---
name: Negativity bias
slug: negativity-bias
category: Judgment and bias
categories: [Onboarding, Retention]
evidence: moderate
page: /principles/negativity-bias/
updated: 2026-10-30
---

# Negativity bias

> A bad moment in your product counts for more than a good one, so fixing what hurts moves more than adding delight.

## What it is

Negative information draws more attention, more processing and more weight in a judgement than positive information of the same size. The asymmetry is an average across people, not a property of every user.

## Evidence

The claim comes from Baumeister, Bratslavsky, Finkenauer and Vohs (2001), who swept everyday events, impressions, learning, close relationships and health for cases where good outweighed bad and “were unable to locate any significant spheres in which good was consistently stronger than bad”. It is a narrative review, not a meta-analysis, and the authors are candid that the evidence “varied widely from one topic to another”. The five-to-one ratio product writing likes to quote comes from the same paper, where it is attributed to Gottman's marital coding work; Heyman and Slep (2001) showed that divorce-prediction equations of that family lose accuracy “precipitously” once they are cross-validated. Treat the direction as well supported and the multiplier as a slogan.

The product-shaped version has real numbers behind it. Mittal, Ross and Baldasare (1998) modelled attribute performance against overall satisfaction and repurchase intention across three large consumer samples in health care and car ownership, and found that negative performance on an attribute moves both outcomes more than positive performance on that same attribute. That is the actual warrant for fixing pain points before adding delight: per attribute, doing it badly costs more than doing it well gains. Keaveney's (1995) critical-incident study points the same way from the churn end — service customers described more than 800 behaviours that made them switch, and they sorted into categories of things that went wrong rather than things that were missing.

In interfaces, the best-measured case is ratings. Chevalier and Mayzlin compared the same books' reviews and sales ranks on Amazon and BarnesandNoble.com and found that “the impact of 1-star reviews is greater than the impact of 5-star reviews”. Their illustration: turn one of a book's four 5-star Amazon reviews into a 1-star review and its rank is expected to fall from 500 to 603, about 20 books a week. Their own reading is diagnosticity rather than a weighting bias — an author can post cheap 5-star reviews but cannot stop a stranger posting a 1-star one — so what a product owes a buyer is the distribution, not the average alone.

The attention half of the effect has been shown at scale and is small per unit. Robertson, Pröllochs, Schwarzenegger, Pärnamets, Van Bavel and Feuerriegel (2023) analysed 22,743 randomised Upworthy headline tests across more than 370 million impressions. For a headline of average length, each additional negative word raised click-through by 2.3% and each positive word lowered it by 1.0%, against a mean click-through of 1.39%. The emotion breakdown cuts against the obvious application: sadness words raised clicks, joy and fear words both lowered them, and anger was null. Negative is not the same as frightening, and in the one large field test the frightening version lost.

It is also not universal. Soroka, Fournier and Nir (2019) recorded skin conductance and heart-rate variability from 1,156 respondents in 17 countries watching the same BBC World News stories. Arousal was reliably higher for negative coverage on average, but only nine of 17 countries reached significance on skin conductance and five on heart-rate variability, and the New Zealand and Sweden samples trended the other way. Country explained under 3% of the variance in individual responses, and the authors conclude that a sizable share of individuals may prefer less negative content — which is why a negative-framed message can win a test overall and still cost you a segment.

The standard UX advice outruns the evidence in two places. First, no controlled test of error-message wording turned up in this research: “write gentle error messages” is a reasonable inference from the asymmetry, not a measured effect. What interfaces do have is field evidence that negative messages get dismissed rather than overweighted — Akhawe and Felt (2013) observed 25.4 million browser warning impressions and found 70.2% click-through on Chrome's SSL warning — and evidence that the fix is layout rather than language: Felt and colleagues (2015) raised field adherence on that warning from 37% to 62% with opinionated design, visual cues that make the safe path obvious, while failing at their primary goal of making the warning understood. Second, on the recovery side Gelbrich and Roschk's (2011) meta-analysis found compensation the salient driver of transaction-specific satisfaction, while cumulative satisfaction — the primary antecedent of loyalty — depends slightly more on interactional justice, how the complaint was handled, than on what was paid out.

Two corrections to the folk version. Fear-based messaging is not ineffective: Tannenbaum and colleagues (2015) pooled 248 independent samples (N = 27,372) and found fear appeals positively affected attitudes, intentions and behaviour at a random-effects d = 0.29, with “very few circumstances under which they are not effective” and no identified circumstances under which they backfire. The reason not to build with fear is ethical and legal, not empirical. And the familiar “losses hurt about twice as much” figure belongs to a different principle: Brown, Imai, Vieider and Camerer's (2024) meta-analysis puts the loss aversion coefficient between 1.8 and 2.1, measured on monetary gambles, and it does not transfer to how a failed upload weighs against a smooth one.

