Global public attitudes towards China deteriorated dramatically after a local COVID-19 outbreak in Wuhan escalated into a global pandemic in 2020. Pew Research Centre surveys revealed that views of China worsened most in Anglophone countries such as the United States, the United Kingdom, Canada and Australia.Footnote 1Against this backdrop, #StopAsianHate demonstrations began in the US in 2021, protesting a perceived increase in hate incidents targeting Asians in general and Chinese people in particular.Footnote 2
These events raise a question: what is the nature of attitudinal bias about Chinese people in the Anglophone world today? While attitudes towards “China” primarily reflect opinions about the Chinese government, they can also colour attitudes towards ordinary Chinese people. In a nationally representative online survey experiment, Americans who perceived a greater “China threat” were more likely to view a hypothetical Chinese person as less trustworthy, less moral, less warm and less competent than other Asian Americans.Footnote 3Nationally representative surveys similarly revealed a significant “spillover” effect, whereby heightened negativity felt towards the Chinese government extended to bias against local Chinese communities, especially among more right-wing British and European publics.Footnote 4
These biases are consequential. Americans living in regions more severely affected by COVID and holding conservative (versus liberal) political views were even less likely to trust Chinese people compared to American, Japanese or Indian people.Footnote 5Scientific article authors were less responsive and less willing to share data with male Chinese requestors, possibly owing to their perceived lower trustworthiness.Footnote 6Some Chinese scientists with tenured or tenure-track positions at US universities felt targeted and considered leaving the US and returning to China.Footnote 7
Understanding Biases about Chinese People
Beyond COVID and the Chinese government, what are the more enduring drivers of biases about Chinese people? Biases are subjective but systematic deviations from rationality in social judgement. Social psychologists view bias as universal. As social animals, all humans identify with social groups, engendering biases towards or against other social groups. Biases can be favourable (for example, intelligent), neutral (for example, feminine) or unfavourable (for example, untrustworthy). Analytically, psychologists divide bias into three categories: stereotypes, prejudice and discrimination – that is, biased beliefs, emotions and behavioural tendencies, respectively.Footnote 8
Social psychologists have learned much about the nature of intergroup bias in general and bias regarding Asians broadly, but less about biases regarding Chinese people in particular. Generally, when viewing other groups (outgroups) relative to one’s own group (ingroup), people show ingroup favouritism and outgroup derogation: they hold more positive evaluations of ingroup members and more negative views of outgroup members.Footnote 9
Generalized intergroup bias does not account for the specific content or nature of biases towards particular outgroups. Asians, for instance, are perceived as “model minorities.” Surveys conducted among mostly White/Anglophone respondents find Asians to be stereotyped as less warm/friendly but more competent/capable,Footnote 10or sometimes moderate on both warmth and competence.Footnote 11
How do Chinese people fit into this broad stereotype about all Asians? In the US, East Asians (as opposed to South or South-East Asians) are the largest Asian minority group and are considered prototypical Asian Americans.Footnote 12With distinct subgroup characteristics, such as lower (self-rated and other-perceived) assertiveness than South Asians, Chinese and other East Asians (but not South Asians) are more likely to be excluded from leadership positions in the US.Footnote 13In the UK, however, it is South Asians who are the largest Asian minority group and who are treated as prototypical British Asians. These different contexts complicate the untangling of biases directed at Asians, East Asians and Chinese in the Anglophone world.
While the extant research offers some insights into biases about Chinese people, a better understanding is still needed of both the specific stereotypes about Chinese people and the prejudicial emotions and discriminatory behavioural tendencies that flow from these stereotypes. Additionally, past studies largely examined public attitudes and behaviour towards Chinese via surveys and experiments; little is known about how Chinese people are depicted in more naturalistic daily language use. How Chinese people are written and read about in English both reflects and shapes the biases surrounding them.Footnote 14
Our Research: Using AI Language Models to Uncover Bias in Natural Language
Using Bidirectional Encoder Representations from Transformers (BERT) models pre-trained on massive corpora of English-language books, news articles, Wikipedia, Reddit and Twitter, and a new method for measuring natural-language propositions, this research report aims to uncover the nature of attitudinal biases about Chinese people in English natural language. We distinguish between semantic propositional representations of cognitive (stereotypes), affective (prejudice) and behavioural (discrimination) biases. Rather than devising hypotheses, we seek to gain insights inductively from the texts produced naturally by people.
