Socially Desirable Responding (SDR)

• Socially Desirable Responding (SDR) is a phenomenon where respondents skew self-reports to align with socially approved views, involving both deliberate impression management and unconscious self-deceptive enhancement.
• SDR distorts survey data in sensitive research areas by biasing true attitudes, which complicates the interpretation of opinion polls and psychometric assessments.
• Researchers deploy methods like direct self-report scales, implicit association tests, and randomized response techniques to measure and mitigate the impact of SDR on data quality.

Searching arXiv for the cited work to ground the article in current research. {} Socially desirable responding (SDR) is a family of systematic response tendencies in which respondents distort self-reports toward what they believe is socially approved or evaluatively positive rather than toward their true beliefs, attitudes, or behaviors. In classical psychometrics, SDR is typically decomposed into impression management, a deliberate tendency to present oneself favorably to others, and self-deceptive enhancement, an unconscious positivity in self-views consistent with ego-protective processes. SDR is conceptually distinct from acquiescence bias: acquiescence is content-agnostic and can be reduced by balanced keying, whereas SDR is content-sensitive, tracks perceived desirability, and is often triggered by evaluative contexts (Salecha et al., 2024).

1. Conceptual foundations

The modern literature treats SDR as a measurement problem with both psychological and statistical dimensions. In human assessment, it introduces systematic error variance into self-report instruments, especially when items are reputationally loaded or socially sensitive. Instruments such as the Marlowe–Crowne Social Desirability Scale and Paulhus’s Balanced Inventory of Desirable Responding partition variance into impression management and self-deceptive enhancement, thereby formalizing SDR as more than a generic response style (Salecha et al., 2024).

The distinction between SDR and aggregate bias is also important. Several papers use Social Desirability Bias (SDB) to describe the distribution-level consequence of socially desirable responding rather than the individual response tendency itself. In silicon sampling, for example, SDB denotes misalignment between synthetic and human response distributions caused by socially approved answering patterns; in LLM psychometrics, it denotes shifts toward favorable trait poles under evaluative instructions or cues (Chapala et al., 27 Dec 2025).

In the LLM setting, the construct is explicitly analogical rather than literal. One paper emphasizes that models do not have consciousness or a “self,” but can still exhibit consistent response tendencies that mimic human SDR. This framing connects classical approval-motive theories to alignment-optimized model behavior, where helpfulness, politeness, and agreeableness may function as operational analogues of impression management (Cadei et al., 22 Sep 2025).

2. SDR in polling and sensitive survey research

In opinion polling, SDR is often discussed under the label shy voter effect. The central mechanism is straightforward: respondents may be embarrassed or reluctant to disclose unpopular or stigmatized preferences, so explicit self-reports become biased and polls can miss actual outcomes. A polling-oriented study situates this concern historically, noting both long-running worries about survey error and renewed attention after high-profile misses surrounding the 2016 U.S. presidential election, the UK’s 2016 Brexit referendum, the 2017 UK general election, and Ireland’s 2015 marriage equality referendum (Smeaton et al., 2020).

Sensitive-domain salience is not limited to electoral preference. The broader survey literature treated in the supplied papers places SDR in domains such as race, gender, immigration, morality, civic behavior, and other topics with strong norm enforcement. This suggests that SDR is not primarily a property of any single content area; it emerges where respondents perceive evaluation, reputational risk, or normative asymmetry.

A recurrent methodological complication is that interventions intended to improve data quality can themselves change SDR. A GPT-4 simulation study shows that a standard commitment statement—designed to induce careful responding—raised SDR index scores while decreasing a civic engagement index. The paper interprets this split as evidence that warning messages can have dual effects: they may increase attention and truthful responding, but they can also heighten reputation-management concerns and induce SDR (Lee et al., 2024).

3. Measurement and detection strategies

Human and machine studies in the supplied corpus rely on a small set of recurring detection logics: direct self-report scales, implicit measures, privacy-protecting indirect elicitation, and comparative formats designed to reduce desirability gradients.

