i

Description

Des rangées de sièges de bois vides dans une salle d'attente de gare.

Source

須田拓樹 · original page

Licence

CC0 Public domain

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

The contrast effect

After a brilliant candidate, the next one seems mediocre; after a weak candidate, the same one seems excellent.

In brief

• What: the perceived value of something depends on what it is compared to right before; the same thing seems better or worse depending on what sits next to it.
• Why: the mind judges by differences, not absolute values; the last example becomes the reference point.
• Detect: ratings that vary with order, a “not as good as the previous one” in comments.
• Avoid: a rubric set in advance, score each candidate or deliverable on its own, mix the order, space out evaluations.

On this page

1. Definition
2. How it works
3. Where you encounter it
   1. In project management
   2. In management
4. Famous cases
   1. Ten years of admissions interviews: order of appearance 2013
   2. Parole decisions: nuance required 2011-2016
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#

Illustration inspired by the person

Fictitious illustration

The contrast effect is the amplification of perceived differences between two items examined one after the other (or side by side): after a high-level stimulus, an average stimulus is judged lower; after a weak stimulus, it is judged higher. Perception is relative: we do not evaluate a thing, we evaluate it in relation to the previous one.

The phenomenon is old and general: a bucket of lukewarm water feels warm after ice and cold after scalding water, and psychophysics experiments have shown this with lifted weights (Sherif et al., 1958). In social psychology, Douglas Kenrick and Sara Gutierres (1980) demonstrated it in attractiveness judgments: students who had just watched a TV show featuring very attractive women judged a photo of an average-looking woman as less attractive, compared with students who had watched another program.

It is close to Anchoring, where the reference is a number, and to The framing effect, where it is wording. Here, it is the previous item that serves as the reference.

How it works#

• A relative judgment. The brain has no absolute scale for most qualities (competence, charm, price, difficulty): it compares with what is available in immediate memory.
• The latest item is the reference. The last example seen occupies working memory; it becomes the implicit yardstick for the next item.
• A shifting scale. After several weak items, the “normal” shifts downward; the same submission then looks good, and a decent submission looks excellent. The opposite happens after several excellent items.
• A need to spread out. When rating several items in sequence, we tend to spread ratings across the whole scale, even if items are close: we exaggerate gaps.

Where you encounter it#

In project management#

• Offer comparison: three vendor offers; one is clearly too expensive, making the other two look “reasonable” (see also Anchoring). The order in which proposal files are read changes ratings.
• Code review: after a catastrophic merge request, an “average” request looks clean; after exemplary work, the same one looks sloppy.
• Estimating: a 5-day task looks small if it follows a 40-day task; it looks large after a 0.5-day task. Planning poker without a stable reference suffers from this.
• Defects: a medium-severity defect is classified as “minor” after a major incident in the same review session, “major” after a string of minor details.
• Demos and acceptance testing: a successful demo right after a failed demo is judged better than it is.

In management#

• Back-to-back hiring interviews: the candidate after a very weak candidate is judged brilliant, and a good candidate following a star is judged average.
• Application screening: after twenty weak resumes, a decent resume looks remarkable; the reverse after a stack of exceptional profiles.
• Evaluation calibration: a solid employee, evaluated right after the best member of the team, gets a harsher rating. A manager with a strong team rates more harshly than one with a weaker team.
• Promotion or internal hiring: the external candidate, compared with the last employee who left, seems better or worse depending on how that person is remembered.
• Raise announcements: a 3% raise seems generous when announced after a frozen year, and stingy after a year of bonuses.

Famous cases#

Detection signals#

• “He is much worse than the previous one”: an evaluation phrased relative to someone else.
• Ratings in the same session spread out across the full scale, even though the work is similar.
• A candidate's rating changes depending on the order in which files are reread.
• After an exceptional case, all following ones seem “okay” or “disappointing” without the rubric being reviewed.
• An estimate or quote seems “reasonable” only because another, more expensive one, was seen right before.
• A strong team judges newcomers more harshly than an average team would.