Theoretical framework
To examine how Chinese people are stereotyped, we used the Stereotype Content Model (SCM) as our theoretical framework. The SCM posits Warmth and Competence as the two fundamental dimensions of stereotype content.Footnote 15The latest SCM work elaborates further by subdividing the Warmth dimension into Sociability/Friendliness and Morality/Trustworthiness facets, and the Competence dimension into Capability/Intelligence and Assertiveness/Dominance facets.Footnote 16We therefore explored all four stereotype facets.
We adopted the Behaviour from Intergroup Affect and Stereotypes (BIAS) map as our theoretical framework to detect semantic representations of emotional feelings (prejudice) and behavioural tendencies (discrimination) towards Chinese people. The BIAS map, as developed from the SCM, posits that distinct patterns of stereotypes on the Warmth and Competence dimensions elicit distinct patterns of intergroup emotions which then shape behavioural tendencies towards different groups (see Figure 1).Footnote 17Specifically, warm and competent (versus cold and incompetent) group stereotypes elicit intergroup emotions of admiration (versus contempt), leading to behavioural tendencies of active and passive facilitation (versus harm), such as helping (versus attacking) and cooperation (versus neglect). By contrast, warm but incompetent (versus cold but competent) stereotypes elicit emotions of pity (versus envy), leading to active (versus passive) facilitation but passive (versus active) harm.Footnote 18In short, the SCM and the BIAS map are comprehensive theoretical frameworks for testing stereotypes, prejudice and discrimination.
Figure 1. Biases about Chinese People in English Language Corpora in the BIAS Map Framework
Methodological framework
To effectively uncover stereotypes, prejudice and discrimination directed at Chinese people in natural language, we employed a new method, the Fill-Mask Association Test (FMAT), which uses BERT models to compute the semantic probabilities of words filling in the masked blank of a designed query (a cloze-like “fill in the blank” sentence).Footnote 19Compared to word frequency and word embedding methods, the FMAT with contextualized BERT models can provide more nuanced insights into social cognition by using phrases and sentences (rather than single words) specifying relational and contextual information. Previous work has established that the FMAT is a reliable and valid tool for measuring attitudes and social cognition at the societal level. It is sensitive to nuanced relational information and is capable of revealing more complex social perceptions, such as intersectional race–gender stereotypes.Footnote 20Thus, the FMAT is well suited for testing semantic propositional representations of intergroup bias towards Chinese people.
Methods
FMAT query design
The FMAT requires a propositional query (i.e. a declarative sentence with a masked word) for a BERT model to understand the context and estimate how likely it is that each word in the model vocabulary might replace the masked word. Following previous research, we used 12 representative BERT models pre-trained on large English-language text corpora, including books, news articles, Wikipedia, Reddit and Twitter (see Table S1 in the online Supplementary Materials).Footnote 21
Table 1 summarizes how we designed FMAT propositional queries to detect linguistic representations of stereotypes, prejudice and discrimination. Each query sentence has an {ATTRIB} placeholder to be substituted, before the fill-mask task, by a word or phrase describing an attribute facet. The attribute words of stereotypes were from an established dictionary, with 12 word pairs for each facet.Footnote 22For intergroup emotions and behaviour, the attribute words/phrases were adopted from the same items as used in the original BIAS map research, with prepositions attached to make the query grammatically correct (Table 1).Footnote 23
Table 1. FMAT Query Design
Notes: To increase the robustness of results, we used four words (people, persons, citizens, individuals) to represent the concept of people, producing four parallel query templates. The [MASK] was left blank for BERT models to estimate the probabilities of the target words (Chinese versus American, British, Canadian, Australian, English, Irish, Scottish, Welsh) appearing in each query sentence.
