AI Glossary / AI and machine learning / Intermediate

Proxy labels

Definition

Proxy labels are stand-in labels used for training when the true outcome you care about cannot be observed directly, so a related, measurable quantity is used in its place.

What it means

Real labels can be expensive, slow or impossible to get. You may want to predict who will be a good employee, but you can only observe who was promoted; you may want to predict health need, but you can only observe past spending. The measurable stand-in is the proxy.

Proxies are everywhere because they are practical. Clicks stand in for interest, repayment stand in for creditworthiness, and arrests stand in for crime. Each is easy to count and only partly the same thing as what you meant.

The gap between the proxy and the real target is where harm hides. A model trained faithfully on a proxy gets very good at the proxy, including every bias in how the proxy was produced, and may be quite poor at the real goal.

For a business, the question to ask of any trained model is simple: what exactly was the label, and is it what we actually want?

How it works

Name the real target in plain words, such as patients who will need extra care.

List what can actually be measured, and pick the closest available stand-in as the proxy label.

Test the proxy against whatever direct measurements exist, even on a small sample, to see where they diverge.

Check the proxy for bias: if one group is less likely to generate the proxy event for reasons unrelated to the target, the model inherits that.

Revisit the choice when better labels arrive, and monitor outcomes against the real goal, not just the proxy.

Three practical examples

• A health system predicts need for extra care using past healthcare cost as the label, though cost also reflects who had access to care.
• A content platform trains on clicks as a proxy for satisfaction, and surfaces attention-grabbing items that people regret.
• A hiring tool uses past promotions as a proxy for performance, reproducing whoever managers promoted before.

Think of it this way

Think of judging a student's understanding by the length of their essay. Length is easy to count and loosely related, but a model trained to maximize it will produce padding, and graders will wonder why quality fell.

Worked example

the cost that was not need

Suppose a hospital ranks patients for a care program using last year's spending as the label. The figures are illustrative.

Two patients are equally ill, but one had fewer visits because of transport and cost barriers, so the model scores that patient as lower need and leaves them out.

The fix is a better label: the team retrains on measured health indicators such as chronic conditions and test results, and checks that the program reaches both groups at similar rates.

The general lesson: a proxy can be accurate about itself and wrong about the goal, and the error falls hardest on people the proxy measures least well.

Illustrative case study

A national lender trained a model to predict good customers using past repayment history as the label. The model performed well on its metric, yet reviewers found it gave thin-file applicants, many of them young or new to the country, low scores because they had no history to repay, not because they would default. The lender added a pilot that approved a small, monitored sample of thin-file applicants, collected their actual repayment, and used that outcome data as a better label. Within a year the revised model approved more applicants at the same loss rate. The lender now documents the label for every model and requires a written note on how it differs from the real goal.

Common misconceptions and mistakes

• A proxy label is just a rougher label. It is a different quantity, and it can carry its own bias and incentives.
• Accuracy on the proxy means success. A model can be excellent at the proxy and poor at the real goal.
• Proxies are always a mistake. They are often the only practical option, and fine when checked and monitored.

Business use

Require every model brief to state the real target, the label actually used and the known gaps between them.

Sample a small set of direct measurements to test the proxy, even if the sample is costly.

Audit proxy-based models for group differences in how the proxy event is generated.

Track outcomes against the real business goal, so drift between proxy and target is noticed.

People also ask

What is a proxy label?

A measurable stand-in used for training when the true outcome cannot be observed directly.

Why are proxies risky?

A model learns the proxy, including its biases, and the proxy may differ from the goal in ways that harm specific groups.

Can proxies be avoided?

Rarely entirely, but they can be tested against direct measures and replaced when better labels become available.

Check your understanding

What is a proxy label, and what question exposes its risk?

Answer: A proxy label is a measurable stand-in for an outcome that cannot be observed directly, used so a model can be trained at all. The question that exposes its risk is what exactly was the label and how does it differ from the real goal, because a model optimizes the proxy faithfully, and any gap or bias in it becomes the model's behavior.

Related terms

• Label
• Ground truth
• Bias
• Label leakage
• target variable
• fairness

Learn next

AI for Business Owners covers asking vendors what their models were really trained to predict: https://aiforbusinessandlife.com/academy/ai-for-business-owners [C2]

More in AI and machine learning

• Proper generalized decomposition - Proper generalized decomposition, or PGD, is a numerical modeling approach that approximates a solution as a sum of products of simpler functions.
• Prototype methods - Prototype methods are machine-learning approaches that represent categories or patterns through selected or learned representative examples, called prototypes.
• Proximal gradient methods for learning - Proximal gradient methods are optimization algorithms that combine a gradient step for a smooth part of an objective with a proximal step for another part...
• Pure function - A pure function is a function whose output depends only on its inputs and which has no side effects: given the same inputs it always returns the same result...
• PVLV - PVLV stands for primary value and learned value.
• Q-function - A Q-function estimates how good it is to take a particular action in a particular situation, measured as the total reward an agent can expect to collect...

Sources

[S2] Google, Machine Learning Glossary; [S12] Aalto University, Dictionary of Machine Learning. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Proximity weighting Next: Proxy metric