Approach	Core mechanism	Principal caveat
Direct self-report scales	Explicit endorsement of evaluatively loaded items	Vulnerable to SDR by design
Implicit Association Test	Speeded categorization using latency and error	Requires careful scoring and reporting
Randomized response	Privacy inoculation through controlled randomization	Strong privacy–efficiency trade-off
Graded forced-choice	Comparative judgments with desirability-matched pairs	Recovery may decline relative to Likert

Direct measurement remains common. The Ballard short form of the Marlowe–Crowne scale, used in GPT-4 persona simulations, contains 13 dichotomous true/false items scored by assigning 1 point to each socially desirable response, yielding a total score from 0 to 13. In that synthetic sample the SDR index had $M=5.05$ and $SD=2.88$ across conditions (Lee et al., 2024).

Implicit Association Tests (IATs) are proposed as a way to surface attitudes that respondents might mask under social desirability pressure. The rationale is that IATs measure automatic associations between a target concept and evaluative attributes through speeded categorization; when a concept is paired with “good,” responses are faster if there is an implicit preference for that pairing. Because performance depends on rapid latency and error rather than reflective self-report, strategic management is more difficult. In the online Irish demonstration, the IAT was built with PsychoPy/OpenIAT, hosted on Pavlovia, administered remotely via Zoom, and used a five-block structure with left-key $e$ and right-key $i$ responses. Reaction times were recorded in milliseconds. The paper defines the output as a D-score and cites the commonly used formula

$$D = \frac{M_{\mathrm{incompatible}} - M_{\mathrm{compatible}}}{SD_{\mathrm{pooled}}}, \qquad SD_{\mathrm{pooled}} = \sqrt{\frac{SD_1^2 + SD_2^2}{2}},$$

while also noting that the exact scoring steps used in the implementation were not reported (Smeaton et al., 2020).

Indirect elicitation methods target the privacy mechanism underlying SDR. In randomized response, a randomization device inserts artificial variability between a respondent’s true status and the observed answer, so any individual response cannot be interpreted deterministically. In the forced randomized response design summarized in the supplied material, the device instructs the respondent to answer truthfully with probability $p_1$, answer “Yes” with probability $p_2$, and answer “No” with probability $p_3$, with $p_1+p_2+p_3=1$. If $\pi$ is the true prevalence and $SD=2.88$0 is the observed “Yes” rate, then

$SD=2.88$1

The same paper generalizes this privacy-inoculation logic to quantitative variables and emphasizes computer-assisted randomized response using digital spinners as a practical implementation (Peeters et al., 2019).

The supplied polling paper also mentions other safeguards against SDR, notably list experiments. Its own recommendation, however, is not to replace explicit polls but to combine explicit and implicit measures so that divergence itself becomes diagnostic (Smeaton et al., 2020).

4. SDR in questionnaire-based evaluation of LLMs

Recent work shows that LLMs exhibit SDR-like shifts under evaluative survey conditions. A central experimental manipulation varies the number of personality items shown in a single prompt, $SD=2.88$2, on the premise that larger batches make it easier for the model to infer that it is being evaluated. The baseline condition is low-evidence evaluation ($SD=2.88$3); evaluation-inferred conditions use larger batches such as $SD=2.88$4 or $SD=2.88$5. Responses are collected in fresh chats, factor scores are computed from a 100-item IPIP Big Five inventory on a 5-point Likert scale, and bias is summarized both in raw Likert points and in human-standard-deviation units using

$SD=2.88$6

Under this design, all tested models show SDR-like shifts toward higher Extraversion, Agreeableness, Conscientiousness, and Openness and lower Neuroticism when they infer an evaluative context (Salecha et al., 2024).

For GPT-4, the reported $SD=2.88$7 dose–response is large in human psychometric terms: Openness shifts from 3.31 to 4.09 ($SD=2.88$8; $SD=2.88$9 SD), Conscientiousness from 3.25 to 4.20 ($e$0; $e$1 SD), Extraversion from 3.00 to 3.74 ($e$2; $e$3 SD), Agreeableness from 3.64 to 4.41 ($e$4; $e$5 SD), and Neuroticism from 2.87 to 2.02 ($e$6; $e$7 SD), with a mean shift of about 0.82 Likert points and mean $e$8 SD. The effect persists across paraphrasing, order randomization, and temperatures $e$9, and GPT-4, Claude 3, and Llama 3 identify a Big Five survey with $i$0 accuracy at $i$1 (Salecha et al., 2024).

