"Garbage in, garbage out" gets said a lot and taught rarely. It usually appears as a one-line caution before a class moves on to the interesting part. But for students trying to understand why AI systems behave the way they do, data quality is most of the explanation.
The thing students assume
Ask a class why an AI system made a bad decision and most will say the program was wrong, or not smart enough yet. The mental model is of a machine following instructions that someone wrote badly.
That model is wrong in a way that matters. Modern AI systems are mostly not given rules. They are given examples and asked to find the pattern. So when the output is wrong, the first question is rarely "what did the programmer say?" It is "what were they shown?"
A model doesn't know anything about the world. It knows about its training data. Everything it appears to know about the world is inherited from that.
Once students hold that idea, a lot of otherwise mysterious AI behaviour becomes predictable and, more importantly, criticisable.
Four ways data goes wrong
It helps students to have names for the failure modes rather than a general sense that data can be "biased".
Missing the people it will be used on
A system trained mostly on one group works worse for everyone else. This is the most documented failure in the field: facial analysis systems with far higher error rates for darker-skinned faces, voice recognition that struggles with regional accents, medical tools validated on one population and deployed on another. Nobody chose the outcome; the data simply didn't contain enough of the people it would later be applied to.
Recording the past faithfully
If a system learns from historical decisions, it learns the historical decisions, including the unfair ones. A hiring model trained on who was hired before will reproduce who was hired before. This is the subtlest one for students, because the data isn't wrong. It is an accurate record of something we don't want repeated.
Measuring a proxy instead of the thing
Data is often a stand-in for what we actually care about. Test scores stand in for learning. Clicks stand in for interest. Arrest rates stand in for crime rates. Every proxy carries assumptions, and a system optimising the proxy will happily walk away from the underlying goal.
Being simply wrong
Duplicate records, typos, broken sensors, mislabelled examples, someone entering height in centimetres into a column expecting metres. Unglamorous and extremely common, and in practice most of the work in any real data project.
Making it concrete
These ideas land much better when students test them than when they are told them. A few approaches that work without any specialist tooling:
- Build a deliberately skewed dataset. Have students collect examples of something with an obvious gap, such as photos taken only indoors or survey responses only from their own year group, then test how a simple classifier behaves on cases the data didn't cover.
- Audit a dataset before using it. Give a class a real spreadsheet and ask what's missing, who is under-represented, when it was collected, and who collected it. The answers are frequently more interesting than the analysis.
- Trace a proxy. Pick something a system optimises, engagement for instance, and ask what it is standing in for, then what happens when the two come apart.
- Break something on purpose. Introduce errors into a clean dataset and watch the output degrade. Students remember the shape of that curve.
- Who is in this data, and who isn't?
- Who collected it, when, and for what purpose?
- What is it actually measuring, as opposed to what it claims to measure?
- What would this data have looked like if it had been collected somewhere else?
These four questions transfer well beyond computing. They are the same questions a good history student asks of a source, and a good science student asks of a method, which is one reason data literacy sits comfortably across subjects rather than only in digital technologies.
Why this is the durable part of AI education
Specific tools date quickly. The model everyone is using this year will be superseded, the interface will change, and any lesson built around the particulars of a product has a short shelf life.
The reasoning does not date. A student who understands that a system's behaviour is downstream of its training data, and who instinctively asks what was in it, who was left out, and what was being measured, has something that will still be useful when the current tools are a footnote.
It is also the part that makes students harder to mislead. Most overclaiming about AI, in marketing and in the press, survives only as long as nobody asks what the system was trained on. Teaching students to ask is a reasonable definition of AI literacy, and it is a good deal more useful than teaching them which buttons to press.