AI Glossary / Governance, ethics and safety / Intermediate

Data proxy

Definition

A data proxy is observed information used to stand in for another quantity or characteristic that is difficult, unavailable, or impossible to measure directly. Its usefulness depends on the relationship between the proxy and the intended target. This entry uses the statistical and measurement sense of the term, rather than a software service that intermediates access to a database.

What it means

Organizations often care about concepts that are not directly recorded. They may use repeat purchases as one indicator of customer satisfaction, or completed practice exercises as an indicator of learning activity. Such measures can be informative, but they are not identical to the concepts they represent. Repeat purchases may reflect lack of alternatives, and completed exercises may not show understanding. A proxy makes a problem measurable at the cost of introducing assumptions. Those assumptions should be stated and tested rather than hidden inside a model's feature list or target definition.

How it works

Begin by defining the target concept and why direct measurement is unavailable. Identify a measurable candidate with a plausible relationship to that concept. Where reference measurements exist, examine how well the proxy tracks them across relevant settings and groups. Consider alternative explanations for the association and whether the relationship changes over time. A proxy can be used as an input feature, an outcome label, or an evaluation metric, with different consequences in each role. Optimizing a model against a proxy teaches it to predict or improve that recorded measure; it does not automatically optimize the underlying goal.

Three practical examples

• A service team uses reopened tickets as one indicator of unresolved problems, while recognizing that some customers may not reopen a ticket despite remaining dissatisfied.
• Climate researchers use natural records such as tree rings to investigate past conditions when direct instrument measurements are unavailable, with appropriate calibration and uncertainty analysis.
• A training platform uses quiz performance as evidence about knowledge, while checking whether the quiz samples the skills needed in practice. Each proxy requires validation for the particular inference being made.

Think of it this way

Think of judging how busy a building is by counting parked cars. The count may be useful, but public transport, car sharing, delivery vehicles, and empty parked cars can break the relationship. The analogy illustrates why a measurable signal is not the target itself. Real applications may combine several proxies and use a statistical model, but additional complexity does not remove the need to explain what makes the inference credible.

Worked example

Suppose a team treats a ticket marked closed as a proxy for a customer's problem being resolved. In a reviewed sample of 100 tickets, seventy are genuinely resolved according to a separate follow-up process and thirty remain unresolved. Of the seventy resolved tickets, sixty-five are marked closed. Of the thirty unresolved tickets, twenty are also marked closed. The closure rate is therefore 85%, while the reviewed resolution rate is 70%. Among closed tickets, 65/85, approximately 76.5%, are resolved. A model that accurately predicts closure may still be a poor tool for predicting resolution. The follow-up process also needs scrutiny, but the discrepancy shows why the proxy cannot simply be renamed as the desired outcome. If staff later receive incentives to close tickets faster, the relationship may change further.

Illustrative case study

Fictional case: Spruce Learning reports engagement using minutes spent on a lesson page. Instructors notice that some learners leave pages open while doing other work, while experienced learners finish quickly and perform well on practical exercises. The team retains time-on-page as an operational signal but stops treating it as a direct measure of learning. It adds reviewed task performance and investigates differences across lesson types. The case illustrates improving the interpretation of a metric without assuming that one replacement measure captures the entire underlying concept.

Common misconceptions and mistakes

Correlation does not make a proxy interchangeable with its target, nor does it establish a causal relationship. A proxy can work in one population and fail in another. Removing an explicitly sensitive attribute also does not guarantee that other variables cannot stand in for it, which can matter when assessing fairness. Conversely, not every correlated variable is inappropriate: relevance, purpose, and consequences require examination. Another mistake is reporting a proxy metric under the name of the desired outcome, such as calling clicks “satisfaction” without evidence that the substitution is valid.

Business use

Maintain a clear map from business objectives to measured variables. For each proxy, document what it is intended to represent, known limitations, validation evidence, and conditions requiring reassessment. Examine errors separately across important contexts rather than relying only on overall correlation. If incentives or automation change how people produce the data, recheck the relationship. When a proxy influences consequential decisions, provide reviewers with enough context to understand what is and is not measured. A model's high predictive accuracy on a proxy label should not be promoted as proof that it measures the underlying goal accurately.

People also ask

Is a proxy always a poor substitute?

No. Some proxies are well validated and extremely useful, especially when direct measurement is impractical.

Can a proxy be an output label?

Yes. Operational labels often stand in for harder-to-observe outcomes, which makes their meaning important during model development.

Does using several proxies solve the problem?

It can improve coverage, but the combination still needs validation and may share the same underlying blind spots.

Check your understanding

A recommendation system increases clicks while users report lower satisfaction. Has it necessarily improved the intended user experience?

Answer: No. Clicks are a measured behavior and may be only a proxy for the intended experience. Investigate why the relationship has changed and evaluate the actual objective.

Related terms

• Measurement Bias
• Correlation
• proxy variable
• measurement validity
• target variable
• construct
• surrogate metric
• fairness assessment

Learn next

Select one metric used in a familiar project and write the sentence “We use this to represent ...”. List three ways the metric could improve while the underlying goal gets worse. Then identify an independent check that would help detect that divergence.

More in Governance, ethics and safety

• Data point - A data point is one unit of information considered in an analysis or machine-learning task.
• Data poisoning - Data poisoning is deliberate interference with data used by an AI system in order to cause unwanted behavior.
• Data privacy - Data privacy concerns how information about people is collected, used, shared, retained, and otherwise processed, and how those activities affect the people...
• Data quality claim - A data quality claim states how well data meets a specified quality requirement.
• Data quality measure - A data quality measure is a variable used to express a measured aspect of data quality.
• Data quality model - A data quality model organizes the characteristics used to define and evaluate data quality.

Sources

NIST SP 1270, Towards a Standard for Identifying and Managing Bias in Artificial Intelligence, proxy terminology: https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=934464 NOAA National Centers for Environmental Information, What Are Proxy Data?: https://www.ncei.noaa.gov/news/what-are-proxy-data

Previous: Data proliferation Next: Data Quality