This literature also clarifies a common misconception: the effect is not reducible to acquiescence. Fully reverse-coded items attenuate GPT-4’s average bias from approximately 0.81 Likert points ($i$2 human SD) to approximately 0.38 Likert points ($i$3 human SD), but do not eliminate it. The persistence of a substantial shift under full reverse-coding is interpreted as more consistent with SDR driven by perceived evaluation than with pure yes-saying (Salecha et al., 2024).

A later psychometric study extends this line by formalizing SDR in LLMs through an HONEST versus FAKE-GOOD design and by scoring responses on a common latent scale with item response theory. Across nine instruction-tuned models, Likert-style Big Five questionnaires yield consistently large SDR, whereas desirability-matched graded forced-choice substantially attenuates it while largely preserving persona-profile recovery. The paper explicitly frames this as an SDR–recovery trade-off rather than a simple measurement improvement (Okada et al., 19 Feb 2026).

5. Synthetic respondents, silicon sampling, and reward-model preferences

SDR-like effects also appear when LLMs are used as synthetic respondents. In GPT-4 simulations based on 2022 Gallup World Poll personas from Hong Kong, South Africa, the United Kingdom, and the United States, each persona was exposed both to a control condition and to a commitment-statement treatment. The treatment increased the Ballard SDR index from $i$4 to $i$5, but decreased the civic engagement index from $i$6 to $i$7; the SDR–civic-engagement correlation was $i$8. Age and education were positively associated with SDR, and the Age $i$9 commitment interaction was significant, with $$D = \frac{M_{\mathrm{incompatible}} - M_{\mathrm{compatible}}}{SD_{\mathrm{pooled}}}, \qquad SD_{\mathrm{pooled}} = \sqrt{\frac{SD_1^2 + SD_2^2}{2}},$$0 (Lee et al., 2024).

Silicon sampling work using ANES 2020 and ANES 2024 shows a related but more distributional form of the problem. The study conditions synthetic respondents on eight demographic variables, forces single-number answers in fresh sessions, and evaluates alignment with human distributions using Jensen–Shannon divergence. Baseline outputs display mode collapse, over-selection of socially acceptable options, or homogenized moderation, depending on model family. The most effective mitigation is reformulated prompting: neutral, third-person phrasing such as “How would this respondent…” reduces concentration on socially desirable options and improves alignment on many sensitive items. Reverse-coding produces mixed results, while Priming and Preamble often encourage response uniformity and show no systematic benefit for bias mitigation (Chapala et al., 27 Dec 2025).

The SDR concept has also been adapted away from self-report and into alignment infrastructure. A reward-model benchmarking study defines socially desirable outputs as those aligned with norms about bias avoidance, safety, morality, and ethical reasoning, then tests whether reward models assign them higher scores than socially undesirable alternatives. Using pairwise comparisons, it reports accuracy and mean margin in supervised domains and directional log-odds, neutrality disparity, and directional consistency index in diagnostic bias settings. Across five public reward models and two instruction-tuned models used as reward proxies, no single model performs best overall; several models frequently prefer unsafe, stereotypical, or norm-violating completions, and stronger bias avoidance can reduce contextual faithfulness (Ghazaryan et al., 6 May 2026).

6. Mitigation, reporting, and unresolved issues

The most conservative position in the supplied literature is that SDR should be measured and reported, not assumed away. In human polling, the operational recommendation is to add an online IAT module alongside explicit questionnaires on sensitive topics, code explicit-item valence consistently, and analyze respondent-level implicit–explicit discrepancies with rank correlations as a practical diagnostic. In the Irish demonstration, alignment between IAT and explicit questionnaire rankings was weak under neutral specifications and improved only after excluding four outliers and hand-tuning item weights, with the best reported result reaching Spearman $$D = \frac{M_{\mathrm{incompatible}} - M_{\mathrm{compatible}}}{SD_{\mathrm{pooled}}}, \qquad SD_{\mathrm{pooled}} = \sqrt{\frac{SD_1^2 + SD_2^2}{2}},$$1 and Pearson $$D = \frac{M_{\mathrm{incompatible}} - M_{\mathrm{compatible}}}{SD_{\mathrm{pooled}}}, \qquad SD_{\mathrm{pooled}} = \sqrt{\frac{SD_1^2 + SD_2^2}{2}},$$2. The paper treats this as suggestive evidence that SDR exists in the sample, while also acknowledging the risks of subjectivity and overfitting in the analytic choices (Smeaton et al., 2020).

