Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.
Advertisement
npj Artificial Intelligence has APC waivers available that can be allocated upon acceptance on an ad-hoc basis. For additional information, contact the Journal Publisher, Ronghua Guo.
npj Artificial Intelligence volume 1, Article number: 16 (2025)
2420
2
Metrics details
This study investigates how stereotypes in online medical crowdfunding campaigns (OMCCs) influence donation intentions. We used both manual and LLM-based methods to extract visual, textual, and campaign features. Two crowdsourcing tasks were conducted: one involving human ratings from 150 participants and the other utilizing an LLM-simulated process. Results showed that campaigns with cover images conveying warmth and competence performed better. Both human and LLM judgments highlighted visual features as critical in predicting these perceptions, with key predictive features including hospital setting, contrasts between vitality and illness, number of images, and facial expressions, particularly smiles. Mediation analysis revealed that smiles indirectly influenced donations via warmth and competence. Images of hospital settings increased donations, while images of healthy subjects had a negative effect. This study highlights the impact of visuals in OMCCs and provides insight into how stereotypes can shape donation behavior. In addition, using LLMs to simulate crowdsourcing tasks overcomes the scale problem and enables efficient analysis of large datasets.
As one of major global health challenges, cancer has a significant impact on both mortality and morbidity. Approximately 22 million people worldwide are expected to receive a cancer diagnosis in 20301. Similar to other chronic diseases, individuals with cancer require prolonged medical care, which places a substantial financial burden on healthcare systems, patients, and their families2,3. In light of such circumstances, online medical crowdfunding has emerged as a supplemental resource when health insurance or personal savings are insufficient to cover healthcare costs. Online medical crowdfunding is a mechanism that relies on donations where campaigners share their requests online to attract public sympathy and secure financial support. Its primary goal is to obtain funds to cover essential medical expenses4. Studies have shown that medical expenses account for roughly one-third of all crowdfunding activities5. For example, GoFundMe hosts approximately 250,000 medical crowdfunding campaigns annually in the United States alone, raising nearly $650 million each year. However, medical crowdfunding projects have a low success rate, with many falling short of their funding goals and trailing behind other categories such as educational scholarships, disaster relief, and nature conservation6,7. This trend is further highlighted by a study which found that only 10% of a sample of 200 randomly selected GoFundMe medical campaigns successfully met their fundraising objectives8.
Stereotypes represent cognitive structures that encompass individuals’ knowledge, beliefs, and expectations about specific social groups9. As a distinct type of social cognitive schema, stereotypes are rigid notions about the characteristics and underlying motivations of certain group members10. A wide range of dimensions have been proposed to elucidate the content structure of stereotypes11, with Fiske et al.’s Stereotype Content Model (SCM) being particularly influential. The SCM evaluates stereotypes along two dimensions: competence and warmth12. Warmth perceptions relate to judgments about the intentions of others, such as their apparent kindness, friendliness, and goodwill13. Competence, on the other hand, is associated with the capacity to carry out one’s intentions, including traits like ability, efficiency, and power14. Building on this, Cuddy and his colleagues developed a BIAS Map, which enhances the practical application of the SCM15. This model posits that groups perceived as warm elicit proactive facilitation (helping behaviors), whereas those perceived as low in warmth provoke proactive harm (aggressive behaviors). Similarly, Cuddy also mentioned that groups perceived as competent lead to passive facilitation (cooperative behavior), while those perceived as incompetent result in passive harm (neglect and exclusionary behavior). Stereotypes have a significant impact on online crowdfunding campaigns, shaping the decisions of potential donors. The perception of warmth and competence is affected by various factors, including the perceiver’s characteristics, the target’s characteristics, their relationship, and the external setting16. These factors can directly impact the outcomes of crowdfunding campaigns. Text-based factors such as language style, social capital, and aversion to overhead cost, affect donor engagement17,18. Effective storytelling and the inclusion of factual details have been shown to increase donations15, while excessive emotional appeals may have opposite effects19. In medical crowdfunding, positive emotional descriptions, detailed textual content, ambitious fundraising goals, and the use of a third-person perspective in project descriptions have been found to improve outcomes20. Visual elements, such as images and videos, also play a critical role in attracting attention and enhancing credibility19,21,22,23. Images convey key attributes, including gender, age, and emotion, which shape donor perceptions24,25,26. Studies on online platforms further highlight the persuasive power of visual content in social interactions27,28,29,30.
Crowdsourcing has become an important method for obtaining research data in social science research and is widely used in data annotation tasks. Numerous online crowdsourcing platforms support social science research, including Amazon MTurk (https://www.mturk.com/), Prolific (https://www.prolific.com/), and CloudResearch (https://www.cloudresearch.com/). However, this approach is typically time-consuming, costly (For example, one study calculated that conducting a typical online experiment with 300 participants and a 5-minute duration on MTurk takes approximately 12 hrs and costs $180 (300 participants × $0.50 per participant × [1 + 0.2]) (Mturk’ s service fee rate is 20%). Moreover, for data annotation tasks in specific domains, crowdsourcing workers often require specialized domain knowledge to ensure annotation accuracy.), and labor intensive31. Large Language Models (LLMs), with their powerful capabilities in various NLP tasks, have also been widely applied to data annotation tasks32,33. Additionally, LLM-based data annotation is not only cost-effective and time-efficient, but also has been proven to be as good as or even better than crowdsourced annotations34. Given their extensive use in scientific research35, large language models (LLMs) are increasingly being applied to online crowdfunding research due to their robust text understanding and generation capabilities36,37,38. Existing studies suggest that ChatGPT’s text annotation capabilities exceed those of humans39. Moreover, the integration of textual and visual features can significantly enhance the accuracy of predicting crowdfunding campaign success40. Thus, we aim to answer the following research questions:
RQ1: To what extent do potential donors exhibit consistent stereotypes (warmth and competence) and donation intentions when evaluating online medical crowdfunding campaigns? How do these stereotypes correlate with perceived donation intentions?