Given a query sentence with the {ATTRIB} substituted by an attribute phrase, the BERT models were then used to estimate the semantic probability of a target word filling in the [MASK] blank. In this research, Chinese was the main target group. However, it was necessary to contrast Chinese people with a meaningful control group because a single nationality group is semantically confounded with a superordinate concept of nationality. The control group should be relevant to the text producers of the training corpora to control for a baseline of social bias embedded in texts. Our supplementary analysis showed that the vast majority of the English-language text producers for the sampled BERT models were from eight Anglophone nations/states, so we used them as the baseline control group: American, British, Canadian, Australian, English, Irish, Scottish and Welsh people (see both Supplementary Analysis #1 and #2 in the online Supplementary Materials, where we also tested bias with each Anglophone nation/state separately as the control group).
Analytic strategy
Following the FMAT analytic approach, we computed the log probability ratio (LPR) of a target word between a pair of attributes to measure their relative association. To balance the data between the Chinese and control conditions, we averaged LPRs across all of the eight Anglophone target words to produce one LPR value for the control group. The raw LPRs were then standardized with the population standard deviation (σ = 1.414), so that the results of target pairwise contrasts can be interpreted as an effect size equivalent to Cohen’s d.Footnote 24Finally, the data presented 4,608 observations of standardized LPRs for stereotypes, 1,536 for prejudice and 1,536 for discrimination, across all combinations of factors.
Data were processed and analysed using R and the R package FMAT.Footnote 25To address the nested data structure with LPRs (Level 1) nested within BERT models (Level 2), we fitted a Linear Mixed Model respectively for stereotypes, prejudice and discrimination, with LPR as the outcome variable and the two [MASK] target conditions (Chinese versus the baseline control), attribute facets, query templates and all their interactions as the predictors.
Results
Reliability analysis
Before the main analysis, we assessed how reliable the FMAT method was by computing: (1) an average-score intraclass correlation coefficient (ICCaverage) to assess the interrater agreement among the 12 BERT models in understanding query sentences and estimating semantic probabilities; and (2) Cronbach’s α (αquery) to assess the internal consistency of LPRs between query templates.Footnote 26
For stereotypes, prejudice and discrimination, respectively, the FMAT showed good interrater agreement among BERT models (ICCaverage = .95, .96, .97) and internal consistency between query templates (αquery = .78, .89, .90). Accordingly, the BERT models and FMAT queries were no longer distinguished when interpreting the main results.
Stereotypes (cognitive beliefs)
The main effect of the [MASK] target was qualified by a significant target × stereotype facet interaction.Footnote 27Compared to the baseline Anglophone control group, Chinese people were stereotyped as less sociable/friendly,Footnote 28less moral/trustworthyFootnote 29and less assertive/dominant,Footnote 30but equally capable/intelligentFootnote 31(see Table 2 and Figure 2A).
Table 2. FMAT Effect Sizes of Biases about Chinese People
*** Notes: Effect sizes are standardized estimates (Cohen’s d) from the linear mixed model of each bias component: 0.2 is considered small, 0.5 medium, and 0.8 large. Standard errors (SE) are shown in parentheses; 95% confidence intervals (CIs) are shown in brackets. * p < .05. ** p < .01. *** p < .001.