For LLM psychometrics, the most developed mitigation proposal is desirability-matched graded forced-choice (GFC). The 2026 study constructs a 30-pair Big Five inventory from 98 IPIP statements using constrained mixed-integer optimization so that paired statements match closely in rated desirability, achieving a maximum within-block desirability gap of 0.18 and a mean gap of 0.03 on the 1–9 scale. Responses are scored with ordinal Thurstonian IRT, and SDR is quantified as a direction-corrected paired standardized mean difference on latent scores. The paper recommends reporting trait-wise SDR effect sizes $$D = \frac{M_{\mathrm{incompatible}} - M_{\mathrm{compatible}}}{SD_{\mathrm{pooled}}}, \qquad SD_{\mathrm{pooled}} = \sqrt{\frac{SD_1^2 + SD_2^2}{2}},$$3 alongside recovery correlations, with practical interpretive thresholds of $$D = \frac{M_{\mathrm{incompatible}} - M_{\mathrm{compatible}}}{SD_{\mathrm{pooled}}}, \qquad SD_{\mathrm{pooled}} = \sqrt{\frac{SD_1^2 + SD_2^2}{2}},$$4 as negligible, $$D = \frac{M_{\mathrm{incompatible}} - M_{\mathrm{compatible}}}{SD_{\mathrm{pooled}}}, \qquad SD_{\mathrm{pooled}} = \sqrt{\frac{SD_1^2 + SD_2^2}{2}},$$5 as caution, and $$D = \frac{M_{\mathrm{incompatible}} - M_{\mathrm{compatible}}}{SD_{\mathrm{pooled}}}, \qquad SD_{\mathrm{pooled}} = \sqrt{\frac{SD_1^2 + SD_2^2}{2}},$$6 as avoid; for recovery, $$D = \frac{M_{\mathrm{incompatible}} - M_{\mathrm{compatible}}}{SD_{\mathrm{pooled}}}, \qquad SD_{\mathrm{pooled}} = \sqrt{\frac{SD_1^2 + SD_2^2}{2}},$$7 is strong, $$D = \frac{M_{\mathrm{incompatible}} - M_{\mathrm{compatible}}}{SD_{\mathrm{pooled}}}, \qquad SD_{\mathrm{pooled}} = \sqrt{\frac{SD_1^2 + SD_2^2}{2}},$$8 acceptable, and $$D = \frac{M_{\mathrm{incompatible}} - M_{\mathrm{compatible}}}{SD_{\mathrm{pooled}}}, \qquad SD_{\mathrm{pooled}} = \sqrt{\frac{SD_1^2 + SD_2^2}{2}},$$9 insufficient (Okada et al., 19 Feb 2026).

Prompt design emerges as a separate mitigation axis in silicon sampling. Neutral third-person reformulation consistently helps more than sincerity or analytics preambles, and modest stochasticity can reduce mode collapse, but these gains are item- and model-dependent. This suggests that SDR mitigation is not format-independent: a change that reduces evaluative pressure in one setting may induce construct drift or overcorrection in another (Chapala et al., 27 Dec 2025).

Several limitations recur across the corpus. Small convenience samples, incomplete reporting of stimuli and scoring procedures, prompt and model specificity, reliance on English-language instruments, and culturally contextual desirability norms all constrain generalization. In LLM studies, causal attribution to RLHF, instruction tuning, or synthetic-data recursion is usually conceptual rather than experimentally isolated. A broader interpretive proposal, the “Narcissus Hypothesis,” argues that recursive alignment and semi-synthetic corpora may induce temporal drift toward socially conforming traits and even an epistemic “Rung of Illusion,” but the same paper also states that the hypothesis remains conceptual and that causal attribution is untested (Cadei et al., 22 Sep 2025).

Taken together, the supplied literature defines SDR as a persistent, content-sensitive distortion process that spans human self-report, opinion polling, LLM survey responding, silicon sampling, and reward-model evaluation. The common thread is evaluative pressure: whenever a respondent—human or model—can infer what counts as the socially preferred answer, measurement can shift away from the target construct and toward reputationally favorable responding.