## When it holds

- The asymmetry is per attribute, not global: negative performance on an attribute moved satisfaction and repurchase intention more than positive performance on the same attribute (Mittal, Ross & Baldasare 1998). The comparison that holds is “this thing done badly against this thing done well”, not “a bug against a new feature”.
- Negative information wins attention mainly when it is the credible signal. Chevalier and Mayzlin attribute the greater weight of 1-star reviews to positive reviews being cheap to manufacture and negative ones not; where your negative signal is obviously self-serving, that argument does not apply.
- Negative is not interchangeable with frightening: across 22,743 headline tests, sadness words raised click-through, joy and fear words lowered it, and anger was null. The per-word effect is small — +2.3% click-through per negative word against a 1.39% mean — and it declined over the 2013–2015 data window (Robertson et al. 2023).
- Individual variation dwarfs group variation. Across 17 countries, country explained 1.5% of the variance in heart-rate-variability coefficients and 2.7% in skin-conductance coefficients, and two country samples trended positive rather than negative (Soroka et al. 2019).
- Warnings are ignored at high rates rather than overweighted: 70.2% click-through on Chrome's SSL warning across 25.4 million impressions (Akhawe & Felt 2013). Adding a negative message to a flow does not mean users will act on it.
- What fixes that is design, not stronger language: field adherence rose from 37% to 62% through visual cues favouring the safe action, while comprehension did not improve (Felt et al. 2015).
- In recovery, payout and manner carry different outcomes: compensation drives transaction-specific satisfaction, while cumulative satisfaction — the antecedent of loyalty — depends slightly more on interactional justice (Gelbrich & Roschk 2011).
- There is no evidence here for the wording advice itself. No controlled study of error-message tone surfaced in this research, so treat gentler copy as a hypothesis to test rather than a finding to apply.

## How to apply

- Rank the backlog by how often each failure fires rather than by how much delight the alternative adds, and count the failures you ship — failed payments, failed uploads, validation dead ends, empty searches, timeouts — as first-class metrics beside activation.
- Put the cause, the next action and the apology in the error message itself rather than in a support article, in the user's words, and never blame the user. Manner of handling is the half of recovery that feeds loyalty.
- Give any destructive or irreversible step opinionated layout: the safe action as the visually obvious control, the risky one as a plain secondary. That is the change that moved field adherence from 37% to 62%; clearer wording did not.
- Delete the warnings that fire on safe states and keep the one that fires on the dangerous one. Seventy per cent click-through on a serious warning is the baseline you are working against.
- Show negative reviews and low ratings in full, with the rating distribution, and let users sort by lowest first. Suppressing them is illegal in the UK and the US, and the rare negative review is the one buyers treat as credible.
- Build the recovery path into the failure screen: retry in place, the entered data preserved, a fallback offered (another payment method, a smaller file, manual entry), and a line on what happened to anything half-finished.
- On cancellation, downgrade and refund flows, resolve the problem first — a stated fix, credit or a pause — and put any offer second. The failure a user already had weighs more than the offer you make now.

## What to avoid

- Hiding or demoting bad reviews. Under paragraph 13 of Schedule 20 of the UK's Digital Markets, Competition and Consumers Act 2024, in force 6 April 2025, publishing reviews “in a misleading way” expressly includes failing to publish or removing negative reviews while publishing positive ones, and giving greater prominence to positive reviews over negative ones. In the US, 16 CFR § 465.7 makes it an unfair or deceptive practice to misrepresent that a review section shows most or all submitted reviews when reviews are suppressed for their rating or negative sentiment.
- Manufacturing the threat. Paragraph 14 of the same Schedule bans making a materially inaccurate claim about the risk to the personal security or safety of the consumer or their household if they do not buy. Inflated breach, data-loss or health warnings sit squarely here.
- Arguing that fear appeals do not work. They do, at d = 0.29 across 248 samples and 27,372 people, with no identified circumstances of backfire (Tannenbaum et al. 2015). The line against them is that frightening your own users to convert them is a cost you impose on them, not an ineffective tactic — and in the headline field test, fear words lowered clicks anyway.
- Quoting a multiplier. The five-to-one ratio traces to prediction work that loses its accuracy under cross-validation (Heyman & Slep 2001), and the 1.8–2.1 loss aversion coefficient is from monetary gambles (Brown et al. 2024). Neither is a conversion rate for your interface.
- Treating the average as the user. Country explained under 3% of the variance in individual physiological responses to negative news, and a sizable share of participants responded in the other direction (Soroka et al. 2019). Segment any negative-framed test before shipping it.
- Padding a flow with warnings, confirmations and risk copy. The measured outcome of over-warning is dismissal at up to 70%, which costs you the one warning that mattered.
- Dressing a system failure as a cheerful empty state. The asymmetry does not care about your illustration; it cares that the thing the user came for did not happen.