Figure 2. FMAT Effect Sizes
Overall, the FMAT revealed low-warmth and low-competence stereotypes of Chinese people in English language texts (Figure 1). The Anglophone text producers perceived Chinese people as less trustworthy, less friendly and less assertive than themselves, but equally as intelligent. Notably, while earlier research found stereotypes of Asians in the US as cold but competent,Footnote 32or sometimes medium on both dimensions,Footnote 33our results are consistent with more recent findings that East (versus South) Asians are perceived to be less assertiveFootnote 34and that Chinese are perceived to be less trustworthy.Footnote 35
Prejudice (emotional feelings)
The main effect of the target group on prejudice was again qualified by a significant target × emotion facet interaction.Footnote 36Since the BIAS map framework placed admiration and contempt on one dimension and envy and pity on the other,Footnote 37we did not test each single emotion but performed planned contrasts for (1) admiration (versus contempt) and (2) envy (versus pity) towards Chinese people (versus control). In doing so, we could determine the relative strength between each pair of intergroup emotions. The results suggest that the Anglophone text producers expressed less admiration and more contempt towards Chinese,Footnote 38but their levels of envy and pity were indistinguishableFootnote 39(see Table 2 and Figure 2B).
In combination with the findings of stereotypes, this pattern of prejudice is consistent with the theoretical prediction of the BIAS map that an emotion of contempt (versus admiration) should follow from the low-warmth/low-competence (versus high-warmth/high-competence) stereotype.
Discrimination (behavioural tendencies)
As with prejudice, the main effect of the target group on discrimination was again divergent between the four behaviour facets.Footnote 40We therefore performed planned contrasts for (1) active facilitation (versus harm) and (2) passive facilitation (versus harm) towards Chinese people (versus control). The results suggest that the Anglophone natural-language text producers were less likely to help/protect and more likely to attack/fight Chinese people,Footnote 41and less likely to cooperate/associate with Chinese people and more likely to exclude/demean themFootnote 42(see Table 2 and Figure 2C).
The FMAT indicated that Chinese people were more likely to be actively and passively harmed (versus facilitated), again supporting the BIAS map prediction. Specifically, originating from the low-warmth/low-competence stereotype, “disliked” (i.e. higher on contempt than admired) Chinese people can elicit a behavioural orientation of either active or passive harm – that is, a bias towards either attacking or excluding Chinese people, rather than helping or cooperating with them.
Robustness checks and supplementary analyses
In addition to these main analyses, we conducted two supplementary analyses to check the robustness of our main findings and to explore any nuances within them. First, to provide nuances in bias directed at Chinese people for different control groups, we tested this bias with each of the eight Anglophone nations/states separately as the control group. Across all eight Anglophone ingroups, we found robust support for the pattern that Chinese people were semantically associated with less moral/trustworthy and less assertive/dominant stereotypes, contempt (versus admiration) prejudice, and both active and passive harm (versus facilitation) discrimination. We also, however, identified nuances that Chinese people were stereotyped as less sociable/friendly than English/Scottish/Welsh people but more capable/intelligent than American/British/Australian/Irish people (see Table S2 from Supplementary Analysis #2 in the online Supplementary Materials).
Second, to explore whether our results were unique to Chinese people or common to all East Asians, we conducted similar tests on biases directed at Japanese and Korean people with all or each of the eight Anglophone nations/states as the baseline control group. While biases directed at Japanese people showed a similar pattern to those directed at Chinese, biases directed at Korean people were distinct: weaker and more positive (see Tables S3 and S4 from Supplementary Analysis #3 in the online Supplementary Materials).
In short, these supplemental analyses revealed that our main findings on cognitive, affective and behavioural biases about Chinese people cannot be reduced to simple outgroup derogation or a common bias about all East Asians.