RQ2: What relationships exist between online information features and various stereotypes? How do these associations differ between human evaluations and LLM-simulated judgments?
RQ3: Can a computational model trained on multimodal features (e.g., visual, textual, personal attributes) accurately predict high/low warmth and competence stereotypes? How does model performance compare when using human-rated vs. LLM-simulated ratings?
RQ4: Which features contribute most significantly to stereotype prediction in OMCCs? Through what mechanisms do specific features mediate the relationship between stereotypes and donation behavior?
To explore these issues, we utilized 250 online medical crowdfunding projects from GoFundMe, each featuring personal images and associated descriptive texts. First, we conducted a survey-based evaluation with human participants to assess their perceptions and donation intentions. Participants provided perception ratings, focusing on two stereotype dimensions, warmth and competence as well as their associated donation intentions toward the campaigns. To address the scale problem, we supplemented this analysis with LLMs to simulate human judgments. This study aims to examine the relationship between online information features and the different stereotypes formed by humans and large language models (LLMs) simulating human identities. Furthermore, we sought to develop a computational model capable of predicting stereotypes formed by humans and LLMs from OMCCs information. By identifying the features that most significantly contribute to stereotype prediction in OMCCs images, we examined how these features affect donation intention. Through this analysis, we aim to provide insights into the role of stereotypes in crowdfunding and contribute to the existing literature in this domain.
This study makes significant theoretical and practical contributions. Theoretically, it introduces an innovative framework that combines human crowdsourcing with large language model (LLM) support and integrates multimodal data, including text and images, to investigate the factors that influence donation behavior in online medical crowdfunding. It extends the research boundaries of research on the role of stereotypes in this field. Practically, it provides actionable strategic guidance for crowdfunding initiators, enabling them to optimize project presentations by leveraging accurate stereotype formation to enhance fundraising effectiveness. Additionally, it provides critical evidence to support the standardization of operations and the strengthening of ethical frameworks for crowdfunding platforms and nonprofit organizations.
The respondents in our experiment were potential donors who were potential recipients of online medical crowdfunding content. Before initiating the analysis, we conducted a validation process to determine whether prospective donors could form consistent stereotypes and donation intentions based solely on the information shared by the campaign initiators. During the crowdsourcing procedures, each item was paired with multiple questions to evaluate consistency. We used Cronbach’s alpha as the metric to measure the internal consistency of ratings for each item. The Cronbach’s alpha values for all aspects exceeded 0.8 (as shown in Table 1), indicating high reliability in the ratings of perceived stereotypes and donation intentions. We subsequently calculated the average rating for each task item and the standard deviation of the five scores per image. Overall, potential donors generally demonstrated consistent evaluations of stereotypes when assessing the information provided by OMCCs.
From the scoring results of LLM-simulated crowdsourcing (as shown in Table 2), only the Cronbach’s alpha coefficient for the donation intention dimension exceeded 0.8, while the coefficients for the warmth and competence dimensions in stereotype impressions did not surpass 0.7. It indicates that the LLM scoring performance was inferior to that of human participants.
To investigate the relationship between stereotypes and donation intention, we conducted linear regression analyses based on the scoring results of human participants (as shown in Table 3). Models 1 through 3 show the relationship between each separate dimension of impression and donation intention. Our findings demonstrate a positive correlation between warmth and the likelihood of donating (p < 0.001). Also, we found competence to be positively associated with the intent to donate (p < 0.001). Based on the findings presented in Table 4, the scoring results derived from LLM-simulated crowdsourcing also exhibit a similar trend.
To better elucidate the differences in scoring between human participants and LLM across the dimensions of warmth, competence, and donation intention, we employed visual comparisons of their respective score distributions in these three aspects (Fig. 1). Additionally, paired-sample t tests were conducted to statistically analyze these differences.
Visualization of Human-LLMs Scoring Comparison.
As shown in the correlation results presented in Table 5, significant positive correlations were observed between human participants and LLM-simulated crowdsourcing results in the dimensions of warmth (p < 0.001) and competence (p < 0.01). However, no correlation was found in donation intention. Furthermore, paired-sample t test results show differences between human participants and LLM-simulated crowdsourcing results across all three dimensions: warmth (p < 0.001), competence (p < 0.001), and donation intention (p < 0.001).
To explore features associated with perceived warmth and perceived competence, we utilized the scoring data from human participants and conducted Spearman’s rank correlation analysis for each feature against the two stereotype dimensions. This method measures the strength and direction of association between two ranked variables. Figure 2 shows the Spearman’s rank correlation between distinct features and the stereotypes, with each row representing a separate feature. Table 6 shows the quantity and proportion of features in each set that have a significant correlation with either type of stereotype dimension (p < 0.05). Age is represented by four dummy variables. Notably, over 50% of the visual features demonstrate a significant correlation with stereotypes.
The top heatmap illustrates the correlation between stereotypes and visual features, the middle heatmap shows the correlation among stereotypes, textual features, and personal attributes, while the bottom heatmap depicts the correlation among stereotypes. *p < 0.05. **p < 0.01. ***p < 0.001.
