Is social sorting the same as profiling?

Profiling constructs a representation or inference; social sorting emphasizes the categories and treatment that organize people's opportunities or experiences.

Can sorting help people?

Yes. Categories may direct support to those who need it, provided their design and consequences are assessed.

Does a difference between groups prove discrimination?

A difference is evidence to investigate. Understanding its cause, justification and any legal significance requires further context and the applicable rules.

AI Glossary / Data science and analytics / Intermediate

Social sorting

Definition

Social sorting is the classification of people or populations into categories that influence how they are treated, what opportunities they receive or what scrutiny they face. In surveillance studies, notably David Lyon's work, it directs attention to the social consequences of data-based classification rather than only to information collection.

What it means

A system may place customers into service tiers, route people through different checks or decide which opportunities to display. Social sorting asks who defines those categories, which evidence is used and what follows from belonging to one category rather than another.

The term does not mean that every classification is harmful. Categories can help allocate assistance or organize services. The important question is whether their purpose, construction and consequences are justified in the setting where they operate.

AI can automate and scale sorting, but the practice predates machine learning. A manually maintained list or a simple rule can also affect people's treatment. A highly accurate model can still support an inappropriate policy if it predicts a target that should not determine the decision.

How it works

The process connects data, classification and action. Records about people are collected or combined; a rule or model produces a category or score; an organization links the output to different treatment.

That action link is crucial. A descriptive grouping used to understand service demand differs from the same grouping used to deny access. Thresholds, exceptions and human procedures can matter as much as the model.

Examine how categories are measured and updated. A geographic indicator may stand in for assumptions about individuals, while historical records may reflect earlier unequal treatment. Missing data can become a category of its own if “unknown” automatically triggers an adverse response.

Review actual outcomes across relevant groups and individual cases. Provide a meaningful correction process and inspect feedback loops: reduced access can change later observations in ways that seem to confirm the original classification.

Three practical examples

A subscription service places customers in support tiers, prompting a review of whether the allocation rule produces acceptable differences in response time.
A training provider uses an eligibility category to allocate extra learning support, checking that missing records do not exclude people who need assistance.
An online platform ranks opportunities for members, requiring examination of who repeatedly receives low visibility and whether the ranking creates self-reinforcing disadvantage.

Think of it this way

Think of a building with several queues whose entry rules are written on an unseen desk. Sorting is not just the label on a visitor's ticket; it is also which door opens, how long the visitor waits and whether mistakes can be corrected.

The analogy is limited because digital categories can travel between organizations, operate invisibly and change over time. People may never see the ticket or realize that a previous classification affects a later service.

Worked example

Tracing a rule to its consequence

Consider a fictional learning service with 200 equally eligible applicants from each of two locations. It gives early access to anyone whose record contains at least three previous online course completions.

Suppose 160 applicants from location A and 80 from location B meet the recorded threshold. The resulting early-access rates are 80% and 40%. The numerical difference is 40 percentage points.

Those figures identify a disparity; by themselves they do not establish its cause or a legal conclusion. The organization must investigate the relevance of prior completions, unequal access to previous courses, missing records and the purpose of early access.

If early access improves the chance of later completion, the rule can also influence future evidence used to allocate access. Evaluating only whether the computer counted previous completions correctly would miss the policy's wider consequences.

Illustrative case study

Fictional case: Cedar Services uses predicted future spending to assign support priority. Customers receiving slower help become less likely to renew, and the next model interprets their lower renewal rates as further evidence that they deserve low priority.

The review team maps the entire process, including queue rules and customer outcomes. It tests whether the support policy contributes to the pattern, establishes a service floor and creates a correction route for inaccurate records.

The team does not assume that removing one demographic field will resolve the issue. Other variables and the policy itself may still produce unequal treatment. This invented case shows why social sorting is examined as an organisational process, not simply as a classification accuracy problem.

Common misconceptions and mistakes

Social sorting requires advanced AI. Rules, databases and manual categories can perform the same social function.
Every grouping is necessarily unjust. The purpose, evidence and consequences require contextual assessment.
Removing sensitive fields guarantees fair treatment. Other variables can act as proxies, and organisational rules can still create disadvantage.
Correct predictions automatically justify a decision. Predictive validity and the legitimacy of the action are separate questions.

Business use

The concept helps businesses inspect customer segmentation, prioritization, eligibility and automated routing. Ask what changes for a person when a label changes, not merely how the label is calculated.

Document the decision owner, intended benefit, affected groups and ways to contest incorrect records or classifications. Review the combined system of model, threshold and workflow. Privacy protections remain important, but limiting access to a database does not by itself address unfair treatment produced through that database's categories.

Check your understanding

A model predicts spending accurately, and its output determines who receives timely support. Has accuracy alone established that the system treats customers appropriately?

Answer: No. Review whether spending should determine support, the minimum service owed, errors and group outcomes, and whether the resulting treatment changes future spending. The action policy matters alongside prediction quality.

Related terms

Social profiling
Classification
Feedback loop
Accountability
algorithmic fairness
proxy variable
eligibility
surveillance

Learn next

Read proxy variable and algorithmic fairness. Draw a simple chain from collected data to category to treatment. Identify who benefits, who may lose access and how an incorrect category can be challenged or revised.

Sources

David Lyon, The End of Privacy?, social-sorting discussion: https://webarchive.ars.electronica.art/en/archiv_files/20071/010_FE_2007_David_Lyon_en.pdf. Lyon, Surveillance, Security and Social Sorting, abstract: https://journals.sagepub.com/doi/10.1177/1057567707306643. NIST, AI Risks and Trustworthiness, harmful bias: https://airc.nist.gov/airmf-resources/airmf/3-sec-characteristics/. NIST, measurement guidance: https://airc.nist.gov/airmf-resources/playbook/measure/. All worked figures and business cases here are fictional illustrations.

Previous: Social profiling Next: Societal Bias