Bias Transfer in AI Systems

• Bias transfer is the phenomenon where biases present in source AI models persist during adaptation to downstream tasks, affecting fairness and performance.
• Studies reveal that mechanisms such as fixed-feature transfer, prompt adaptation, and multilingual alignment result in varied bias propagation across domains.
• Mitigation strategies, including controlled model retention, modified adaptation objectives, and data-centric debiasing, are critical to manage bias transfer.

Searching arXiv for recent and foundational papers on bias transfer across transfer learning, prompting, multilingual transfer, and debiasing. arXiv search query: "bias transfer transfer learning prompting multilingual debiasing"

Bias transfer denotes the persistence or propagation of a learned bias under adaptation, reuse, or transfer. In the contemporary fairness literature, it usually refers to the way social, demographic, dataset, or spurious-feature biases in a source model, representation, prompt, or training corpus reappear in downstream models, tasks, languages, or even human-authored text after AI assistance. In transfer learning and optimization, the same expression is also used in an older and broader sense for the deliberate transfer of an inductive bias or structural prior from previous training runs or pre-trained parameters to a new task. The literature therefore treats bias transfer as a family of phenomena rather than a single mechanism, ranging from harmful persistence of stereotypes to beneficial retention of source-task structure (Salman et al., 2022, Mackraz et al., 2024, Li et al., 2018, Pelikan et al., 2012).

1. Conceptual scope

A widely cited formalization defines bias transfer as the phenomenon in which biases present in a pre-trained source model persist even after the model is adapted to a downstream target task (Salman et al., 2022). That definition was developed in transfer learning settings involving pre-training and fine-tuning, but closely related formulations now appear in prompt adaptation, cross-lingual transfer, multi-task learning, and human–AI interaction.

In prompt-adapted causal LLMs, the relevant question is framed by the bias transfer hypothesis: whether intrinsic biases in a pre-trained model correlate with biases expressed after zero-shot or few-shot prompting (Mackraz et al., 2024). In fair multi-task learning, bias transfer is defined as the increase in fairness violation for a task under joint training relative to single-task learning, and is operationalized through the discrimination gain

$$DG(t)=F_{viol}^{(t)}(\mathcal{M})-F_{viol}^{(t)}(\mathcal{H}),$$

where positive $DG(t)$ means that multi-task learning increases unfairness for task $t$ (Roy et al., 2023).

A distinct but historically important usage appears in work on transfer learning with CNNs, where bias transfer refers to the leverage of inductive bias from a pre-trained source model during target-task training (Li et al., 2018). In hierarchical BOA, prior runs are used to bias future model building through soft distance-based statistics (Pelikan et al., 2012). This suggests that the term has two major senses: unwanted propagation of unfairness or shortcut structure, and deliberate transmission of useful priors.

2. Mechanisms and conditions

The mechanisms proposed for bias transfer depend strongly on the adaptation regime. In the analytical framework of transfer learning developed by Salman et al., a source model can remain sensitive to a bias direction that is insufficiently represented in the span of target data; fixed-feature transfer is especially vulnerable because retraining only the last layer leaves biased feature directions largely intact, whereas full-network fine-tuning lowers attack success rate but can still leave significant residual bias (Salman et al., 2022). The same study shows that bias transfer can arise in realistic settings such as ImageNet pre-training and can occur even when the target dataset is explicitly de-biased (Salman et al., 2022).

Prompt adaptation exhibits a different mechanism. For pronoun co-reference resolution with Mistral, Falcon, and Llama, occupation-level selection biases in the intrinsic model and the prompted model are strongly correlated, with $\rho \geq 0.94$ for zero- and few-shot prompting, $\rho \geq 0.92$ even when models are explicitly prompted to be fair or biased, and $\rho \geq 0.97$ when few-shot length and stereotypical composition are varied (Mackraz et al., 2024). In that setting, prompting modulates surface behavior but does not disrupt the underlying structure of occupational gender bias.