The association between visual features and stereotypes is elucidated detailed in Fig. 2. For the warmth dimension, a notable positive correlation (r = 0.35) was observed between warmth and the presence of a smile, indicating that the images portraying a smiling person are positively correlated with the perception of warmth. Similarly, a significant positive correlation was found between smiling and competence (p < 0.001), suggesting that beneficiaries who upload smiling images on online medical crowdfunding platforms are more likely to convey a sense of competence to donors. The GoFundMe platform not only supports the upload of cover images but also allows fundraisers to upload additional images on the details page. In a pertinent exploration, the number of images uploaded by fundraisers was negatively correlated with the perception of competence.
For example, in Figs. 3 and 4 illustrate two scenarios based on whether the beneficiary is smiling or not. In Fig. 3, the beneficiary smiles at the camera, conveying high warmth and competence, resulting in elevated scores for both dimensions. Conversely, in Fig. 4, the beneficiary does not smile and is shown lying in bed with an oxygen tube, which conveys low warmth and competence. This demonstrates how the absence of a smile can negatively impact perceptions of both dimensions of warmth and competence.
To protect the privacy of individuals, all identifiable information, including faces, names, and specific details, has been masked in the image.
To protect the privacy of individuals, all identifiable information, including faces, names, and specific details, has been masked in the image.
The correlation between textual features and stereotypes is also illustrated in Fig. 2. The findings indicate a positive correlation (p < 0.05) between the length and complexity of crowdfunding project descriptions and perceived warmth. This suggests that longer and more complex textual descriptions are more effective in expressing the enthusiasm of beneficiaries. Similarly, complex textual descriptions are positively correlated (p < 0.05) with the competence of the recipients, as beneficiaries with higher competence tend to produce more detailed and sophisticated descriptions. An interesting discovery pertains to perspective from which medical crowdfunding information is posted. When the information is posted by the beneficiary themselves, it exhibits a negative correlation (p < 0.01) with both warmth and competence. While the information is posted by friends, it shows a positive correlation (p < 0.001) with competence.
In summary, the correlation coefficient heatmaps provided in this study shed light on the relationships between visual features, textual elements, and stereotypes. However, these correlations should not be misinterpreted as indicators of causality. Further study is necessary to examine the key characteristics that shape the perception of warmth and competence. In the subsequent study, we explored how these critical variables influence warmth and competence, striving for a deeper understanding of their impact.
For example, Figs. 5 and 6 present two contrasting scenarios. In Fig. 5, it is visually evident that the appeal was written by the beneficiary’s friend or sister, which indirectly conveys the impression that the beneficiary has strong interpersonal relationships. This, in turn, conveys a sense of high warmth and competence among the donors, resulting in higher scores on both the warmth and competence dimensions. In contrast, Fig. 6 shows that the appeal is initiated by the beneficiary themselves. Unlike in Fig. 5 the beneficiary is unable to directly emphasize the importance of having strong interpersonal relationships, leading to lower scores on both warmth and competence dimensions for the project.
To protect the privacy of individuals, all identifiable information, including faces, names, and specific details, has been masked in the image.
To protect the privacy of individuals, all identifiable information, including faces, names, and specific details, has been masked in the image.
Employing the same methodology, we analyzed and compared the simulated crowdsourced scoring data generated by the LLM. Figure 7 illustrates the Spearman rank correlation between each feature and stereotype dimensions, with each row representing an independent feature. Table 7 presents the quantity and proportion of features within each group that exhibit significant correlations with either type of stereotype (p < 0.05). A comparison of Table 6 and Table 7 reveals that the scores derived from the LLM-simulated crowdsourcing process demonstrate a greater number of significant variables when exploring the correlations between visual features, textual features, personal attributes, perceived warmth, perceived competence, and donation intention. Specifically, over 40% of the visual features exhibit a significant correlation with stereotypes, while more than 40% of the textual features show a significant correlation with perceived warmth. Additionally, age and gender differences also exhibit significant correlations with stereotypes. These findings surpass those presented by human participants’ ratings.
The top heatmap illustrates the correlation between stereotypes and visual features, the middle heatmap shows the correlation among stereotypes, textual features, and personal attributes, while the bottom heatmap depicts the correlation among stereotypes. *p < 0.05. **p < 0.01. ***p < 0.001.
Consistent with ratings from human participants, the presence of a smile on the beneficiary in the cover image is significantly positively correlated with perceived warmth (p < 0.001) and perceived competence (p < 0.01). However, notable differences emerge when considering the impact of Beauty Score and the shooting location depicted in the cover image as evaluated by LLM-simulated crowdsourcing. Specifically, when the beneficiary in the cover image possesses a higher Beauty Score (p < 0.05), LLM-simulated crowdsourcing tends to assign higher scores for perceived warmth and perceived competence. Conversely, when the shooting scene depicted in the cover image is within a hospital setting, LLM-simulated crowdsourcing assigns lower scores for perceived warmth (p < 0.05) but higher scores for perceived competence (p < 0.001).