Avoidance reflexes#

1. Set the rubric before seeing anything. Written criteria, levels, and examples: each item is judged against the rubric.
2. Rate each item on its own, immediately. Write down a candidate's rating and facts before moving to the next one, without revisiting previous ratings.
3. Change the order. When possible, vary the order in which files are reviewed or the order people go through; in a committee, rotate evaluators.
4. Space things out. Avoid a burst of interviews or reviews without breaks; spacing makes it easier to return to the rubric.
5. Use absolute references. A scale anchored in behaviors (“delivered a full work package end-to-end independently”), level examples, rating history.
6. Evaluate first, compare afterward. Comparing candidates has its place, but after individual ratings; it decides, it does not rate.
7. Be wary of the “best in the batch.” An item that dominates a weak batch is not necessarily good; check it against the rubric.
8. Use blind rating. Have multiple evaluators review separately, in different orders, then compare ratings.

With AI: detect, fix, mitigate#

Prompts to integrate into your practices#

1. Isolated evaluation against a rubric. A fresh session per candidate or per deliverable; the rubric is provided before the item.

Here is an evaluation rubric: [criteria, levels, and behavior examples].
Evaluate the item below <strong>only</strong> against this rubric. You must not compare this item to any other item, or to any assumed average.
For each criterion: the rating, the specific facts from the document that justify it, and what is missing to assess it. If information is missing, do not assume it.

Item: [anonymized resume / review / deliverable]

2. Detect an order effect. After a series of interviews or reviews, using anonymized data.

Here are ratings in order of appearance, with criteria and associated comments: [table].
1. Calculate, for each item, the previous rating and the gap.
2. Indicate whether ratings seem more tied to the previous item's rating than to the facts cited in the comments.
3. Quote comments that judge by comparison (“better than,” “worse than”).
4. State what you can assert from this data and what requires more evidence (sample size, non-random order).

How to integrate them#

• Before the session: set the rubric, optionally with AI help to anchor it with examples.
• During: each evaluator rates alone, before seeing others' ratings; AI provides an independent rating via prompt 1.
• After: prompt 2 is used for process review, not to judge evaluators.

Precautions#

• Resumes and evaluations are personal data: anonymize and use only an approved tool; automated candidate evaluation is regulated by law.
• Never evaluate two candidates back-to-back in the same conversation.
• AI reproduces biases from its data (name, gender, school): it does not replace a reasoned human decision.
• It may cite facts not present in the document: verify every justification.

Limitations#

• Contrast is not always unfavorable: comparison is sometimes useful (choosing among three offers means comparing them). Bias starts when comparison replaces the standard.
• The opposite effect exists: assimilation. When items are perceived as belonging to the same category, judgment moves closer to the reference instead of farther away from it (Mussweiler, 2003).
• The order effect in successive decisions, widely discussed, is debated (see the parole case); the MBA interview study, meanwhile, is based on observed data from a single program.
• In some contexts (competitive exams, panels), contrast is amplified by the requirement to rank, which pushes ratings apart.

Sources and verification#

• Sherif, Taub & Hovland (1958), “Assimilation and contrast effects of anchoring stimuli on judgments,” Journal of Experimental Psychology, 55 — lifted weights.
• Kenrick & Gutierres (1980), “Contrast effects and judgments of physical attractiveness: When beauty becomes a social problem,” Journal of Personality and Social Psychology — the TV series experiment.
• Simonsohn & Gino (2013), “Daily horizons: Evidence of narrow bracketing in judgment from 10 years of MBA admissions interviews,” Psychological Science — order of appearance in interviews.
• Danziger, Levav & Avnaim-Pesso (2011), “Extraneous factors in judicial decisions,” PNAS — controversial study.
• Glöckner (2016), “The irrational hungry judge effect revisited: Simulations reveal that the magnitude of the effect is overestimated,” Judgment and Decision Making, 11(6) — simulation.
• Weinshall-Margel & Shapard (2011), “Overlooked factors in the analysis of parole decisions,” PNAS — methodological critique.
• Mussweiler (2003), “Comparison processes in social judgment: Mechanisms and consequences,” Psychological Review — contrast and assimilation.
• Echterhoff et al. (2024), “Cognitive Bias in Decision-Making with LLMs” — cognitive biases in language models.

See also#