Is unawareness the same as anonymization?

No. Excluding a sensitive field does not establish that individuals cannot be identified.

Does a proxy have to perfectly predict the omitted attribute?

No. Partial associations can still affect outcomes.

Does unawareness guarantee demographic parity?

No. Equal selection rates do not follow from omitting an input.

AI Glossary / Governance, ethics and safety / Advanced

Unawareness (to a sensitive attribute)

Definition

Unawareness to a sensitive attribute means omitting that attribute from the data used to train a model, even though the characteristic exists in the population. It prevents direct use of that particular field during training. It does not guarantee that predictions are independent of the characteristic or that the resulting system is fair.

What it means

A sensitive attribute is a characteristic given particular consideration for ethical, social, personal, or legal reasons. Which attributes matter depends on context. In this glossary sense, unawareness describes an information restriction, not a model's consciousness or a person's knowledge.
Suppose a training table excludes a group identifier. Other columns may still carry information associated with that group. A model can exploit those relationships without receiving the identifier itself.
The phrase “fairness through unawareness” sometimes describes the idea that exclusion alone will deliver fairness. That conclusion is not justified. A feature list says what inputs are available; it does not establish how benefits, errors, or burdens are distributed.

How it works

The developer removes or withholds the sensitive field from the training inputs. The remaining features are used to learn the prediction rule. If those features correlate with the omitted characteristic, predictions can still vary systematically across groups.
A proxy need not be a perfect substitute. Several weakly informative features may jointly convey information that no single column reveals clearly. Historical labels and unequal representation can also affect outcomes independently of whether the sensitive field is present.
Assessment therefore needs to examine the context and relevant outcomes. Where appropriately available, sensitive-attribute information can be kept separate for evaluation rather than supplied to the predictor. Its collection and use require appropriate governance; unawareness is not a blanket instruction to collect or delete such information everywhere.

Three practical examples

A course-recommendation model omits a demographic field but uses previous attendance. If access to earlier courses differed across groups, the model can reproduce an access pattern while remaining unaware of the demographic field.
A service-routing model excludes a group label but uses the channel through which a request arrives. If groups use channels in different proportions, routing outcomes can differ even though the same rule is applied to every request.
An analyst receives only predictions and an overall accuracy score. With no suitable subgroup information, the analyst cannot infer equal performance across groups. Missing evidence does not establish either fairness or unfairness.

Think of it this way

Think of hiding the category column in a spreadsheet while leaving several related columns visible. Someone may still recognize patterns linked to the hidden category. Removing a label changes the available information, but it does not necessarily remove the relationships embedded in the remaining data.

Worked example

Consider an invented system that prioritizes requests submitted through an online channel. It receives no sensitive group identifier and applies the same rule to everyone: online requests receive priority; other requests do not.
There are 100 requests from Group A and 100 from Group B. Eighty A requests arrive online, compared with 20 B requests. The priority rates are therefore 80/100 = 80% for A and 20/100 = 20% for B.
The difference is 60 percentage points. The model never reads the group identifier, yet its outputs are associated with that identifier because channel use differs between groups.
Now suppose an analyst reports only the combined result: 100 of 200 requests receive priority, or 50%. That aggregate does not reveal the group difference.
The example demonstrates that unawareness does not imply equal selection rates. It does not, by itself, determine which fairness criterion is appropriate or prove why channel use differs. Deciding whether the rule is justified requires the service's purpose, consequences, and relevant context. No legal conclusion follows from this arithmetic alone.

Illustrative case study

Fictional company Willow Learning develops a tool to suggest employees for optional workshops. The team omits a sensitive demographic field and initially describes the system as fair because it cannot access that field.
A review finds that the model relies heavily on attendance at previous workshops. Earlier sessions were offered at times that some employees could rarely attend. The team investigates whether the attendance feature measures interest, opportunity, or a mixture of both.
It reviews the workshop-selection purpose, examines relevant outcome differences, and considers changing both the recommendation approach and the scheduling process. It also separates information used for evaluation from the tool's prediction inputs where appropriate.
This fictional case illustrates why removing one field does not settle the broader assessment. It does not claim that every attendance feature is inappropriate or that adding the omitted field automatically fixes the problem.

Common misconceptions and mistakes

“The model cannot use information about a field it does not receive.” Related inputs may carry some of that information.
“The same rule guarantees fair outcomes.” Groups can encounter the rule under different conditions.
“A disparity proves intentional discrimination.” The observation alone does not establish intent or a complete causal explanation.
“Sensitive attributes must always be predictor inputs.” The appropriate approach depends on the task and applicable constraints.
“Removing evaluation data improves fairness.” It may instead make relevant differences harder to investigate.

Business use

When a supplier claims a model is fair because it excludes sensitive fields, ask what outcome assessment supports that claim. Examine the remaining inputs, the target being predicted, and how the model enters the actual workflow.
Distinguish three questions: what the model receives, what evaluators can examine, and what outcomes people experience. Keep documentation clear about those differences. Investigate potential proxies and historical patterns without assuming that correlation alone identifies the correct remedy. Evaluate proposed changes against the relevant objectives and harms rather than treating feature removal as a complete solution.

Check your understanding

Why do the worked-example priority rates differ?

Answer: Channel use differs across groups, and priority depends on channel.

What does the combined 50% rate conceal?

Answer: The 80% and 20% subgroup rates.

Why separate prediction inputs from evaluation information?

Answer: A field may be excluded from predictions while still being relevant to assessing outcomes under appropriate governance.

Related terms

Learn next

Choose a hypothetical prediction task. Identify an omitted sensitive attribute, a possible related input, and an outcome measure worth comparing across groups. Explain what the comparison would reveal and what further context it would still require.

Sensitive attribute

Sources

Google, Machine Learning Glossary, unawareness to a sensitive attribute and sensitive-attribute proxies: https://developers.google.com/machine-learning/glossary/responsible-ai Fairlearn, Fairness in Machine Learning, context, harms, and assessment: https://fairlearn.org/main/user_guide/fairness_in_machine_learning.html

Previous: Unavoidable Redundancy Next: Unbiased estimator