Consistent with ratings from human participants, the length (p < 0.001) and complexity (p < 0.001) of crowdfunding project descriptions are positively correlated with perceived warmth. However, when the crowdfunding information is self-posted, it exhibits negative correlations with warmth (p < 0.001) and competence (p < 0.01). Conversely, when the information is posted by friends of the beneficiary, it correlates positively with competence (p < 0.05). Moreover, the simulated crowdsourcing scores from LLM also demonstrate significant correlations with other variables. For instance, the proportion of health-related terms in the project description text exhibits significant negative correlations with perceived warmth (p < 0.05) and perceived competence (p < 0.05), while the proportion of emotion-related terms correlates positively with perceived warmth (p < 0.05). Furthermore, LLM-simulated crowdsourcing results reveal that the cancer stage of the beneficiary is significantly negatively correlated with perceived warmth (p < 0.05) and perceived competence (p < 0.05). Additionally, when the crowdfunding information is posted by friends of the beneficiary, it positively correlates with perceived warmth (p < 0.01).
In the analysis of features related to the personal attributes of the beneficiaries, human participants’ ratings did not show any significant correlations with stereotypes. However, the scores generated by LLM-simulated crowdsourcing demonstrated significant correlations with gender and age. Regarding age, minors (age<18) showed significant positive correlations with perceived warmth (p < 0.05) and perceived competence (p < 0.05). Young adults (age 18–35) only exhibited a significant positive correlation with perceived warmth (p < 0.05), while middle-aged adults (age 36–60) showed significant negative correlations with perceived warmth (p < 0.01) and perceived competence (p < 0.01).In terms of gender, when the beneficiary was male, significant negative correlations were found with perceived warmth (p < 0.001) and perceived competence (p < 0.05).
This section evaluates the potential of predictive computational models to identify potential donors’ stereotypes in OMCCs by analyzing images and text. In this section, the study also categorized and trained classifiers based on the source of ratings. The specific computational model selection and its parameters are detailed in Table 8.
The stereotype prediction results based on crowdsourced ratings from human participants are presented in Table 9 and Fig. 8. Table 9 summarizes the performance of the four classifiers in distinguishing “high/low competence” and “high/low warmth” categories using various evaluation metrics. During the entire training process, the average and maximum model performance for each classifier were recorded using 10-fold cross-validation. Overall, the imbalance in the dataset has led to lower performance of the classification models, resulting in relatively low average metrics across most models. This is a common challenge encountered in binary classification tasks. In terms of average performance, the SVC model with an RBF kernel and the decision tree model exhibited relatively lower performance. However, SVC models with a linear kernel consistently demonstrated notably superior average performance. When examining the best-performing models, as measured by F1-score and AUC metrics, the SVC model with a linear kernel and the random forest model performed exceptionally well in predicting levels of warmth and competence. For the classification of “high/low competence” and “high/low warmth,” the SVC model with a linear kernel achieved F1-score values of 0.800 and 0.759 respectively, and AUC values of 0.682 and 0.713, respectively. The random forest model achieved F1-score values of 0.750 and 0.813 respectively, and AUC values of 0.721 and 0.827, respectively. Considering both average and best performance, the SVC model with a linear kernel was selected as the machine learning model for the prediction task based on crowdsourced ratings from human and participants.
The “SVC(linear)-warmth” and “SVC(linear)-competence” curves illustrate the ROC values of the linear kernel Support Vector Classifier (SVC) for predictions of “high/low warmth” and “high/low competence”. The “SVC(rbf)-warmth” and “SVC(rbf)-competence” curves show the ROC values of the RBF kernel SVC for the same classifications. The “RF-warmth” and “RF-competence” curves demonstrate the Random Forest ROC values for these classifications. The “DT-warmth” and “DT-competence” curves exhibit the Decision Tree ROC values for the same classifications.
Table 10 presents the performance of four classifiers based on simulated crowdsourcing scores from LLM in distinguishing between “high/low competence” and “high/low warmth” categories using various evaluation metrics. In terms of average performance, the SVC models with RBF kernel and Decision Tree models exhibit relatively lower performance. However, the SVC model with a linear kernel demonstrates superior average performance. When examining the models performing at the top, considering F1-score and AUC metrics, the SVC models with RBF kernel perform well in terms of AUC, but exhibit poorer performance in F1-score, especially in predicting perceived competence. For the classification of “high/low competence” and “high/low warmth”, the SVC models with a linear kernel achieve F1-score values of 0.700 and 0.727, respectively, along with AUC values of 0.800 and 0.769 respectively. Considering both average and optimal performance, a SVC model with a linear kernel is selected as the machine learning model for stereotypes prediction based on simulated crowdsourcing scores from LLM.
By combining the information obtained from Fig. 8 and Fig. 9, and comprehensively analyzing the ROC curves, AUC values, and F1-score values, the SVC model with a linear kernel displays slightly higher classification performance. The red curves show how well the linear kernel SVC model performs in classifying “high/low competence” and “high/low warmth”. Both curves exhibit a higher area under the curve (AUC), indicating superior performance. In this study of 29 features, the SVC model demonstrated its robustness, particularly in dealing with datasets containing a significant number of features.
The “SVC(linear)-warmth” and “SVC(linear)-competence” curves illustrate the ROC values of the linear kernel Support Vector Classifier (SVC) for predictions of “high/low warmth” and “high/low competence”. The “SVC(rbf)-warmth” and “SVC(rbf)-competence” curves show the ROC values of the RBF kernel SVC for the same classifications. The “RF-warmth” and “RF-competence” curves demonstrate the Random Forest ROC values for these classifications. The “DT-warmth” and “DT-competence” curves exhibit the Decision Tree ROC values for the same classifications.
