Lost in Translation: Why Lower Bias Scores Don't Mean Fairer Models

Harsh Kumar

cs.CLOct 14, 2026

Lost in Translation: Why Lower Bias Scores Don't Mean Fairer Models

Authors: Harsh Kumar

Organizations: Columbia University

Abstract

Bias audits of multilingual language models usually compare a model's average stereotype score in English with its average in another language and read the difference as a difference in fairness. We ask how much of that difference is noise. Auditing ten open models in eight languages, we measure how far scores move when English test items are merely reworded, test every language shift item by item, and check that the test catches shifts we inject on purpose. In all 70 model-language pairs the stereotype score falls below its English level, and for Hindi, Korean and Swahili the drop clearly exceeds what rewording English produces; most of it comes from StereoSet items. Judging every translation, we find that imperfect ones drop more, but faithful translations alone drop nearly as much. Still, a lower average does not show that a model is fairer. An item's English score predicts its translated score far less well than its reworded score (correlation 0.23, 0.27 on faithful translations, vs. 0.48), while scores stay as far from neutral, models agree on translated items, and within Hindi or Spanish a translated item's score correlates with its own rewording at 0.84 and 0.82. The average cannot distinguish lower bias from a different pattern of item-level scores. We also document three common pipeline errors that silently erase bias signal, and release our code.

Figures & tables

Figure 1: The test catches shifts it should catch. We inject shifts of known size into reworded-English pairs, where the true shift is zero, and rerun the full procedure. Shifts of 0.04 or more are found in most runs; there are no false alarms. The shaded band is the range of average gaps per language that we observe across languages (§ 4.1 ).

Language	soft	binary	mean	95% CI	soft	binary	slope	same choice	Fluency
	Score		Drop vs. English		Reliable drops		Carry-over
English	0.536	0.63	–	–	–	–	–	–	0.68
Spanish	0.528	0.57	0.007	[0.002, 0.011]	0/10	1/10	0.37	63%	0.60
French	0.519	0.57	0.017	[0.012, 0.021]	1/10	2/10	0.38	62%	0.59
Chinese	0.517	0.54	0.019	[0.012, 0.027]	3/10	8/10	0.36	58%	0.63
Arabic	0.516	0.56	0.020	[0.012, 0.027]	5/10	6/10	0.18	59%	0.59

Table 1: Results per language, averaged over the ten models. Score : mean stereotype score (soft) and share of items on which the stereotyped continuation wins (binary). Drop : how far the mean score falls below English (95% interval across models). Reliable drops : models whose drop survives a permutation test with false-discovery correction, using soft or binary scores. Carry-over : slope of the translated score on the English score for the same item (1 = fully kept), and how often the model picks the same continuation in both versions. Fluency : how well the model tells the two candidate continuations from the unrelated one (excluding BBQ, whose third option is an “unknown” answer). Last row: the same measures when English items are only reworded (median over models and the three machine rewordings).

Figure 2: How far each language’s average stereotype score falls below English, averaged over the ten models (95% interval across models). The dashed line is the typical noise floor from rewording English. Above each bar: how many of the ten models show a drop that survives the permutation test with false-discovery correction.

Figure 3: Average stereotype score per model and language (0.5 = no preference; red leans toward the stereotype). English is the reddest column for every model; languages are ordered by the size of their average drop.

Figure 4: The same item, changed in two ways. Each point is one item scored by one model, before (x-axis) and after (y-axis) the change; darker means more items. Left: the English item reworded. Right: the item translated. If the model’s preference carried over perfectly, points would lie on the dashed line. The orange line is the least-squares fit: slope 0.66 after rewording, 0.26 after translation.

Figure 5: Two checks of what the drop means. (a) How often the model picks the same continuation as in English: about 70% when the item is reworded, much closer to chance when it is translated. (b) How strongly the model prefers one continuation, whichever it is: as strong in other languages as in English (Hindi’s lower value is a length effect, § 4.2 ).

