Is linking bias the same as population bias?

No. Population bias concerns who is represented; linking bias concerns the relationships represented among participants, although both can coexist.

Does it require a machine-learning model?

No. A simple network analysis can make the same interpretive mistake.

Can a graph still be useful?

Yes. It can answer a well-defined question when its link meanings and limitations are explicit.

AI Glossary / Governance, ethics and safety / Intermediate

Linking Bias

Definition

Linking bias describes systematic differences in the networks inferred from people's recorded connections, interactions or activity. A network built from one kind of online link may not faithfully represent the relationships a researcher or AI system intends to study.

What it means

A social network diagram contains nodes and links, but the meaning of a link is a design decision. Following an account, exchanging messages and knowing someone personally are different relationships. Treating them as interchangeable can produce misleading conclusions about influence, trust or community structure.

The source taxonomy places linking bias within behavioral biases in social data. It does not mean that adding a web hyperlink is inherently biased, or that every observed network is wrong. A network can be suitable for one question and unsuitable for another.

The key is to state what the graph represents. A map of public interactions should not silently become a map of all social relationships, especially when private or offline interactions are absent.

How it works

Researchers choose participants, define links and decide how to aggregate activity. A follow might create a directed link; a mutual exchange might create an undirected one; the number of messages might become a weight. Each choice changes the resulting graph.

The platform also influences which links are possible and visible. People may follow public figures without personal acquaintance, interact differently in private channels or maintain important relationships outside the platform. The observed network therefore reflects both social behavior and the recording mechanism.

Compare alternative link definitions where appropriate and examine what the available data omits. A centrality score is a property of the constructed graph under its assumptions. It is not automatically a direct measurement of real-world authority or influence.

Three practical examples

A company avoids treating a public reply network as a complete map of collaboration when much work happens in private channels.
A recommender team distinguishes following an account from having a reciprocal relationship with its owner.
A community study tests whether apparent clusters depend on using message counts, unique contacts or mutual exchanges as links.

Think of it this way

Think of mapping friendships from seating positions at one event. The seating chart is real, but it may reflect ticket allocation, arrival time or available chairs rather than enduring relationships.

The analogy shows why a recorded connection needs interpretation. Online links can carry meaningful social information; the point is to establish which relationship they support rather than dismiss them altogether.

Worked example

Two graphs from the same four people

In an invented group, A publicly follows B, C and D. There are no other public follow links. A therefore has three outgoing links in that graph, while the others have none.

Private work records show that B exchanges messages with C and D, while A has no work exchanges with anyone in the group. In a graph of those exchanges, B is connected to two colleagues and A is isolated.

Neither graph must be fabricated. They represent different relationships. Calling A the central workplace collaborator because of the public-follow graph would substitute one meaning of connection for another. The lesson is to define links in relation to the question before interpreting a network score.

Illustrative case study

Fictional company Willow Teams uses an AI tool to identify knowledge-sharing hubs from public discussion channels. Employees who answer public questions receive high network scores, while colleagues who support newcomers through private mentoring appear less connected.

The company reviews what the graph actually captures and changes the description to public discussion activity. It explores additional, appropriately collected evidence before making broader claims about knowledge sharing. It also consults staff about work that the platform does not record.

The team keeps the graph for its narrower purpose rather than presenting a single visible network as a complete account of everyone's contribution.

Common misconceptions and mistakes

A link has the same meaning on every platform. Connection mechanisms and social conventions differ.
A high centrality score proves real-world influence. The score depends on the chosen network and metric.
More recorded activity always means a stronger relationship. Activity may reflect conflict, administrative duties or platform mechanics.
Using the same participants removes linking bias. The relationships recorded among them can still differ systematically.

Business use

Linking bias matters when businesses analyze collaboration, recommendations, information flow or customer communities. Decisions based on network metrics can affect visibility and opportunity, so the interpretation deserves scrutiny.

Ask what each edge means, which interactions are missing and whether the score is being used beyond that meaning. Compare results under reasonable alternative definitions and investigate surprising classifications with contextual evidence. Avoid turning absence from a particular network into a judgment about a person's willingness to cooperate or value to an organization.

Check your understanding

An employee has few public-message links but handles extensive private mentoring. Can the public graph establish that they rarely collaborate?

Answer: No. It measures the selected public interactions. A broader collaboration claim requires evidence that covers the relevant forms of work.

Related terms

Learn next

Read graph and construct validity. Write a plain-language definition of one edge in a network your team uses. Then list relationships that people might incorrectly assume the edge represents.

Sources

Olteanu, Castillo, Diaz and Kiciman, Social Data: Biases, Methodological Pitfalls, and Ethical Boundaries, section 3.5: https://www.frontiersin.org/journals/big-data/articles/10.3389/fdata.2019.00013/full. This is the source taxonomy for linking biases; the workplace examples and four-person graphs are original teaching illustrations.

Previous: Linked data Next: Lipschitz continuity