To understand the relative importance of individual features in distinguishing high/low warmth and high/low competence, we conducted feature-importance analyses based on the feature coefficients. Feature coefficients represent the weights that measure the influence of each feature on the classification results in the SVC model. In this study, we used a linear kernel function to train the SVC model. As a result, the feature coefficients indicate the contribution of each feature to the decision boundary. The larger the absolute value of the feature coefficient, the greater its contribution to the classification results. This means that features with the higher coefficients have a stronger impact on the classification outcome and therefore it is more important for predicting the results. By examining the magnitude of the feature coefficients, we can identify the features with the highest predictive power in distinguishing high/low warmth and high/low competence.
Table 11 lists the top ten important features predicted for each stereotype based on crowdsourced ratings from human participants. In terms of competence, among all features, whether the beneficiary is in hospital has the highest predictive contribution, indicating that the beneficiary’s location (whether he/she is in the hospital) has a significant impact on the perception of competence. Other crucial factors in predicting competence include Vitality versus Illness, indicating whether the images uploaded by beneficiaries portray themes of health. Additional factors considered are the number of images uploaded by the beneficiaries and whether they exhibit a smile. For warmth, the smile feature has the highest predictive contribution among all features, suggesting that the presence or absence of a smile has a significant influence on the perception of warmth. Additionally, features such as Is in the Hospital, Vitality versus Illness and Interaction with Surroundings also play important roles in predicting high/low warmth. In the age dimension, the significance of age in predicting high/low warmth and high/low competence is notable only when the group is categorized as Younger Adults (ages 18–35) or Older Adults (age > = 60). Regarding the dimension of the perspective from which the information is posted, both “Posted by Self” and “Posted by Friend” play crucial roles in predicting contributions.
Table 12 presents the top ten important features predicted for each stereotype based on simulated crowdsourcing scores from LLM. For the competence dimension, among all features, the “Is in Hospital” feature contributes the most to the prediction, consistent with the results obtained from crowdsourced ratings from human participants. Other key factors predicting competence include the number of images uploaded by the beneficiary and whether the information is posted by a third person. For the warmth dimension, “Vitality versus Illness” has the highest predictive contribution among all features, indicating that whether the cover image uploaded by the beneficiary is health-themed has a significant impact on predicting perceived warmth. Additionally, variables such as whether the cover image uploaded by the beneficiary interacts with surroundings and whether the scene is in a hospital are also crucial variables for predicting perceived warmth.
Figure 10 illustrates the importance of three types of variables derived from crowdsourced ratings from human participants and GPT-4 LLM in predicting stereotypes, namely various visual features, textual features, and personal attributes, in predicting stereotypes. In both human-rated and LLM-rated evaluations, visual features contribute the most to predicting high/low competence and high/low warmth. Table 13 summarizes the top 10 features with the highest overall contribution when simultaneously predicting high/low competence and high/low warmth based on simulated crowdsourcing scores from Human and LLM. Based on the results of Human-rated, the top three features with the highest overall predictive contribution are Is in Hospital, Vitality versus Illness, and Smile. While based on the results of LLM, the top three features with the highest overall predictive contribution are “Is in Hospital,” “Vitality versus Illness,” and “Image Count.” Comparatively, Smile ranks fourth in overall predictive contribution. Overall, there is minimal difference in the ranking of feature importance for prediction between the two crowdsourcing approaches mentioned above.
(Human and LLM).
In the previous regression analysis, we established the relationships between competence, warmth, and stereotypes. Using machine learning based on crowdsourced ratings from human participants, the three most important variables predicting warmth and competence were identified as “Is in Hospital,” “Vitality versus Illness,” and “Smile.” From the results of correlation analysis, we observed that only “Smile” is significantly correlated with both warmth and competence. Therefore, we constructed the following parallel mediation models to analyze how this key feature impacts warmth, competence, and donation intention. Similarly, based on simulated crowdsourcing scores from LLM, the three most important variables predicting warmth and competence were identified as “Is in Hospital,” “Vitality versus Illness,” and “Image Count.” Considering the corresponding correlation analysis results, we observed that the variable “Image Count” does not exhibit significant correlation with either warmth or competence. As a result, this variable was excluded from the subsequent analysis. In the parallel mediation models, we systematically employed “Smile,” “Is in Hospital,” and “Vitality versus Illness” as independent variables, with warmth and competence serving as mediators, and donation intention as the dependent variable. Data analysis using Statistical Product and Service Solutions (SPSS) Version 26.0 for Windows (IBM Corp., Armonk, NY, USA). The mediating effect was evaluated using 5000 bootstrap samples via Model 4 (parallel mediation) of the SPSS PROCESS macro. Respectively, the variables “Smile,” “Is in Hospital,” and “Vitality versus Illness” were coded as binary (yes = 1, no = 0).
As shown in Fig. 11, we found that when the beneficiary in the image is smiling, it has a significant positive relationship to competence (B = 0.26; SE = 0.07; t = 3.75; p < 0.001) and warmth (B = 0.43; SE = 0.07; t = 5.95; p < 0.001). Competence (B = 0.41; SE = 0.08; t = 5.07; p < 0.001) and warmth (B = 0.39, SE = 0.08; t = 5.02; p < 0.001) exhibit positive and significant effects on donation intention. However, the direct effect of a smile on donation intention is not significant (B = -0.05; SE = 0.07; 95% CI: −0.18 to 0.09), indicating that whether the beneficiary smiles or not in the uploaded images does not directly impact the donor’s willingness to contribute. Nevertheless, considering the overall model, we found that the total effect of smile on donation intention is significantly positive (B = 0.23; SE = 0.08; 95% CI: 0.06 to 0.39). The total overall indirect effects from warmth and competence are also positive and significant (B = 0.27; SE = 0.06; 95% CI: 0.16 to 0.39). The specific indirect effects are as follows: the indirect effect from warmth is 0.1666 (SE = 0.04; 95% CI: 0.09 to 0.26), and the indirect effect from competence is 0.1060 (SE = 0.03; 95% CI: 0.04 to 0.18). Furthermore, there is no significant difference in the indirect effects contributed by warmth and competence (B = 0.06; SE = 0.05; 95% CI: −0.04 to 0.16).