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Source	Categories (items)	Items
StereoSet	gender, profession, race, religion (50 each)	200
BBQ	age 25, gender identity 24, nationality 25, race/ethnicity 25, religion 26	125
Written	gender–profession 15, race 7, gender trait 5, age 5, religion 5	37
Total English items		362

Appendix

Table 2: English test items after removing the 25 BBQ items whose roles could not be recovered. Each is translated into seven languages.

All items
Rewording	Sim.	Same	Floor	Slope	Agree
via German	0.93	28%	0.002	0.74	74%
via Finnish	0.91	16%	0.008	0.70	71%
via Japanese	0.90	9%	0.010	0.53	65%
92-item subset (LLM rewrites)
via German	0.92	28%	0.030	0.72	72%

Appendix

Table 3: How the noise floor depends on the rewording. Sim. : embedding similarity to the original; Same : share of fields returned unchanged; Floor : median over models of the gap between English and its rewording; Slope , Agree : carry-over of the English preference, as in Table 1 .

Model	Focus	soft	binary	Gap	Floor	Gap $-$ floor [95% CI]	soft	binary	CLCI
		English score					Reliable drops
Aya-23-8B	Multiling.	0.521	0.62	0.010	0.006	$+0.004$ [ $-0.009$ , $+0.015$ ]	0/7	4/7	0.990
Aya-Expanse-8B	Multiling.	0.526	0.59	0.010	0.008	$+0.003$ [ $-0.012$ , $+0.017$ ]	0/7	1/7	0.990
Gemma-2-9B	English	0.537	0.62	0.015	0.007	$+0.008$ [ $-0.008$ , $+0.019$ ]	1/7	0/7	0.985
Llama-3.1-8B	English	0.534	0.64	0.017	0.008	$+0.009$ [ $-0.008$ , $+0.021$ ]	2/7	3/7	0.983
Yi-1.5-9B	Regional	0.533	0.62	0.019	0.001	$+0.018$ [ $+0.000$ , $+0.027$ ]	4/7	5/7	0.981

Appendix

Table 4: Results per model, ordered by gap. Gap : mean distance between a language’s average score and English’s, over the seven languages (the cross-lingual score shift, CSS). Floor : the same distance between English and reworded English. Gap

-

floor : with a 95% bootstrap interval over items; ∗ marks intervals above zero. Reliable drops : languages (of seven) whose drop survives the permutation test with false-discovery correction. CLCI

=1-

mean gap. These numbers describe consistency with English, not fairness, and should not be used to rank models for deployment.

Model	Spanish	French	Chinese	Arabic	Korean	Swahili	Hindi	Noise only
Aya-23-8B	0.004	0.013	0.010	0.009	0.010	0.011	0.012	0.007
Aya-Expanse-8B	0.002	0.009	0.011	0.005	0.002	0.028	0.015	0.010
BLOOMZ-7B	0.021	0.019	0.021	0.010	0.024	0.005	0.038	0.010
Gemma-2-9B	0.004	0.017	0.014	0.010	0.022	0.012	0.024	0.009
Llama-3.1-8B	0.008	0.014	0.017	0.017	0.019	0.023	0.024	0.008
Mistral-7B	0.004	0.016	0.027	0.032	0.038	0.038	0.035	0.007

Appendix

Table 5: How far each language’s average stereotype score falls below English, for every model (negative values: above English). Bold: the drop survives the permutation test with false-discovery correction. Noise only : the distance from zero expected if the item-level changes were pure noise, averaged over languages.

Figure 6: Each model’s average gap minus its own noise floor, with 95% bootstrap intervals over items. Intervals entirely right of zero indicate a gap larger than rewording English produces.

Items	Sp	Fr	Ch	Ar	Ko	Sw	Hi	Avg.
All	.007	.017	.019	.020	.026	.029	.030	.021
Faithful	.003	.010	.012	.016	.026	.022	.031	.017

Appendix

Table 6: Drop below English with all translations and with faithful ones only (back-translation similarity

\geq 0.9

, a subset of the first row). Avg. : equal weight per language. The reduction is 0.004 (95% paired bootstrap interval over items [-0.002, 0.009]).

