i

Description

Un banc de poissons nageant tous dans la même direction dans une eau claire.

Source

Milada Vigerova · original page

Licence

CC0 Public domain

Original photo, unretouched. The background is a version of it stylised by the site.

Similarity Bias

The candidate with whom we “click” is very often the one who resembles us, and we mistake that ease for proof of competence.

In brief

• What: with equal competence, we favor those who share our background, hobbies, style, and path.
• Why: similarity is reassuring, makes exchanges smoother, and gets confused with “good chemistry” and “cultural fit.”
• Detect: a very homogeneous team, reasons for choices drawn from personality rather than work, “we could've had a beer together.”
• Avoid: criteria set in advance, structured interview, varied evaluators, observable facts before impressions.

On this page

1. Definition
2. The mechanism
3. Where you encounter it
   1. In project management
   2. In management
4. Well-known cases
   1. Rivera: hiring for “cultural fit” in elite firms 2012
   2. Amazon: a resume-screening tool that reproduces past hiring 2014-2018
5. Detection signals
6. Avoidance reflexes
7. With AI: detect, fix, mitigate
   1. Detect
   2. Fix
   3. Mitigate
   4. Prompts to integrate into your practices
   5. How to integrate them
   6. Precautions
8. Limitations
9. Sources and verification
10. See also
    1. Combines with, or reinforces
    2. Related biases
    3. Counterweights and related reflexes

Definition#

Fictitious illustration

Similarity bias (similarity bias, or similar-to-me effect) is the tendency to evaluate people more favorably when they resemble the evaluator: same school, same region, same hobbies, same humor, same way of speaking or reasoning. It requires neither an intention to discriminate nor awareness of the phenomenon: the evaluator sincerely feels they are judging the person.

It extends a classic finding in social psychology, the link between similarity and attraction (Byrne, 1960s-1970s). In organizations, sociologist Rosabeth Moss Kanter described it as early as 1977 under the name homosocial reproduction: leaders tend to promote those who resemble them, which perpetuates the composition at the top.

The mechanism#

• Fluency mistaken for talent. Talking with someone who shares our references feels comfortable. That comfort is interpreted as a sign of the person's quality, even though it describes the relationship.
• Similarity as a guarantee of predictability. A profile close to ours seems easier to manage, less risky: we can anticipate their reactions.
• Vague criteria. The vaguer the criterion (“cultural fit,” “potential,” “leadership”), the more the evaluator fills it with their own references. Similarity then comes in through the door left open.
• Self-esteem. Preferring people like us is, in part, a way of valuing our own life choices.

Where you encounter it#

In project management#

• Team building: the same people are brought back from project to project, those whose habits we know. The team gains fluency and loses diversity of viewpoints.
• Choosing a vendor or expert: the counterpart who speaks our language (same jargon, same technical culture, same alma mater) seems more credible than someone who speaks differently.
• Code and design review: code written “the way we would've written it” is judged clean; correct code in a different style is judged suspicious. Style similarity gets confused with quality.
• Requirements gathering: we listen more readily to users who resemble us, which biases the sample of needs taken into account.

In management#

• Hiring: “cultural fit” becomes the decisive criterion, even though it is the least defined.
• Promotions and high-visibility assignments: we assign visible topics to people we're close to; they progress faster, which seems to confirm the choice.
• Mentoring and management time: we spend more time with team members with whom exchanges are easy, so we give them more feedback and opportunities.
• Evaluation: with equal work, differences in tone from one evaluation to another come from the degree of affinity.

Well-known cases#

Detection signals#

• “He/she will fit in well” is the main reason, with no behavior cited.
• The team is very homogeneous (same school, same age, same background) while the work to be done is varied.
• Reasons for a choice focus on personality, hobbies, shared history, rather than work.
• “We could've had a beer together” appears in an interview report.
• Those who receive visible assignments are those with whom the manager has the most affinity.
• A highly competent but “different” candidate is judged “not a cultural fit.”

