Completeness

Completeness means: nothing is missing that should be there. A class list is complete if every student is on it. A table of customer data is complete if every field in every row is filled in. In the computing world, completeness is a measurable quantity, not a feeling. One counts how many of the expected pieces of information are actually present. If 50 out of 1000 birth dates are missing, completeness stands at 95 percent. What matters here is always the question of what is expected in the first place — without that specification, nothing can be counted.

Gaps that don’t show up in the result

The most dangerous thing about missing data is that it remains invisible. A program keeps calculating even with incomplete information and delivers a number. That number looks just as solid as a correct one. The error only becomes apparent once someone checks the underlying basis.

For AI systems, this has a double effect. A model learns from the examples it is shown. If an entire group is missing from these examples, the model may later assess it poorly. Well-known cases involve facial recognition that worked reliably for light skin tones and noticeably worse for dark ones. The cause lay not in the program but in the incomplete training data.

Completeness also counts strictly in the financial sector. Anyone publishing a business report must provide certain mandatory information. If it is missing, the report is not merely thin but legally deficient. Auditors therefore check not only whether the figures are correct, but also whether all of them are present.

How it is counted and filled in

First, one defines which fields are mandatory. This rule is called a schema, i.e. a blueprint for the data. A program then checks every row against this blueprint. It counts the empty spots and outputs a rate. Such automatic checks run daily for large datasets.

There are three common ways to deal with the gaps found. One can discard the affected rows entirely, which costs information. One can estimate missing values, for instance using the average of all other values. Or one can explicitly mark the gap as unknown and deliberately not calculate with it. The third path is often the most honest, because it invents nothing.

A common misconception: completeness does not mean correctness. A field can be filled in and still be wrong. If a birth year reads 1900, the record is formally complete but nonsensical in content. Data quality therefore consists of several metrics side by side, including completeness, correctness, and timeliness.

Completeness in chatbots, contracts, and reports

The term is encountered most directly with language models like ChatGPT. When such a system has only learned fragments about a topic, it often fills the gap with a plausible-sounding invention. This is called a hallucination. Answers then appear complete even though the underlying basis was not.

In companies, completeness appears in contracts and audit reports. Database providers guarantee a certain completeness rate, for example 99 percent of all mandatory fields. If this is not met, contractual penalties may apply. Regulatory authorities, too, explicitly require complete datasets for reports submitted by banks.

In news reports, the term is usually read in connection with AI rules. The EU committee on Artificial Intelligence requires training data for high-risk systems to be complete and representative. Representative means: all groups that the system will later affect are also represented in it. Completeness has thus turned from a technical detail into a legal requirement.