Smile: 0 = the individual does not smile, 1 = the individual smiles. Values are unstandardized coefficients based on crowdsourced ratings from human participants. *p < 0.05. **p < 0.01. ***p < 0.001.
As shown in Fig. 12, we found that when the beneficiary in the image is in hospital, it has a significant positive relationship to competence (B = 0.27; SE = 0.08; t = 3.42; p < 0.001) and a significant negative relationship to warmth (B = −0.19; SE = 0.09; t = −2.00; p < 0.05). Competence (B = 0.26; SE = 0.07; t = 3.44; p < 0.001) and warmth (B = 0.29, SE = 0.06; t = 4.55; p < 0.001) exhibit positive and significant effects on donation intention. And the direct effect of Is in Hospital on donation intention is significant (B = 0.38; SE = 0.07; 95% CI: 0.23 to 0.52), indicating that if the cover image uploaded by the beneficiary depicts a scene in a hospital, it can have a direct positive and significant impact on the willingness of potential donors to donate. The total overall indirect effects from warmth and competence are not significant (B = 0.02; SE = 0.04; 95% CI: −0.07 to 0.10). The specific indirect effects are as follows: the indirect effect from warmth is -0.0537 (SE = 0.03; 95% CI: −0.11 to -0.01), and the indirect effect from competence is 0.0696 (SE = 0.05; 95% CI: 0.06 to 0.27). It can be inferred that “Is in Hospital” exerts an indirect positive effect on Donation Intention through Competence, while also exerting an indirect negative effect through Warmth. The positive and negative effects offset each other, resulting in an overall indirect effect that is not statistically significant.
Presence in Hospital: 0 = the individual is not in hospital, 1 = the individual is in hospital. Values are unstandardized coefficients based on crowdsourced ratings from LLM. *p < 0.05. **p < .0.01. ***p < 0.001.
As shown in Fig. 13, we found that whether the cover image exhibits a significant negative correlation with competence in portraying health narratives (B = −0.20; SE = 0.08; t = -2.61; p < 0.01), while it does not show a significant relationship with warmth. Competence (B = 0.28; SE = 0.07; t = 3.89; p < 0.001) and warmth (B = 0.28, SE = 0.06; t = 4.45; p < 0.001) exhibit positive and significant effects on donation intention. And the direct effect of Vitality versus Illness on donation intention is negatively significant (B = −0.37; SE = 0.07; 95% CI: −0.50 to −0.24), indicating that if the beneficiary uploads a cover image with a health narrative theme, it can have a significant negative impact on the donation intention of potential donors. The total overall indirect effects from warmth and competence are not significant (B = −0.10; SE = 0.04; 95% CI: −0.09 to 0.07). The specific indirect effects are as follows: the indirect effect from warmth is 0.05 (SE = 0.02; 95% CI: −0.10 to −0.02), and the indirect effect from competence is 0.10 (SE = 0.02; 95% CI: 0.06 to 0.15). It can be inferred that “Vitality versus Illness” exerts an indirect positive effect on Donation Intention through Competence, while also exerting an indirect negative effect through Warmth. The positive and negative effects offset each other, resulting in an overall indirect effect that is not statistically significant.
Vitality versus Illness: 0 = the cover image has an unhealthy narrative theme, 1 = the cover image has a healthy narrative theme. Values are unstandardized coefficients based on crowdsourced ratings from LLM. *p < 0.05, **p < 0.01, ***p < 0.001.
We conducted a study using both manual and LLMs to annotate images and texts in OMCCs to investigate the influence of visual features, textual features, personal attributes, perceived warmth, and perceived competence of beneficiaries on potential donors’ intention to donate to these campaigns. From a theoretical perspective, this study extends the understanding of donation behavior in OMCCs, specifically by exploring how potential donors make donation decisions based on stereotypes formed from online information disseminated by beneficiaries. By integrating a multimodal analysis of visual and textual elements, this study provides a novel analytical framework for crowdfunding researchers to assess and predict user behavior. Moreover, the application of machine learning models, specifically SVC model, not only validates the effectiveness of these techniques in predicting donor perceptions of warmth and competence, but also provides an empirical foundation for the development of predictive models in crowdfunding and other areas of social computing. Additionally, we simulated the human crowdsourcing rating process using LLMs, supplementing visual and textual features that were not addressed in the human participant crowdsourced ratings. On the practical application front, this study offers strategic insights for OMCCs beneficiaries, particularly in optimizing the stereotypes of potential donors through the design of cover images and the crafting of textual content, thereby increasing the success rate of fundraising. The findings underscore the importance of conveying a positive image in crowdfunding activities and provide crowdfunding platforms and other nonprofit organizations with specific guidance on how to design more engaging fundraising campaigns. At the same time, the study addresses critical ethical considerations in crowdfunding, including the prevention of negative stereotypes and the protection of donor information, which are critical to fostering a more equitable and inclusive crowdfunding environment.