Discussion
Using the new FMAT method with BERT models pre-trained on large English-language text corpora, we investigated natural-language propositions that suggest how English speakers think, feel and are inclined to behave towards Chinese people. Overall, the FMAT uncovered relative semantic associations of Chinese people (versus Anglophone control groups) with (1) low-warmth (less trustworthy and less friendly) and somewhat low-competence (less assertive but equally intelligent) stereotypes, (2) contempt (versus admiration) prejudice, and (3) both active and passive harm (versus facilitation) discrimination. This pattern locates Chinese people in the bottom-left quadrant of the BIAS map framework (see Figure 1), consistent with the theoretical prediction that groups in different quadrants of stereotypes should be associated with one intergroup emotion and two behavioural tendencies.Footnote 43
Implications
These findings have implications for theory, methodology and interventions to reduce prejudice and discrimination against Chinese people. Theoretically, our findings support the SCM and the BIAS map as comprehensive frameworks for gaining a better understanding of stereotypes, prejudice and discrimination directed at Chinese people, who constitute one-fifth of the world’s population. Recent studies have documented varieties of biases against Chinese: dishonest,Footnote 44distrustFootnote 45and exclusion.Footnote 46Adding novel natural-language evidence to this literature, we show that bias against Chinese people is semantically represented in English-language books, newspapers, Wikipedia and Twitter. Associating Chinese people with lower trustworthiness and lower assertiveness, Anglophone publics tend to dislike (versus admire), attack (versus help) and exclude (versus cooperate with) Chinese people. These semantic representations of anti-Chinese bias can reveal non-intentional bias in the real world that interviews, surveys and laboratory experiments may not be able to capture. Together, our findings integrate past observations into the BIAS map theoretical framework, providing a more systematic understanding of attitudinal biases about Chinese people in real-life English language use.
Methodologically, the validity of FMAT in identifying cognitive, affective and behavioural biases in natural language highlights its potential for studying complex intergroup processes. Large language models (LLMs) are increasingly used as proxies for human participants, simulating human populations at the group level.Footnote 47The FMAT, as a new methodological framework, provides a more fine-grained group-level measurement of psychological, social and cultural constructs, especially those that cannot be well captured by single words.Footnote 48Expanding the scope of FMAT applications, this research report demonstrates its flexibility and validity to measure emotional feelings and behavioural tendencies in natural language involving perceivers and targets, with their relational information specified by verb phrases. This method, therefore, can advance research on other complex intergroup relations in natural contexts outside the laboratory.
Additionally, our findings can suggest interventions to reduce bias against Chinese people. The #StopAsianHate protests opposed hateful behaviour against Asians, and Chinese people in particular. Biased intergroup behaviour (discrimination) derives from intergroup emotions (prejudice) that originate from cognitive beliefs about social groups (stereotypes).Footnote 49Therefore, to change emotional feelings and behavioural orientations towards Chinese people, we need to combat the most detrimental stereotypical beliefs about Chinese people: that they are immoral, untrustworthy and dishonest.
A recent controversy over Chinese dis/honesty is instructive. Although stereotypes are mostly social constructions, they may sometimes have some factual basis. So, when a group of scholars used one measure (emailing the wallet owner to return a lost wallet) in a large cross-national behavioural study and concluded that Chinese people were the least honest people in the world, this empirical finding could reinforce stereotypes of Chinese as exceptionally untrustworthy.Footnote 50Another team of scholars then used a more culturally sensitive measure of dis/honesty for collectivistic cultures (safekeeping the wallet), and found the wallet recovery rates of Chinese people to be little different from the email response rates of other national groups.Footnote 51This finding undermined the factual basis for the “untrustworthy” stereotype given to Chinese people. To reduce prejudice and discrimination, we need more such attempts to counter detrimental stereotypes about Chinese people.
