When the training set tips over

CONCEPT

In Class 7, you named three places bias can enter an AI system. Data was one of them. This chapter makes that mechanism concrete.

Suppose one kind of training example is far bigger than another. The system leans on the bigger pile. Faced with something new or unclear, it defaults to whichever class it saw more of. That is not understanding the object. That is matching patterns.

An artists' group in Seoul tested this. They took photos of clouds that faintly looked like faces and fed the photos to an AI face-detector. The program marked every single cloud as a detected face. Its training data had far more lighter-skin-colour male faces in it than any other kind, so it had learned an over-broad "face" pattern — broad enough to catch a cloud.

The same shape appears with plain objects. Train an AI mostly on cricket-bat photos, with far fewer badminton-racket photos. On a new or unclear photo, it will guess "cricket bat" most of the time. It has no idea what either object even is. More examples of one class simply made it better at telling that class apart, and worse at the other.

The rule underneath: how right the system is tracks training-set balance, not the true count of the object in the world.

two worked cases, one mechanism: the majority class wins
Cloud photos fed to a face-detectortrained mostly on lighter-skin-colour male faces — every cloud is reported as a detected face
recognition accuracy tracks training-set balance, not real-world frequency
A new object photo, cricket bat vs. badminton rackettrained mostly on cricket-bat images — a new or unclear photo is guessed 'cricket bat' most of the time

↑ Back to top

Measured, not hypothetical

CONCEPT

The clouds and the cricket bats are worked examples, useful for seeing a mechanism clearly. What happened to computer scientist Joy Buolamwini is not a worked example. It is a real test result.

Buolamwini tested commercial face-recognition systems built by major tech companies. She found a real gap in how right they were. The systems worked well for light-skinned men. They made far more errors for women, and for people with darker skin tones.

The cause traced back to the same place it always does: the training data. The datasets these systems had learned from held far more photos of some groups than others. In one test, a system failed to detect Buolamwini's own face — until she wore a white mask.

This is not a made-up example. It is a measured fact: the imbalanced-training-set mechanism produced real, unequal treatment in systems that companies had already deployed.

a measured accuracy gap, traced to the training data
TEST — commercial face-recognition systems from major tech companies
FINDING — high accuracy for light-skinned men; substantially more errors for women and darker skin tones
CAUSE — training datasets contained far more photos of some groups than others
DEMONSTRATION — one system failed to detect Buolamwini's own face until she wore a white mask
a real, measured result — the imbalanced-training-set mechanism showing up in deployed systems, not only in a classroom example

↑ Back to top

The same mechanism, a different shape

CONCEPT

The same mechanism can hide inside something as ordinary as a translated sentence.

Take the Hindi sentence "वह डॉक्टर है" — roughly, "[they] are a doctor." Hindi's grammar does not require a gender here; the sentence itself is silent on it. Feed it to an AI translation system, though, and it comes back as "He is a doctor."

Here is the trap: a translation that reads fluently, in perfect grammar, must have understood the sentence correctly. It did not. The gender was never in the source at all. It was supplied by the training data — which had learned from more examples of male doctors than female ones — and inserted with total confidence.

This is exactly the kind of pattern Nila would catch. She sorts the evidence before she names anything, and a hundred "doctor" sentences turning out male would jump out at her fast. But spotting the pattern is only half the job. The harder question is different: is this pattern about the world, or about an imbalance in what the system was shown? Ask that before trusting what a fluent answer seems to confirm.

gender-neutral in, gendered out
BEFORE
Hindi source: "वह डॉक्टर है"grammatically gender-neutral — does not specify "he" or "she"
the training data's own imbalance supplied the gender, not the source sentence
AFTER
AI translation: "He is a doctor."gender inserted by the AI, traced to more male-doctor examples in training data

↑ Back to top

A worked case: your own school's sports AI

CONCEPT

Picture your own school's sports-recommendation AI, built the way most of these systems are: from a table of who played what.

The table says this. Cricket: 55 boys, 15 girls. Badminton: 10 boys, 20 girls. That is the entire record — nothing about who was actually good at either sport, only who turned up and played it.