Language	avg.	sum	avg.	sum	avg.	sum
	Drop		Reliable		Slope
Spanish	0.007	0.037	0	2	0.37	0.38
French	0.017	0.038	1	1	0.38	0.36
Chinese	0.019	0.065	3	4	0.36	0.32
Arabic	0.020	0.057	5	5	0.18	0.24
Korean	0.026	0.057	5	5	0.14	0.16

Appendix

Table 7: Main results with the per-token average log-probability (avg.) and with the summed log-probability (sum), which does not dilute a one-word contrast in a long continuation. Reliable : models (of 10; last row: pairs of 70) with a drop that survives the corrected permutation test. Summed scores are more extreme, so drops are larger; the drop and the overall carry-over hold, while carry-over by language shifts (Swahili falls, Arabic and Hindi rise).

	Models agree
Version	score	change	Spread	Center
English	0.50	–	–	–
Spanish	0.58	0.45	0.21	0.515
French	0.55	0.45	0.20	0.506
Chinese	0.51	0.38	0.22	0.503
Arabic	0.53	0.41	0.20	0.510

Appendix

Table 8: Are item-level scores outside English systematic or noise? Models agree : average correlation, over the 45 pairs of models, of their scores on the same items ( score ) and of the change from English ( change ). Spread : standard deviation of the item-level change. Center : fitted score of an item on which the model has no preference in English.

		Reliable drops		Drop	Carry-over		Same choice
Items	$n$	soft	binary	[95% CI]	transl.	rew.	transl.	rew.
StereoSet	200	41/70	18/70	0.027 [ $0.016$ , $0.037$ ]	0.15	0.49	56%	67%
BBQ	125	0/70	0/70	0.011 [ $-0.009$ , $0.033$ ]	0.32	0.72	61%	72%
Written	37	0/70	0/70	0.026 [ $-0.019$ , $0.068$ ]	0.24	0.69	63%	82%
BBQ + Written	162	0/70	1/70	0.015 [ $-0.004$ , $0.034$ ]	0.30	0.71	61%	74%

Appendix

Table 9: Results by source of the items. Reliable drops : model–language pairs (of 70) whose drop survives the permutation test with false-discovery correction within that source. Drop : average drop below English with a 95% bootstrap interval over items (the change is averaged over models and languages within each item). Carry-over and Same choice : for translations and for English rewordings. The last row excludes StereoSet entirely.

Language	beyond rew.	$q$	kept	beyond rew.	$q$	beyond rew.	$q$
	All items		Fully preserved only			BBQ + written
Spanish	.006 [ $-$ .009, .021]	.439	266/349	.001 [ $-$ .017, .020]	.887	$-$ .012 [ $-$ .043, .020]	.635
French	.016 [.000, .032]	.050	295/360	.004 [ $-$ .013, .021]	.758	.009 [ $-$ .020, .038]	.636
Chinese	.018 [.001, .036]	.050	251/361	.011 [ $-$ .008, .032]	.491	.022 [ $-$ .010, .054]	.420
Arabic	.018 [.003, .034]	.038	272/361	.008 [ $-$ .009, .025]	.491	.005 [ $-$ .023, .034]	.717
Korean	.025 [.008, .041]	.008	197/361	.022 [.002, .042]	.139	.023 [ $-$ .007, .054]	.420

Appendix

Table 10: Item-level test of each language against English rewording. The change is first averaged over the ten models within each item, so that items, not model–item pairs, are the unit. beyond rew. : how much more the language drops below English than rewording does, with a 95% bootstrap interval over items;

q

: paired sign-flip permutation test over items, false-discovery corrected. Fully preserved only : translations whose contrast was judged fully preserved ( kept : fully preserved out of all judged translations). BBQ + written : all items from these two sources, without StereoSet.

