplete picture.

You do not need advanced statistics to see the issue. Ask whether the mistakes are evenly spread or clustered in one type of case.

The Denominator Is Part of the Claim

“95% accurate” means something like 95 correct decisions per 100 tested decisions, but only for the actual test set used. A percentage without the number and composition of test cases can hide fragility.

Compare:

19 correct out of 20;
190 correct out of 200;
9,500 correct out of 10,000.

All equal 95%, but the amount and diversity of evidence differ. More cases do not automatically make a test representative, yet the denominator still matters.

Duplicates Can Make “New” Data Less New Than They Look

Suppose the test set contains a photograph that is almost identical to one in the training set: same object, same background, same camera angle, perhaps cropped slightly differently. Technically the files are different. Scientifically, the test may be less independent than the label “new data” suggests.

This is similar to a school learner doing the same question with the numbers changed. Success can be useful, but it is weaker evidence of transfer than success on a genuinely unfamiliar representation or context.

Why “More Data” Does Not Automatically Solve Everything

If 100,000 training images mostly show one easy kind of object, adding more of the same kind may not prepare the system for rare but important conditions. NIST research on AI testing has examined how the composition and frequency of combinations in data can affect model performance. The key Primary Science habit is simple: count and variety are not the same thing.

The 95%-Accuracy Audit

What exact job was tested? Classification of what, under what conditions?
How many cases were tested? Find the denominator.
Were the test cases used during building or tuning?
Were there near-duplicates or strongly related cases across the split?
Do test cases represent expected real use?
Which kinds of cases were easy and which were hard?
What mistakes make up the remaining 5%?
Has performance been checked in the field, not only on a benchmark?
Is the public claim limited to the conditions actually tested?

Worked Case 1: The Reused Worksheet Problem

A classifier is built with 500 labelled images and scores 98% when tested on those same 500 images. A second system scores 92% on 200 entirely new images from a separate photo session. Which percentage gives stronger evidence about future new images?

The second. Its headline number is lower, but its test better matches the question “How well does the system handle images it did not build itself around?”

Worked Case 2: New but Unrepresentative

A leaf classifier is tested on new photographs, but every photograph is taken in bright daylight. It will be used in a greenhouse where many images are taken under dim artificial light. The test is separated from training, but what evidence is still missing?

Performance under the intended greenhouse lighting conditions. Independence alone does not establish real-world representativeness.

Worked Case 3: Same Overall Accuracy, Different Errors

System A and System B are both 95% accurate. System A makes five random mistakes across many object types. System B makes every mistake on one important category. Are the systems scientifically equivalent?

Not from the overall percentage alone. The pattern of errors changes what the number means for a particular use.

Worked Case 4: Benchmark Versus Field

A system is 97% accurate on a benchmark of clean images and 81% accurate during a field test with shadows and partial obstruction. An advertisement quotes only 97%. Which scientific question should a learner ask?

Which test conditions resemble the advertised use? If the field environment is the intended use, quoting only the cleaner benchmark can overstate expected performance.

Tempting Reasoning That Fails

“95% accurate means every future set of 100 cases will contain exactly five errors.” No. The percentage summarises a particular test; future performance can vary.
“A bigger training set always means a better model.” Data quality, coverage and relevance matter as well as amount.
“Held-out means representative.” A test can be independent yet unlike real use.
“Overall accuracy tells me which cases fail.” It hides the distribution of mistakes.
“If the field score is lower, the system cheated.” Not necessarily. The field may simply be harder or different from the benchmark.
“AI is special, so ordinary science rules do not apply.” The opposite: sampling, measurement, comparison, uncertainty and evidence boundaries remain essential.

What Evidence Would Strengthen the 95% Claim?

a clearly described test set kept separate from training and tuning;
enough cases to include meaningful variation;
test conditions that resemble expected use;
performance broken down across important subgroups or conditions;
documentation of test methods and mistakes;
field evaluation when the system is intended for real-world deployment;
repeat evaluation as conditions, equipment or data sources change.

PSLE-Style Transfer Case

A student creates a rule to identify whether seeds are likely to float or sink using their size and shape. She designs the rule after examining 40 seeds. She then reports that her rule is 90% correct because it correctly classifies 36 of those same 40 seeds.

Question: Suggest one way to test the rule more strongly.

Strong answer: Test the rule on a new set of seeds that was not used to create the rule, and compare its predictions with the actual floating or sinking results. The new seeds should represent the kinds of seeds the rule is supposed to classify.

Practice Lab

Practice A

A model scores 100% on 30 training cases and 84% on 100 new test cases. Which result is more useful for estimating performance on future unfamiliar cases?

Answer: The 84% result on new test cases, provided the test cases are relevant to future use.

Practice B

A classifier is 96% accurate on 1,000 indoor photographs but will be used outdoors at night. What is missing?

Answer: Evidence of performance under outdoor night conditions or a test set representative of those conditions.

Practice C

A system gets 950 of 1,000 cases right. All 50 errors occur on wet objects. What extra claim is now unsafe?

Answer: It is unsafe to say the system is 95% accurate under wet conditions. The overall score hides poor performance in that subgroup.

Practice D

A developer repeatedly checks the official test set and changes the model after each result. Why can the final test score become less independent?

Answer: Information from the test results is now influencing model choices. The test set has partly become a tuning guide rather than untouched evidence.

Delayed Independent Return

Next time you see an accuracy percentage for a scientific classifier, AI tool or automated detector, write the percentage down—then refuse to interpret it until you can complete this sentence:

“It was ___% accurate on ___ cases, which were / were not used during development, under ___ conditions.”

If the public claim does not tell you those things, you have found the missing evidence questions.

Where to Route Next

Teaching Guide for Parents and Tutors

Use familiar learning analogies carefully. Give a child ten practice questions and let them study the answers. Then ask whether a perfect score on those same ten questions proves mastery. Most learners immediately see why a new test is stronger. Transfer that insight to a scientific classifier: development examples and evaluation examples have different evidence jobs.

Then add the second distinction: new does not mean representative. A learner can succeed on ten new questions that are almost identical to practice and still struggle when the representation changes. Likewise, a classifier can perform well on clean new images and struggle in the field.

Finally, ask the learner to open one overall accuracy score and inspect the errors. This prevents the number from becoming a magic badge. The goal is not suspicion of AI. It is the same evidence discipline used in any good Science investigation: define the test, inspect the sample, examine the conditions, and limit the conclusion.

Authoritative Sources

The Quiet Return

“95% accurate” is not meaningless. It is incomplete by itself.

The scientific habit is to attach the number back to the test that created it: the cases, the conditions, the separation from development and the pattern of mistakes. Once you do that, the percentage becomes evidence instead of decoration.

Small Group Tutorials