Data: What Machines Actually Learn From

A model does not learn "reality"

A model does not learn "reality." It learns statistical structure available through particular data, objectives, and feedback.

That sentence is the whole lesson. The rest is what it forces you to notice.

AI is a category of machine-based systems. [nist-ai-glossary] Machine learning is one family of methods inside that category, not the definition of AI. This page is about the data those methods, and some hybrids, consume. It is not a claim that every AI system "learns from examples." For the methods that sit beside learning, see AI Is More Than Machine Learning.

Data is not the world

A spreadsheet of hospital visits is not "health." A photo dataset is not "vision." A pile of resumes is not "talent."

Sensors and Data already said the physical point: a sensor turns a property into a voltage or current, and the stored record is limited. It is not the whole event.

This lesson is the next step. Once a method starts learning from that record, the limits stop being a philosophical footnote. They become the thing the method can actually see.

The dataset is a sample from a process, not the process. A clinic's 2020 visits are not medicine. They are the visits that clinic recorded that year, under those billing codes, for the people who showed up.

Five pieces of a learning example

When people say "the data," they usually mean a bundle of five different things. Name them separately.

1. Example. One row, one image, one trip, one resume. The unit the method is shown.
2. Features. What you actually feed the method. A hospital visit might become a handful of numbers and codes. The raw chart is not what went in.
3. Label or feedback. The target, if there is one. "Spam." "Interview." "Tumor present." Unsupervised methods have no label. They still consume features from a sample.
4. Who sampled it, from where, when. A dataset is collected from some population, in some place, in some years. It is not "people" or "the world."
5. Who decided the label, and under what rule. A spam tag is a choice. A medical label depends on criteria and on who applied them. The target is not a fact of nature waiting in the file.

If you cannot name those five, you do not yet know what the system saw.

Objectives and feedback change what is "learned"

The same photos can train different systems.

Give the method "is there a tumor?" and you are asking it to pick up statistical structure that happens to correlate with that label in this sample. Give it "which hospital took the scan?" and the same pixels become a hospital detector. The pixels did not contain a single meaning waiting to be extracted.

The objective is the question you asked. The feedback is how you scored the answer. Change either one and you changed what counts as "learned," even if the files did not move.

That is why a fluent demo is not an evaluation. Fluency is a style of output. It is not evidence that the system saw the world you care about.

Hold some examples back

A common practice is to split examples into training, validation, and test.

• Training examples are the ones the method may look at while it fits.
• Validation examples, when used, are for comparing settings. They are not a trophy.
• Test examples are held out so you can notice memorization. They are a check for generalization: did the method pick up structure that still shows up on examples it was not shown?

The split is a method, not a guarantee. There is no ratio that is a law of nature. If test examples leaked into training, the check is fake. The score then describes memory of the test, not behavior on new cases.

This page will not walk through benchmarks, leaderboards, or experimental leakage. For how to inspect a claim about a named system, see How to Check an AI Claim. For how training actually runs, see Machine Learning Fundamentals.

Quality is not "more rows"

More rows can help. More rows can also repeat the same hole.

Watch for:

• Duplicates. The method saw one story many times and treated it as many stories.
• Missing groups. A clinic, a language, a job title, a neighborhood never entered the sample.
• Stale years. A 2020 file is not 2026.
• Scraped text with unknown license. You may not know who wrote it, who is represented, or whether you may use it.
• Labels from exhausted raters. People get tired. The "truth" in the file is still a decision they made under a rule.
• Synthetic rows generated by the same kind of model you will deploy. Generated examples are produced under a procedure. They are not automatically more representative of the world. They can fill a gap. They can also clone the generator's failures.

Provenance is the trail: where the data came from, how it was labeled, how it was filtered, who maintains it. Documentation of that trail matters. It is not something every lab publishes.

Everyday example: a hiring resume screener

A company wants a system that flags resumes to interview.

Each example is one resume. The features might be years of experience, school names, and keywords. The PDF the applicant uploaded is not what the method sees.

The label is "interview" or "no interview." That is a decision a recruiter already made, not a fact of nature that this person is qualified. Two recruiters can disagree. The rule might have been "looks like people we already hired."

Who was sampled? People who applied to this company in these years, not everyone who could do the job, and not people who never saw the posting.

A 2020 file is not 2026. Job titles moved. The school list changed. A system that looked sharp on last decade's applicants can fail on this year's.

Now add synthetic extra resumes, generated by a model, to "balance" the file. Those rows are not extra applicants. They are text produced under a procedure. They might cover a title the sample lacked. They might also photocopy the old hiring pattern and call it data.

The same logic holds for a toy photo task that asks "smiling" or "not smiling." A smile is a judgment. The sample is who was photographed. A 2020 photo set is not 2026. Extra generated faces are not extra people.

What you can say after this lesson

• A model does not learn "reality." It learns statistical structure available through particular data, objectives, and feedback.
• An example is a row, an image, a trip. Features are what the method is actually fed. Labels are decisions. Sampling and provenance say who is in the file and who applied the rule.
• Holding out a test set is a check for generalization. If those examples were in training, the check is fake.
• Quality is not more rows. Duplicates, missing groups, stale years, unknown scrapes, tired labels, and synthetic clones each change what the system can honestly be said to have seen.
• Related lessons: AI Is More Than Machine Learning for "learning is one method." Machine Learning Fundamentals for how training actually runs. Sensors and Data for the physical record. Hardware, Software, and Models for where a model sits. How to Check an AI Claim for a named-system claim. What is AI? for the category.