Sort this table fast and you get: mostly boys chose cricket, mostly girls chose badminton. Done. That is the trap. A recommendation trained on real past data must reflect what students are actually good at — surely real numbers cannot lie.

But the table never measured ability. It only recorded which sport past students happened to enjoy, sorted by gender. Now a new girl joins the school. The system's recommendation follows the pattern it learned — boys toward cricket, girls toward badminton — not her own ability or interest. The table never recorded that in the first place, so the system has nothing of hers to go on.

the training table a sports-recommendation AI learns from
STUDENTS WHO ENJOYED IT →
SPORT ↓
Boys
Girls
a new girl student joins — the system's recommendation follows the table's boys-cricket/girls-badminton pattern, not her own ability

↑ Back to top

Wherever AI decides, the same risk follows

CONCEPT

AI systems now decide or advise in places far beyond face-detection and photo apps: suggesting videos, approving loans, helping doctors flag cases, selecting job applicants. Wherever training data carries an old imbalance, the same risk follows it into the decision.

Take hiring. A company built a recruitment AI from its own past hiring records. Most of the engineers it had previously hired were men. The system learned to favour resumes that looked like those men's — regardless of whether that likeness had anything to do with being a good engineer.

Here is the trap: surely a hiring tool that decides from data, not personal opinion, is more fair than a human recruiter. It is not, on its own. Data built from past human decisions carries those decisions forward. The AI did not choose to discriminate. It simply repeated a pattern already present in whom the company happened to hire before — without being told to correct for it.

AI should be built to help everyone. It should not simply repeat who was chosen in the past.

one mechanism, several deployed settings
VIDEO RECOMMENDATION — what gets suggested next
LOAN APPROVAL — who is judged creditworthy
HEALTHCARE SUPPORT — whose case an AI flags for a doctor
HIRING — a company's AI learns to favour resumes resembling its own past (mostly-male) hires
the AI did not choose to discriminate — it reproduced a pattern already present in the historical data it was given
CONCEPT

Recommendation is the same risk in a different outfit. A recommender learns from what people already watch, like, and share — and that habit can amplify an existing majority instead of serving everyone equally.

Take a video platform. Its recommender is built on watch, like, and share history. English- and Hindi-language videos already carry very high total view counts, so the system tends to suggest them first to new users. A video in a smaller regional language never had the same chance to build up views. It appears less often in recommendations — regardless of how good it actually is.

Watch what this does over time. Already-popular content gets shown more. Being shown more makes it more popular still. Meanwhile, a regional-language creator's work stays harder to discover, through no fault of the content itself. The loop that decides what gets seen started from an unequal head start, and kept that gap open.

how a popularity-trained recommender narrows toward the majority
TRAINING SIGNAL — watch/like/share history, dominated by high-view-count English/Hindi videosRECOMMENDATION — new users are mostly shown the already-popular languagesOUTCOME — regional-language videos appear less often, regardless of quality
OUTCOME — regional-language videos appear less often, regardless of quality → TRAINING SIGNAL — watch/like/share history, dominated by high-view-count English/Hindi videos  the loop closes
the loop reinforces itself: already-popular content is shown more, which makes it more popular still

↑ Back to top

Three strategies that make AI fairer

RECAP

None of this is hopeless. Three strategies make an AI system fairer, and none of them happens on its own — each takes a choice someone makes on purpose.

The first is a bias check. Test the AI program directly: does it treat different groups in different ways, rather than assuming it is neutral?

The second is human supervision. A person reviews the AI's decisions instead of trusting them blindly — this matters most wherever the decision affects someone's chances.

The third is transparency. Understand how the system actually works and how it reaches its decisions. Only then can a reviewer judge the reasoning behind an answer, not just the answer itself.

three deliberate strategies — none of them automatic
BIAS CHECK — test whether the AI treats different categories differently
HUMAN SUPERVISION — a person reviews AI decisions rather than trusting them blindly
TRANSPARENCY — understand how the system works and decides, so the reasoning can be evaluated
fairness does not happen automatically — it requires careful design, continuous testing, and responsible use
Bias and Fairness in AI — Grade 8 (CT & AI) · projected from the LATTICE via prism_html.py · register: school-g8

↑ Back to top