	Models agree on change
Language	raw	$-$ length	$-$ length, freq.	$R^{2}$
Spanish	0.45	0.48	0.47	0.097
French	0.45	0.47	0.46	0.121
Chinese	0.38	0.38	0.37	0.122
Arabic	0.41	0.41	0.41	0.067
Korean	0.47	0.48	–	0.049

Appendix

Table 11: Do surface properties drive the shared item-level changes? Average correlation between models of the change from English, before and after regressing out differences in token count between the two continuations and in overall length ( length ), and additionally in word frequency ( freq. ; no frequency data for Korean and Swahili).

R^{2}

: variance in the change explained by these properties. Last two rows: items whose two continuations have the same number of tokens in both languages, and all pairs.

	All translated items			Fully preserved only
Items	$r$ [95% CI]	SD ratio	slope [95% CI]	$r$ [95% CI]	SD ratio
All languages	0.23 [0.20, 0.27]	1.13	0.26 [0.22, 0.31]	0.27 [0.23, 0.31]	1.04
Spanish	0.31 [0.23, 0.38]	1.21	0.37 [0.28, 0.46]	0.32 [0.23, 0.40]	1.16
French	0.32 [0.25, 0.39]	1.17	0.38 [0.29, 0.46]	0.35 [0.27, 0.42]	1.10
Chinese	0.28 [0.22, 0.33]	1.32	0.36 [0.29, 0.43]	0.33 [0.26, 0.38]	1.22
Arabic	0.18 [0.12, 0.24]	1.01	0.18 [0.12, 0.25]	0.19 [0.11, 0.26]	0.92

Appendix

Table 12: Association between an item’s score before and after a change: correlation

r

, ratio of standard deviations, and least-squares slope (

=r\times

SD ratio), each with a 95% bootstrap interval over items. Translated rows compare English with the translation; English rewording compares English with its rewordings; the Hindi and Spanish rewording rows compare a translation with its rewording in the same language. Low slopes after translation come from low correlation; the SD ratio does not fall consistently below 1.

Dimension	$n$	Drop [95% CI]	$r$ tr.	$r$ rew.	Rev.
Gender	94	.039 [.017, .061]	0.20	0.58	21%
Race/eth.	82	.013 [ $-$ .010, .036]	0.27	0.53	21%
Religion	81	.012 [ $-$ .008, .030]	0.19	0.35	24%
Profession	50	.016 [.000, .032]	0.09	0.26	22%
Age	30	.014 [ $-$ .027, .057]	0.33	0.40	26%
Nationality	25	.034 [ $-$ .002, .072]	0.22	0.57	35%

Appendix

Table 13: Results by social dimension (pooled over sources). Drop : mean drop below English, bootstrap over items.

r

: correlation of item scores with English after translation and after rewording. Rev. : share of model–item pairs leaning stereotyped in English (score

>0.55

) that lean counter-stereotyped after translation (

<0.45

Stage	Count
StereoSet intrasentence items sampled	200
BBQ ambiguous items sampled / kept (roles recovered)	150 / 125
Items written by the authors	37
English items	362
Translated items (7 languages)	2,534
removed: contrast collapsed	37

Appendix

Table 14: Items at each stage. Every translated or reworded item is compared only with its own English original, scored by the same model.

Change	Slope	Same	Items
Hindi items (129)
Hindi $\rightarrow$ reworded Hindi	0.83	83%	129
English $\rightarrow$ Hindi	0.16	–	129
English $\rightarrow$ reworded English	0.69	–	129
Spanish items (101)
Spanish $\rightarrow$ reworded Spanish	0.82	81%	101

Appendix

Table 15: Top: carry-over of item-level preferences within a language (each translated item reworded in the same language by Aya-Expanse-32B, which is not among the audited models; 129 Hindi and 101 Spanish items passed validation), across languages, and within English by machine round-trip rewording, all on the same items. Same : same choice as before the change. Bottom: StereoSet items scored as sentences on their own, without the template that contains the blank.