In addition, given our supplementary results showing diverse patterns of bias towards different East Asian subgroups, it is crucial to consider the nuances among the Chinese, Japanese and Korean targets when designing interventions to combat intergroup bias. In one study, East Asians (versus South Asians) were perceived as more represented by the #StopAsianHate movement and more credible when described as victims of a hate crime; this link was mediated by perceived prototypicality.Footnote 52However, biases about East Asians are complex and mixed with both “model minority” and “Yellow Peril” racist stereotypes.Footnote 53Our results from Supplementary Analysis #3 highlight both similarities and differences regarding bias towards East Asian subgroups. Thus, rather than using “Asian” as a broad label, policymakers should address the nuanced situations faced by Chinese, Japanese and Korean targets of bias.Footnote 54
Future directions
While our research contributes to an understanding of biases about Chinese people in the real world, several open questions warrant future investigation. First, more research is needed about the two facets of the competence stereotype. On the one hand, Chinese people were depicted as just as capable/intelligent as ingroup Anglophones, which is consistent with the “model minority” stereotype. On the other hand, Chinese people were seen as less assertive/dominant, perhaps explaining why Chinese men were seen as easy targets of verbal and physical aggression during COVID-19, and why Chinese people are underrepresented in leadership positions in the US.Footnote 55Thus, both favourable and unfavourable biases underlie stereotypes about the competence of Chinese people. Might the greater salience of one or the other facet of competence shape the nature of any subsequent prejudice and discrimination?
Second, our use of the FMAT to measure stereotypes, prejudice and discrimination in natural language could be dynamic rather than static. Our findings are based on BERT models pre-trained on pre-COVID text corpora. Future studies could use similar methods to track how biases about Chinese people have changed over the past centuries, or use BERT models trained on pre-COVID and post-COVID text corpora to test whether the nature of bias has shifted due to the pandemic.
Third, our study focuses on biases about Chinese people, with results that integrate previous findings specific to them. Future studies could use the same English-language corpora to further explore biases about other Asian subgroups. For instance, what are the similarities and differences in Anglophone biases about East, South-East or South Asian people?
Fourth, future research could switch the target of bias from Chinese people to China and its government. Such studies would need to draw from international relations (IR) theories, which are more appropriate than social psychological theories for understanding Anglophone views of foreign countries and governments. For instance, building on Immanuel Kant’s work on the “democratic peace,” liberal IR theories today suggest that democratic publics in the Anglophone world should be more biased against China’s authoritarian government than against Japan’s democratic one.Footnote 56
Fifth and finally, similar methods can be used to explore Chinese biases. By utilizing and analysing large Chinese-language corpora, what can we learn about the stereotypes, prejudice and discrimination that Chinese people may hold towards Americans, Europeans, Africans and other national/cultural groups around the world?
Concluding Remarks
Bias – the subjective but systematic deviation from rationality in social judgement – is universal but takes on specific contents for specific social groups. What is the nature of biases about Chinese in particular? Using a new methodological approach with the FMAT method and pre-trained BERT models, we explored bias regarding Chinese people in English language use. Anglophone text producers tended to: (1) perceive Chinese people to be less trustworthy, less friendly and less assertive, while holding them to be as equally intelligent (stereotypes); (2) scorn (versus admire) Chinese people yet equally envy and pity them (prejudice); and (3) either attack (versus help) or exclude (versus cooperate with) Chinese people (discrimination). Consistent with the BIAS map’s theoretical prediction, these findings suggest how multiple facets of anti-Chinese bias are semantically represented in real-life English language texts, providing a social psychological framework for practitioners and policymakers to reduce anti-Chinese bias and hate crimes in the real world.
Supplementary material
Supplementary materials, including all data and analysis code, are available at the Open Science Framework (OSF) https://osf.io/9wgkb/ and at https://doi.org/10.1017/S0305741025100532
Acknowledgements
The authors thank Tao Wang and Xiaobing Wang for their feedback and comments on preliminary findings. They also wish to thank Kai Hung Lee, whose generous gift to the University of Manchester endowed the Manchester China Institute, enabling this research. The first author was also supported by the Shanghai Pujiang Programme (grant number: 24PJC025).
Competing interests
None.
Han-Wu-Shuang BAO is an assistant professor at the East China Normal University. His research uses natural language processing and computational intelligent methods (for example, the FMAT) to understand social cognition and historical psychology.
Peter GRIES is professor of Chinese politics and the Lee Kai Hung chair and director of the Manchester China Institute at the University of Manchester. He studies the political psychology of international affairs, with a focus on China and the United States.
Open access