Cross-lingual representation learning introduces alignment-dependent mechanisms. In multilingual embeddings, the magnitude of gender bias changes when embeddings are aligned to different target spaces, and the alignment direction can influence downstream bias transfer (Zhao et al., 2020). Aligning to gender-rich target spaces reduces source bias, while aligning gender-rich languages to English increases their bias (Zhao et al., 2020). By contrast, the CLIP study finds no consistent trend linking upstream and downstream social bias, and attributes that inconsistency to convergence of representation spaces during downstream adaptation: after adaptation to the same frozen LLM, mean cosine similarity between model embedding spaces rises from $0.94$ to $0.994$, while the standard deviation shrinks from $0.017$ to $0.003$ (Ramos et al., 25 Aug 2025). Bias transfer is therefore neither universal nor monotone.

3. Evaluation frameworks

Because bias transfer is instantiated differently across settings, its measurement is heterogeneous. In source-to-target transfer with synthetic or natural interventions, the core metric is the attack success rate

$DG(t)$0

which quantifies how often a bias-inducing intervention flips a correct prediction (Salman et al., 2022).

In prompt-adapted LLMs, the key fairness metric is selection bias,

$DG(t)$1

and bias transfer is quantified by the Pearson correlation coefficient $DG(t)$2 between vectors of occupation-level SBs in the pre-trained and adapted models (Mackraz et al., 2024). In multilingual embeddings, the literature separates intrinsic and extrinsic views: intrinsic bias is measured in the shared embedding space via distances between occupation words and gender seed sets, while extrinsic bias is measured as the average absolute accuracy gap in downstream occupation prediction between masculine and feminine biography groups (Zhao et al., 2020).

Domain class imbalance in transfer learning is evaluated through accuracy and F1 scores reported separately for each class, rather than only on average. The central diagnostic is the gap in minor-class F1 between the source and transferred target domains under different source/target class ratios such as $DG(t)$3, $DG(t)$4, $DG(t)$5, $DG(t)$6, and $DG(t)$7 (Li, 2021). Cross-lingual sentiment analysis instead uses counterfactual evaluation, with bias defined as the average difference in predicted sentiment for paired sentences differing only in a protected demographic variable (Goldfarb-Tarrant et al., 2023). In fair multi-task learning, discrimination gain compares fairness violation under MTL and STL (Roy et al., 2023). These protocols share a common principle: aggregate accuracy is insufficient when the transferred bias is subgroup-specific, class-specific, or localized in feature space.

4. Natural language and multilingual systems

In text classification under transfer learning, bias transfer has been studied through domain class imbalance (DCI), where the class ratio in the source domain differs from that in the target domain. The reported effect is that classifiers can achieve deceptively high accuracy by over-predicting the majority class, while minor-class performance deteriorates; standard adaptation methods do not automatically correct class-ratio shifts, and both traditional and deep transfer learning models can propagate or amplify the resulting class bias (Li, 2021). The proposed Distance-based Backpropagation ADDA (DBA) addresses this by reweighting gradients using target–source feature similarity and improves minor-class performance under imbalance without sacrificing major-class performance, with the largest gains at $DG(t)$8, $DG(t)$9, and $t$0 ratios (Li, 2021).

Cross-lingual transfer can also worsen social bias. Across Japanese, Simplified Chinese, Spanish, German, and English, systems using zero-shot cross-lingual transfer usually become more biased than their monolingual counterparts, and racial biases are reported as more prevalent than gender biases (Goldfarb-Tarrant et al., 2023). The study attributes most of the increase to multilingual pre-training rather than English supervision alone and releases 1,525 distinct sentiment models together with intermediate checkpoints and evaluation code (Goldfarb-Tarrant et al., 2023).

At the representation level, multilingual embeddings contain measurable gender bias in English, Spanish, German, and French, and the downstream occupation-prediction bias reflects the bias of the chosen shared embedding space (Zhao et al., 2020). Aligning English to Spanish reduces bias by about $t$1, whereas mapping Spanish to English increases bias from $t$2 to $t$3; pre-align debiasing of English and subsequent alignment lower downstream bias with minimal performance drop, and out-of-the-box M-BERT can still show notable bias in cross-lingual transfer (Zhao et al., 2020). A separate mBERT study finds that cross-lingual transfer of debiasing techniques is feasible across English, French, German, and Dutch, with SentenceDebias reducing bias by an average of $t$4 and no performance disadvantages for the non-English languages included in the analyses (Reusens et al., 2023).