In the realm of medical crowdfunding, our results highlight that perceived warmth and competence bolster donation intent. Donors tend to gravitate towards individuals who exude warmth and competence. The role of stereotypes in dictating donation behaviors in online initiatives becomes clear through this observation. For instance, when potential donors perceive high competence in a beneficiary, their donation actions are principally rational. This stems from the belief that individuals with strong capabilities are better equipped to overcome challenges, thus making them more likely to garner support. Conversely, perceived warmth in a beneficiary emotionally motivates donors by meeting their need for connection. These findings suggest that stereotypes substantially shape donation intentions, with cognitive and emotional factors intertwining to influence their decision-making within the crowdfunding context41.
In both human participant crowdsourced ratings and LLM-simulated crowdsourced ratings, visual features emerged as the most important predictors of competence and warmth levels, compared to visual, textual, and personal attributes. Visual features, such as emotional expression, body language, and the beneficiary’s appearance in an image, serve as immediate representations of critical information. Since that images are processed rapidly by the human brain, they may inherently influence potential donors’ stereotypes. Interestingly, we found that whether the beneficiary is depicted in a hospital setting in an image emerged as a critical predictor of perceived competence, while the presence of a smile was of the highest importance for predicting warmth. Furthermore, in the combined prediction tasks, the Vitality versus Illness ranked second in importance. These findings shed light on the significant role visual cues, particularly related to the patient’s medical condition and context, play in shaping potential donors’ perceptions. In contrast, textual features demand a higher level of cognitive engagement and time to process and interpret. It is noteworthy to discuss that this study revealed an interesting observation: a smile does not have a significant direct impact on donation intention. Instead, it indirectly influences potential donors’ willingness to contribute by enhancing their perception of warmth and competence in the beneficiary. While some earlier research postulates the positive impact of a smile on crowdfunding outcomes42,43, others argue that smiling has no significant effect on the success rates44. From the perspective of LLM-simulated crowdsourcing results, the predictive model for stereotypes based on LLM-simulated crowdsourced ratings closely resemble human ratings, with the difference being that the variable “Image Count” enters the top three in terms of predictive importance, while ranking fourth in the predictive task based on crowdsourced ratings from human participants. Additionally, we have enriched the mechanism of key variables affecting donation intention. The variable “Is in Hospital” presented in cover image which is uploaded by the beneficiary has a direct impact on donation intention, but this direct effect is offset by the opposite indirect effects brought about by the two mediating variables, warmth and competence. Whether the beneficiary is in a hospital has a significant direct effect on donation intention. However, this effect is accompanied by two opposing indirect effects, mediated by perceived warmth and perceived competence, which counteract each other and result in a non-significant total indirect effect. “Vitality versus Illness” only has a significant negative impact on competence and does not significantly affect warmth. Furthermore, “Vitality versus Illness” can directly and significantly negatively affect donation intention, while the overall indirect effect is also offset by the opposite indirect effects brought about by the two mediating variables, warmth and competence.
Building upon these studies, our research further explores the mechanisms through which the variables of “smile,” “Is in Hospital,” and “Vitality versus Illness” influence donation intention. What’s more interesting is that we discovered a positive impact of a smile on both the dimensions of warmth and competence, establishing that the perceived warmth and competence have a positive influence on potential donors’ willingness to contribute. A smile serves as a positive emotional expression, conveying warmth and friendliness. When potential donors witness a smile from the beneficiary, they experience a positive emotion, fostering a sense of closeness and building a favorable impression of the beneficiary. This, in turn, is likely to stimulate the psychological engagement of potential donors. Moreover, a smile can convey the beneficiary’s optimistic and resilient mindset. This emotional signal influences potential donors, making them more likely to pay attention to and empathize with the beneficiary’s situation. Consequently, this enhances the willingness of potential donors to contribute.
Our research found that the features of the written information released by Online Medical Crowdfunding Campaigns are also related to the stereotypes of potential donors of beneficiaries regardless of whether it is based on human participant crowdsourced ratings or LLM-simulated crowdsourced ratings. First, the more complex the text description, the more effective it is in expressing the perceived warmth of the beneficiaries. Detailed textual descriptions commonly provide a wealth of information, which, in turn, facilitate donors’ deeper understanding of the beneficiary’s circumstances and requirements. These detailed descriptions often include the beneficiary’s medical condition, background, accomplishments, and other pertinent factors. Consequently, such detailed descriptions tend to facilitate an emotional rapport between the beneficiary and potential donors, thereby heightening the perceived warmth attributed to the beneficiary. In contrast, overly simplistic text descriptions may lack specific information and emotional elements, making it challenging to evoke a sense of empathy in potential donors toward the beneficiary. The difference lies in that LLM appears to pay more attention to textual information. When the described text involves more emotional vocabulary, the depiction of emotional words triggers an enhanced emotional connection with the beneficiary, evoking empathy and compassion, thereby increasing perceived warmth. On the other hand, an increase in descriptions related to health vocabulary leads to a decrease in perceived warmth and competence. This may be due to health-related terms often involving complex medical conditions, which can be difficult to understand, potentially leading to difficulties in comprehending the text content, reducing the perceived warmth towards the beneficiary. Moreover, these terms often reflect the beneficiary’s physical condition, which also reduces the perception of their competence.