Avoidance reflexes#

1. Define “fit” through behaviors. “Shares information,” “accepts being challenged,” “asks for help in time”: criteria that can be observed and rated, instead of a feeling of affinity.
2. Structured interview. Same questions, same order, same rubric for all candidates. Schmidt and Hunter (1998) show that structured interviews predict performance much better than unstructured interviews.
3. Evaluate the work before the person. Technical test, practical exercise, anonymized code review before the interview.
4. Vary evaluators. A panel made up of different profiles makes similarity less decisive, because no one resembles the whole.
5. Rate before discussing. Each evaluator rates independently, then compare. The first person to express their “feeling” steers the group.
6. Ask the counterpart question. “If this person had a different profile (age, school, background) with the same file, would my rating be the same?”
7. Distribute management time. Plan one-on-ones, feedback, and visible assignments explicitly, so affinity does not decide their distribution.

With AI: detect, fix, mitigate#

Prompts to integrate into your practices#

1. Turn “cultural fit” into observable criteria. Before opening applications.

Our team works like this: [description of actual practices: code review, daily standup, support, client relationship].
Propose 4 to 6 observable collaboration criteria that explain success in this context. For each: a one-sentence definition, two behaviors that show it, two that contradict it, and one interview question asked identically to everyone.
Exclude any criterion based on hobbies, age, gender, background, school, or general “energy.”

2. Hunt for affinity in a report. On anonymized text, with no names or identifying details.

Here are my interview notes (anonymized): [notes]. Classify each sentence into one of these categories:
- observable fact related to the role;
- impression of affinity or feeling (“nice,” “we got along well,” “from our world”);
- inference about personality with no fact.
For each affinity or inference sentence, propose the question that would allow the corresponding fact to be found.
Do not judge the person and do not propose any decision.

3. The double-check question. To challenge your own decision before validating it.

I am considering selecting this person for this role: [role criteria, factual elements from the file, no names]. I noted a strong affinity with this person.
Play devil's advocate: list the factual elements that should count, those I may have underestimated in a different profile, and three questions to ask to verify that my choice stands without affinity.

How to integrate them#

• Before opening files: prompt 1, then fixed rubric.
• After each interview: prompt 2 on your own notes before sharing your opinion with the group.
• Before the decision: prompt 3 by the decision-maker, who keeps responsibility for the choice and its written justification.

Precautions#

• Personal data: do not send any names or details that can identify a person; check that the tool is authorized in your organization.
• Legal framework: in Europe, using AI to select or evaluate people is classified as high-risk (Regulation 2024/1689): human oversight, information for individuals. Ask your legal or HR department.
• Do not have AI rank people. It inherits biases from its training data.
• It readily agrees with what you think: explicitly ask it to contradict you.

Limitations#

• Some degree of affinity is legitimate: teamwork requires possible collaboration. Bias starts when affinity replaces competence or remains an undefined criterion.
• Rivera's study focuses on three elite sectors in the United States; it is qualitative. It shows a mechanism but does not measure its magnitude in an IT department or an SMB.
• Bertrand and Mullainathan's study measures discrimination linked to perceived background: it overlaps with resemblance without being limited to it.
• Profile diversity alone does not guarantee better results; it mainly reduces the risk of blind spots. Criteria still need to be defined.

Sources and verification#

• Rivera (2012), “Hiring as Cultural Matching: The Case of Elite Professional Service Firms,” American Sociological Review 77(6) — the study of elite firms.
• Byrne (1971), The Attraction Paradigm, Academic Press — similarity and attraction.
• Kanter (1977), Men and Women of the Corporation, Basic Books — homosocial reproduction.
• Bertrand & Mullainathan (2004), “Are Emily and Greg More Employable than Lakisha and Jamal?,” American Economic Review 94(4).
• Schmidt & Hunter (1998), “The validity and utility of selection methods in personnel psychology,” Psychological Bulletin — the advantage of structured interviews.
• Dastin (2018), “Amazon scraps secret AI recruiting tool that showed bias against women,” Reuters, October 10, 2018.
• Regulation (EU) 2024/1689 on artificial intelligence — Annex III: employment and worker management.

See also#