Prompt adaptation extends the problem beyond supervised fine-tuning. In causal LLMs, intrinsic gender biases in Mistral, Falcon, and Llama remain strongly correlated with downstream prompted behavior on WinoBias even under fairness-inducing prompts (Mackraz et al., 2024). In LLM-assisted student writing, gender-biased assistance transfers into human-authored career-plan essays: the biased condition yields a larger agentic gap and a higher stereotype congruence rate than the control and neutral conditions, and the effect is asymmetric because agency is suppressed in essays about female subjects while male-target writing remains largely unaffected (Hossain et al., 14 Jun 2026). Neutral prompting reduces stereotyping below the control condition in that experiment, but prompt engineering does not eliminate the broader concern that structural biases survive prompt-based deployment (Mackraz et al., 2024, Hossain et al., 14 Jun 2026).

5. Vision, audio, and multi-task learning

Vision studies show both strong and inconsistent forms of bias transfer. Salman et al. report synthetic backdoor transfer, natural co-occurrence bias transfer, demographic bias transfer in CelebA, and inherited ImageNet biases that affect downstream classifiers even when target data do not contain the source spurious feature (Salman et al., 2022). In that sense, bias transfer is not confined to explicit fairness annotations; it includes naturally occurring shortcut features such as background or object co-occurrence (Salman et al., 2022).

The CLIP analysis complicates that picture. It distinguishes global from local bias measurements and finds that a model may appear less biased overall while being more biased in semantically coherent local clusters (Ramos et al., 25 Aug 2025). Across 29 CLIP models and downstream VQA and captioning settings, most Spearman correlations between pre-training and downstream bias metrics are weak or statistically insignificant, and the strongest reported correlation is $t$5 with $t$6, which is not conventionally significant (Ramos et al., 25 Aug 2025). Better upstream fairness therefore does not guarantee better downstream fairness in that adaptation paradigm (Ramos et al., 25 Aug 2025).

Related observations appear in audio transfer learning. For instrument recognition with VGGish, OpenL3, and YAMNet, pre-trained audio embeddings differ in cross-dataset generalization and embed dataset-identity and genre-distribution biases. Bias is quantified through the cosine similarity between a dataset-separating LDA direction and an instrument classifier direction, and post-processing debiasing by projecting out dataset- and genre-separating subspaces yields modest improvements in cross-domain ROC-AUC while usually preserving within-domain performance (Wang et al., 2023).

In fair multi-task learning, bias transfer is explicitly tied to gradient conflict. FairBranch groups related tasks by parameter similarity using CKA and corrects fairness-gradient conflicts within branches. On PUMS 18–19 it reports $t$7 and $t$8, and on CelebA gender tasks $t$9 and $\rho \geq 0.94$0, outperforming six state-of-the-art MTL baselines on the joint fairness–accuracy trade-off (Roy et al., 2023). This formulation makes bias transfer an analogue of negative transfer: fairness degradation caused by shared updates.

6. Mitigation and controlled transfer

Mitigation strategies fall into three broad classes: controlling source-model retention, modifying the adaptation objective, and altering the data presented during or after transfer. In transfer learning with CNNs, explicit regularization toward the pre-trained weights replaces standard weight decay with

$\rho \geq 0.94$1

so that the pre-trained model acts not only as initialization but also as a reference point (Li et al., 2018). $\rho \geq 0.94$2-SP and $\rho \geq 0.94$3-SP-Fisher consistently outperform standard $\rho \geq 0.94$4 in target-task accuracy, especially with less target data; $\rho \geq 0.94$5-SP-Fisher is particularly useful when preserving source-task accuracy is important (Li et al., 2018). In hBOA, soft distance-based bias transfer similarly introduces empirical priors from previous runs and yields speedups from about $\rho \geq 0.94$6 to $\rho \geq 0.94$7, and $\rho \geq 0.94$8 when combined with sporadic model building in one reported setting (Pelikan et al., 2012). These are controlled forms of beneficial bias transfer.

For unfair or spurious bias, adaptation-objective modifications include DBA, which applies distance-weighted backpropagation,

$\rho \geq 0.94$9

with $\rho \geq 0.92$0 determined by inverse feature distance, and optional class-frequency weighting under heavy imbalance (Li, 2021). In LLM debiasing, masked language modeling unlearning performs gradient ascent on harmful content and shows cross-domain effects: unlearning only gender hate speech reduces CrowS-Pairs bias from $\rho \geq 0.92$1 to $\rho \geq 0.92$2 for gender, from $\rho \geq 0.92$3 to $\rho \geq 0.92$4 for race, and from $\rho \geq 0.92$5 to $\rho \geq 0.92$6 for religion, while Wikitext-2 perplexity changes only marginally from $\rho \geq 0.92$7 to $\rho \geq 0.92$8 after 50 unlearning steps (Lu et al., 2024).

