i

Description

Un halo lumineux en anneau autour du soleil dans un ciel bleu.

Source

Joshua Tree National Park (NPS / Emily Hassell) · original page

Licence

PD Public domain

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

The halo effect

A favorable impression on one point rubs off on all the others, and vice versa.

In brief

• What : a salient trait (ease, degree, reputation) influences judgment on unrelated traits.
• Why : the brain prefers a coherent image of a person; it fills in the blanks.
• Detect : almost identical ratings across all criteria, vague justifications (“good feeling”).
• Avoid : criteria defined beforehand, rated separately, independent evaluators, dated facts.

On this page

1. Definition
2. The mechanism
3. Where you encounter it
   1. In management
   2. In project management
4. Famous cases
   1. Theranos: prestige as a substitute for proof 2003-2022
   2. Stavisky and the Bayonne bonds: borrowed respectability (France) 1931-1934
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. Limits
9. Sources and verification
10. See also
    1. Combines with, or reinforces
    2. Related biases
    3. Counterweights and related reflexes

Definition#

Illustration inspired by the person

The halo effect is the tendency to let a general impression, or a salient quality, influence judgment about distinct characteristics. A candidate who is comfortable speaking is also judged to be more technically competent; an engineer who is brilliant on one topic is assumed to be good at everything.

The phenomenon was described by psychologist Edward Thorndike in 1920: officers tasked with rating their subordinates gave ratings that were far too correlated across qualities (physique, intelligence, leadership, character) that should be independent. The reverse phenomenon, one flaw that darkens everything, is sometimes called the horn effect.

The mechanism#

• Coherence before accuracy. A person who is “good at one thing and bad at another” is uncomfortable to think about. The mind smooths things out: it makes the whole picture coherent.
• What you see is all there is. Kahneman calls this WYSIATI (what you see is all there is): we judge based on available information as if it were complete, and we extrapolate.
• Unawareness of the effect. Nisbett and Wilson (1977) had people evaluate the same instructor, presented as either warm or cold: participants also judged his appearance and accent differently depending on the version, and denied that his warmth had influenced their ratings.

Where you encounter it#

In management#

• Recruiting : interview ease, the prestige of a school or a previous employer rubs off on the assessment of technical competence.
• Annual review : a recent, visible success pushes all ratings up; an incident pulls them down.
• Promotions : the person you know well, or who speaks well in committee, is judged “ready” for responsibilities different from those they currently handle.

In project management#

• Vendor and tool selection : the publisher's reputation, the quality of the sales presentation, or the marketing site leads people to assume product or support quality that nothing has demonstrated.
• Code and design review : a proposal is read differently depending on whether it comes from someone reputed to be excellent or from a beginner.

Famous cases#

Detection signals#

• A person's ratings across different criteria are almost identical.
• “This is an excellent person, they'll definitely be able to do it.”
• Justifications are vague: “good feeling,” “good profile,” “you can tell he's solid.”
• A negative judgment on one point, drawn from a single observation, extends to everything else.
• Everyone's opinion on the same candidate changes depending on who spoke first.

Avoidance reflexes#

1. Define the criteria first. Write down, before meeting anyone, what matters for the role or decision, and how to recognize it.
2. Rate one criterion at a time. Evaluate each criterion separately, for all candidates, before moving on to the next one. Kahneman applied this method to reform Israeli army interviews, with a clear gain in validity.
3. Evaluate independently before discussing. Each evaluator writes their ratings alone; then they are compared. The first person to speak has less influence on the group.
4. Rely on dated facts. “He delivered this module on time in March” is better than “he is reliable.” Record examples throughout the year, not the day before the interview.
5. Mask what can be masked. Technical test before the interview, code review without the author's name, comparison of responses to an RFP without logos.
6. Look for the reverse halo. Ask yourself: “What am I accepting because I like this person? What am I rejecting because I dislike them?”
7. Calibrate among evaluators. Compare ratings on shared cases to identify those who systematically smooth everything out.