## Measuring it

- Primary metric: failure rate per critical flow — payment, upload, search, import, auth — and the completion rate of that flow conditional on a failure having occurred. Fixing a pain point shows up as the gap between those two closing.
- Secondary: recovery in the failure state itself — retry taken in place, data preserved, fallback used — plus support contacts per thousand failures. Alongside it, retention and repurchase split by whether a user hit a failure in their first session, where the attribute-level asymmetry predicts a larger decrement than any positive feature produces as a gain.
- Backfire metric: dismissal rate per warning type. A dismissal rate climbing toward the 70% seen in the browser field data means the warning has become wallpaper and should be removed or redesigned, not reworded harder.
- Backfire metric: for any negative-framed or risk-framed message, conversion by segment plus unsubscribes, complaints and uninstalls — individual variation is large enough that an overall win can hide a segment reacting the other way.
- Backfire metric on reviews: conversion and return rate after you expose the full rating distribution. A lower conversion rate with better-matched purchases is the expected trade; a falling conversion rate with a flat return rate means the display, not the honesty, is the problem.

## Live examples

- Lex gives the delete-account confirmation the only filled control to its safe answer — “Nevermind!” on yellow against a plain “Yes, delete” — and on the page behind it “Deactivate instead” is the filled green button while “Delete account” is a plain outline, so the layout, not the wording, points at the reversible path.
- Zomato names the amount, the card and what happens to the money in the failure message itself — a ₹0.90 payment on the card ending 2652, refunded within 24 hours if it was deducted — and offers “Retry payment” and “Try another payment method” on the same sheet, so the recovery sits where the order broke.
- Warby Parker states the cause of a billing failure inside the order it belongs to — the card was declined when the order was billed — and puts “Update card info” in the same red-bordered box, above the $21.98 order no. 546203419 and its items, so the fix is reachable from the report of the problem.
- Etsy shows the whole distribution behind an item's rating — 88% five star, 7% four, 2% three, 1% two, 2% one across 188 reviews — and lists “Lowest rating” beside “Highest rating” in the sort sheet, so the worst reviews are one tap from the average.

## References

- Baumeister, Bratslavsky, Finkenauer & Vohs (2001). Bad Is Stronger Than Good. Review of General Psychology. — https://doi.org/10.1037/1089-2680.5.4.323
- Heyman & Slep (2001). The Hazards of Predicting Divorce Without Crossvalidation. Journal of Marriage and Family. — https://doi.org/10.1111/j.1741-3737.2001.00473.x
- Mittal, Ross & Baldasare (1998). The Asymmetric Impact of Negative and Positive Attribute-Level Performance on Overall Satisfaction and Repurchase Intentions. Journal of Marketing. — https://doi.org/10.1177/002224299806200104
- Keaveney (1995). Customer Switching Behavior in Service Industries: An Exploratory Study. Journal of Marketing. — https://doi.org/10.1177/002224299505900206
- Chevalier & Mayzlin (2006). The Effect of Word of Mouth on Sales: Online Book Reviews. Journal of Marketing Research. — https://doi.org/10.1509/jmkr.43.3.345
- Chevalier & Mayzlin (2003). The Effect of Word of Mouth on Sales: Online Book Reviews. NBER Working Paper 10148. — https://www.nber.org/papers/w10148
- Robertson, Pröllochs, Schwarzenegger, Pärnamets, Van Bavel & Feuerriegel (2023). Negativity Drives Online News Consumption. Nature Human Behaviour. — https://doi.org/10.1038/s41562-023-01538-4
- Soroka, Fournier & Nir (2019). Cross-national Evidence of a Negativity Bias in Psychophysiological Reactions to News. PNAS. — https://doi.org/10.1073/pnas.1908369116
- Akhawe & Felt (2013). Alice in Warningland: A Large-Scale Field Study of Browser Security Warning Effectiveness. USENIX Security Symposium. — https://www.usenix.org/conference/usenixsecurity13/technical-sessions/presentation/akhawe
- Felt, Ainslie, Reeder, Consolvo, Thyagaraja, Bettes, Harris & Grimes (2015). Improving SSL Warnings: Comprehension and Adherence. CHI. — https://doi.org/10.1145/2702123.2702442
- Gelbrich & Roschk (2011). A Meta-Analysis of Organizational Complaint Handling and Customer Responses. Journal of Service Research. — https://doi.org/10.1177/1094670510387914
- Tannenbaum, Hepler, Zimmerman, Saul, Jacobs, Wilson & Albarracín (2015). Appealing to Fear: A Meta-analysis of Fear Appeal Effectiveness and Theories. Psychological Bulletin. — https://doi.org/10.1037/a0039729
- Brown, Imai, Vieider & Camerer (2024). Meta-analysis of Empirical Estimates of Loss Aversion. Journal of Economic Literature. — https://doi.org/10.1257/jel.20221698
- Digital Markets, Competition and Consumers Act 2024, Schedule 20, paragraphs 13 and 14 (UK banned practices: misleading review publication; inaccurate personal-safety claims). — https://www.legislation.gov.uk/ukpga/2024/13/schedule/20
- 16 CFR § 465.7 — Review suppression (FTC Rule on the Use of Consumer Reviews and Testimonials). — https://www.ecfr.gov/current/title-16/part-465/section-465.7

## Similar principles

- Loss aversion — /agents/loss-aversion.md
- Peak-end rule — /agents/peak-end-rule.md
- Framing effect — /agents/framing-effect.md
- Social proof — /agents/social-proof.md
- Regret aversion — /agents/regret-aversion.md
- Scarcity — /agents/scarcity.md