Data-centric mitigation frequently uses transfer itself as a debiasing instrument. BiaSwap identifies bias-guiding and bias-contrary samples without bias labels, localizes easy-to-learn shortcut regions with CAMs, and creates bias-swapped images; on Colored MNIST with 99% bias ratio it reports $\rho \geq 0.92$9 unbiased test accuracy versus $\rho \geq 0.97$0 for Vanilla and $\rho \geq 0.97$1 for LfF, and on bFFHQ it achieves $\rho \geq 0.97$2 bias-contrary accuracy versus $\rho \geq 0.97$3 for LfF (Kim et al., 2021). BLADE similarly requires no prior knowledge of bias or bias-conflicting samples, translating images across bias domains and adaptively refining representations; it exceeds the closest baseline by around $\rho \geq 0.97$4 on corrupted CIFAR-10 under the worst group setting and by more than $\rho \geq 0.97$5 on Waterbirds worst-group accuracy (Arora et al., 5 Oct 2025).

Style transfer has also been used directly for textual debiasing. A masked-language-modeling style-transfer architecture combines latent content encoding with explicit keyword replacement and, on the Jigsaw dataset, reports content preservation of $\rho \geq 0.97$6, perplexity $\rho \geq 0.97$7, and style-transfer accuracy $\rho \geq 0.97$8 (Tokpo et al., 2022). In Arabic mental-health text, a pretraining-free diffusion model reframes gender-bias mitigation as male-to-female style transfer on the CARMA corpus, which has female:male $\rho \geq 0.97$9; across five dataset constructions it achieves BERTScore $0.94$0–$0.94$1 with ROUGE-1,2,L below $0.94$2, indicating strong semantic preservation with non-trivial stylistic change (Mankarious et al., 20 Jan 2026). In visual classification, transfer learning on stylized images and domain-adversarial training are used to increase shape bias; robustness to stylized test images improves substantially, although no gain in base accuracy is reported (Brochu, 2019).

7. Debates, asymmetries, and open problems

The main controversy is not whether bias transfer exists, but when it should be expected. Some studies report strong persistence across adaptation regimes: transfer learning from biased source models (Salman et al., 2022), zero- and few-shot prompting of causal LLMs (Mackraz et al., 2024), multilingual embedding alignment (Zhao et al., 2020), and cross-lingual sentiment transfer (Goldfarb-Tarrant et al., 2023). Other studies report weak or inconsistent upstream–downstream coupling, especially when downstream adaptation homogenizes representations, as in CLIP backbones connected to the same frozen LLM (Ramos et al., 25 Aug 2025).

A second recurring issue is asymmetry. In human–AI collaboration, transferred bias can manifest not by amplifying privileged-group advantages but by suppressing language associated with marginalized groups, as in the reduced agency of female-target essays under biased LLM assistance (Hossain et al., 14 Jun 2026). In class-imbalanced transfer, the transferred harm may appear primarily in the minor class while aggregate accuracy remains deceptively high (Li, 2021). In multilingual settings, languages with grammatical gender can react differently to alignment and debiasing than less gender-marked languages (Zhao et al., 2020, Reusens et al., 2023).

A third issue is locality. The CLIP study shows that global and local bias measurements can diverge sharply, so a globally fairer model may still be locally worse on particular clusters (Ramos et al., 25 Aug 2025). This aligns with recommendations from several strands of the literature: evaluate per class rather than only by macro-average under DCI (Li, 2021), evaluate both intrinsic and extrinsic bias in multilingual transfer (Zhao et al., 2020), and audit both model outputs and downstream human behavior in AI-assisted settings (Hossain et al., 14 Jun 2026).

Taken together, the literature implies that bias transfer cannot be treated as a single scalar property of a source model. It depends on the bias type, the transfer operator, the measurement protocol, the geometry of the representation space, and the degree to which downstream optimization preserves or overwrites source structure. A plausible implication is that robust bias governance requires auditing at pre-training, alignment, adaptation, and deployment time rather than assuming that any one intervention will remain effective after transfer.