With AI: detect, fix, mitigate#

Prompts to integrate into your practices#

1. Criteria grid before seeing candidates.

Here is the job description: [text]. Propose an evaluation grid with 5 to 7 criteria independent from one another.
For each criterion: a one-sentence definition, three observable behaviors that show mastery, three that show lack of it, and one interview question asked identically to all candidates.
Do not include any criterion based on appearance, general social ease, age, gender, or origin.

2. Separate facts from impressions. On anonymized notes.

Here are my notes on an evaluated person (name removed): [notes]. Classify each sentence into one of these categories:
- observable, dated fact;
- impression or judgment without supporting fact;
- inference (deduction from another trait).
List adjectives or judgments not based on any fact and propose, for each one, the question I should ask myself to find the corresponding fact.
Do not make any judgment about the person.

3. Consistency check of ratings. On a table without names.

Here is an anonymized ratings table (one row per person, one column per criterion): [table].
Use code to calculate, for each evaluator: the standard deviation of ratings for the same person and the correlation between criteria.
Flag profiles whose ratings are almost identical across all criteria (possible halo) and evaluators whose ratings are systematically homogeneous.
Specify that these are signals to verify, not conclusions.

4. Fact-based evaluation feedback.

Based on the following dated facts: [facts, criterion by criterion], write structured evaluation feedback by criterion.
One sentence per fact, with no personality adjectives (no “brilliant,” “reliable,” “motivated” without an associated fact).
End each criterion with an improvement point phrased as a behavior to change, not a trait.

How to integrate them#

• Recruitment: prompt 1 before screening applications; the grid is locked before opening files.
• Evaluation: prompt 2 while preparing annual reviews; prompt 3 by HR, on anonymized data, before calibration.
• Keep the human decision and make it traceable: AI prepares, the evaluator decides and justifies.

Precautions#

• Personal data: do not send names or any identifying information to an AI tool; always anonymize, and verify that the tool is approved by your organization.
• Legal framework: in Europe, using AI to select candidates or evaluate employees falls under high-risk systems (AI regulation); these uses require human oversight and informing the people concerned. Check with your legal or HR department before automating anything.
• AI-specific bias: do not have it rank or rate people. Use it for method (criteria, consistency, wording).
• It readily agrees with what you already think (sycophancy): explicitly ask it to challenge your ratings.

Limits#

• Some qualities are genuinely correlated: an apparent halo may sometimes reflect a real correlation. The bias is assuming it without checking.
• Overly formal criteria can create false objectivity: they reduce halo, not other biases (similarity, authority, anchoring).

Sources and verification#

• Thorndike (1920), “A constant error in psychological ratings,” Journal of Applied Psychology — the original description.
• Nisbett & Wilson (1977), “The halo effect: Evidence for unconscious alteration of judgments,” Journal of Personality and Social Psychology.
• Schmidt & Hunter (1998), “The validity and utility of selection methods in personnel psychology,” Psychological Bulletin — the advantage of structured interviews.
• Kahneman (2011), Thinking, Fast and Slow, chapters on consistency and hiring interviews.
• Wilson & Caliskan (2024), “Gender, Race, and Intersectional Bias in Resume Screening via Language Model Retrieval” — AAAI/ACM Conference on AI, Ethics, and Society.
• Regulation (EU) 2024/1689 on artificial intelligence — Annex III: high-risk systems (employment, worker management).
• Sharma et al. (2023), “Towards Understanding Sycophancy in Language Models” — language model sycophancy.
• Axios (2018), “SEC charges Theranos CEO Elizabeth Holmes with "massive fraud"”
• CNBC (2015), Wall Street Journal investigation into test reliability
• Wikipedia, “Elizabeth Holmes” — trial, conviction (January 2022), and sentence (November 2022).
• Carreyrou (2018), Bad Blood: Secrets and Lies in a Silicon Valley Startup — the Wall Street Journal reporter's investigation.
• H-France Review, review by Paul F. Jankowski, Stavisky: A Confidence Man in the Republic of Virtue (Cornell University Press, 2002) — describes Stavisky's manufactured respectability, trial postponements, and political interventions.
• Herodote.net, “January 9, 1934: death of Stavisky” — timeline of Stavisky's death, the government's fall, and the February 6, 1934 riot.

See also#