	English		Below		Slope
Errors present	soft	bin.	En.	Rel.	rew.	trans.
All three errors	0.493	0.42	35/70	0/70	0.47	0.21
Only label errors	0.490	0.44	33/70	0/70	0.56	0.28
Only scorer error	0.534	0.59	70/70	28/70	0.48	0.21
None (ours)	0.536	0.63	70/70	28/70	0.66	0.26

Appendix

Table 16: The same analysis under each combination of errors. English : mean soft and binary score on English items. Below En. : model–language pairs scoring below English; Rel. : of these, reliable after correction. Slope : carry-over for rewordings and translations. With the label errors, every gap looks like noise.

Model	En	Es	Fr	Zh	Ar	Ko	Hi	Sw
Qwen-2.5-7B	0.93	0.72	0.47	0.73	0.74	0.52	0.46	0.41
Aya-23-8B	0.83	0.32	0.42	0.88	0.78	0.70	0.43	0.54
Mistral-7B	0.83	0.51	0.50	0.77	0.42	0.54	0.47	0.33

Appendix

Table 17: One item in eight languages: “The elderly woman tried to set up her new smartphone.” followed by “She was confused by the interface and needed constant help.” (stereotyped) or “She configured it independently using the online guide.” Bold: the model prefers the stereotyped continuation. Every model leans strongly toward it in English; after translation the preference holds in some languages, reverses in others, and differs between models.

Language Model Bias Evaluation

Language model bias evaluation tests how models’ outputs, decisions, and representations vary across social groups, cultures, languages, and contexts.

See all 473 papers

Redirected, Not Removed: Task-Dependent Stereotyping Reveals the Limits of LLM Alignments
Apr 3, 2026Divyanshu Kumar, Ishita Gupta, Nitin Aravind Birur +3Social Bias in Language Models Language Model Bias Evaluation
Enkrypt AI
LLM Evaluators are Biased across Languages
Jul 16, 2026Ej Zhou, Lucas Resck, Zheng Hui +1Multilingual Language Model Evaluation Language Model Safety Evaluation
Language Technology Lab, University of Cambridge
Side-by-side Comparison Amplifies Dialect Bias in Language Models
May 23, 2026Kritee Kondapally, Claire J. Smerdon, Pooja C. Patel +5Social Bias in Language Models Algorithmic Fairness
University of Southern California, USA

Apr 3, 2026J/K move · Enter open · S save

Redirected, Not Removed: Task-Dependent Stereotyping Reveals the Limits of LLM Alignments

Divyanshu Kumar, Ishita Gupta, Nitin Aravind Birur, Tanay Baswa+2

Enkrypt AI

Social Bias in Language Models Language Model Bias Evaluation LLM Safety Alignment Linguistic Bias in Language Models

Open paper AskPDF arXiv

How biased is a language model? The answer depends on how you ask. A model that refuses to choose between castes for a leadership role will, in a fill-in-the-blank task, reliably associate upper castes with purity and lower castes with lack of hygiene. Single-task benchmarks miss this because they capture only one slice of a model's bias profile. We introduce a hierarchical taxonomy covering 9 bias types, including under-studied axes like caste, linguistic, and geographic bias, operationalized through 7 evaluation tasks that span explicit decision-making to implicit association. Auditing 7 commercial and open-weight LLMs with \textasciitilde45K prompts, we find three systematic patterns. First, bias is task-dependent: models counter stereotypes on explicit probes but reproduce them on implicit ones, with Stereotype Score divergences up to 0.43 between task types for the same model and identity groups. Second, safety alignment is asymmetric: models refuse to assign negative traits to marginalized groups, but freely associate positive traits with privileged ones. Third, under-studied bias axes show the strongest stereotyping across all models, suggesting alignment effort tracks benchmark coverage rather than harm severity. These results demonstrate that single-benchmark audits systematically mischaracterize LLM bias and that current alignment practices mask representational harm rather than mitigating it.

Abstract

Language Model Bias Evaluation

Explore similar work

Redirected, Not Removed: Task-Dependent Stereotyping Reveals the Limits of LLM Alignments

LLM Evaluators are Biased across Languages

Side-by-side Comparison Amplifies Dialect Bias in Language Models