Our study primarily focused on the static features of medical crowdfunding campaigns, analyzing visual elements (e.g., number of images, depiction of illness versus vitality, presence of a smile, and interaction) as well as textual features and personal attributes (e.g., content of the text, age, and gender of the beneficiary). However, real-world donation behaviors are influenced by dynamic factors such as the number of ongoing donations, comments, likes, and shares. Future research could extend our approach by incorporating these dynamic elements to better capture how evolving campaign interactions impact donor perceptions and behaviors. Additionally, longitudinal data collection would allow us to analyze how stereotypes and donation intentions shift over time, providing deeper insights into the temporal dynamics of medical crowdfunding success.
While our findings underscore the predictive power of visual features, the influence of textual descriptions remains underexplored. Future studies should place greater emphasis on textual pathways, identifying which specific textual elements have the most significant impact on perceived warmth and competence. This could involve a more granular analysis of narrative structures, linguistic styles, and the interplay between emotional and informational content. A deeper understanding of textual influence would further refine strategies for optimizing campaign descriptions to enhance donor engagement.
Furthermore, our study’s findings are based on data and participants from the United States, which raises concerns about generalizing these results to other cultural contexts, such as collectivist cultures (e.g., Asian countries) or regions with different healthcare systems. Cultural variations in perceptions of warmth and competence, as well as donation behaviors, may lead to different outcomes in non-Western settings. Future studies are needed to replicate this study in diverse cultural contexts to validate its findings and uncover potential cross-cultural differences.
Although sufficient for initial analysis, the sample size of 250 online medical crowdfunding campaigns may limit the generalizability of our findings. Expanding the dataset to include a larger and more diverse range of campaigns will enhance the robustness of predictive models and provide stronger empirical foundations for stereotype research in crowdfunding. Additionally, while this study focuses on medical crowdfunding, it overlooks the diversity of campaign goals, such as paying for medical bills, funeral costs, or other expenses. These differences may significantly influence donor behavior, and future research could further explore how different crowdfunding goals affect donor decisions.
In OMCCs, donors’ decisions are influenced by various factors, with stereotypes playing a critical role. However, the precise identification and quantification of these stereotypes, as well as the specific mechanisms linking them to the characteristics of crowdfunding campaigns, such as visual features (e.g., image aesthetics, facial expressions, interaction with surroundings) and textual features (e.g., description length, vocabulary use, health-related terms), remain unclear. Previous studies have often focused on the analysis of single factors, lacking a systematic approach that considers the integration of multiple factors. This study aims to address these issues by collecting and analyzing an informative dataset from the GoFundMe platform. Using both human and large language model (LLM)-based methods, we investigate the complex relationship between stereotypes and donation intentions. Through a detailed analysis of personal attributes (e.g., age, gender), visual features, and textual features, we employ machine learning models to predict and explain how these factors collectively influence donors’ stereotypes and their willingness to donate. This approach provides a more comprehensive and nuanced understanding of both research and practice in the field of online medical crowdfunding. The specific research process is shown in Fig. 14.
To protect the privacy of individuals, all identifiable information, including faces, names, and specific details, has been masked in the image.
We used data from the GoFundMe platform, the largest online medical crowdfunding platform in the United States. According to East China Normal University’s Institutional Review Board policy, this study was exempt from an ethics review, as it exclusively utilized publicly available data. We created a Python web-scraping tool to gather data on cancer-related crowdfunding campaigns from a platform, covering the period from January 1, 2019, to July 12, 2021, resulting in 156,551 entries. This data included 18 textual fields such as campaign type, ID, donation count, description, country, title, goal amount, current balance, percentage of goal reached, YouTube URL, description and title length, zip code, creation and update times, and comments.
In analyzing the dataset, we found that the majority of online medical crowdfunding projects predominantly feature individual beneficiaries. To ensure consistency, our dataset only includes projects with individual beneficiaries and single-person cover images. The stereotype content model suggests that warmth and competence are independent dimensions. At the individual level, warmth and competence tend to be positively correlated. At the group level, however, they are often negatively correlated, meaning that many groups are rated high on one dimension and low on the other45. From a social relations perspective, because the cases we collected predominantly featured individual beneficiaries, the descriptions of the images in the cases were broadly categorized into two scenarios: single-person (the beneficiary alone) and multi-person (the beneficiary with family, friends, etc.). Multi-person images often involve more complex social relationships. Research suggests that conflicting or ambiguous relationships between social groups contribute to the formation of negative stereotypes, whereas cooperative relationships contribute to positive stereotypes46. The selection of single-person images as project covers serves the purpose of highlighting the beneficiaries and mitigating potential confusion or distraction that may arise from the presence of unrelated individuals in the images. Finally, our data has 19,225 crowdfunding projects with a single image as the project cover.
To manage the impracticality of using a large dataset for our research due to time, effort, and resource constraints, we employed a random sampling method to select 250 online crowdfunding projects from the original dataset. This approach balanced the need to avoid the issues of a too-large dataset with the risk of a too-small dataset that lacks meaningful insights. This selection aimed to reduce bias from stereotypical impressions due to intergroup relations and to maintain the consistency of our experimental data46.
In this study, we adopted the Stereotype Content Model (SCM) to evaluate two critical dimensions of stereotype impressions: warmth and competence. The warmth dimension encompasses attributes such as friendliness, helpfulness, sincerity, trustworthiness, and morality. Warmth relates to social acceptance and the sense of connection vital for survival. Research has shown that perceptions of warmth can influence interpersonal interactions and decisions to assist others, as people are more inclined to form friendships with warm individuals. According to the BIAS Map theory15, individuals are more likely to engage in helping behaviors toward groups perceived as high in warmth. In the context of donations, Cuddy47 mentioned that beneficiaries characterized by high warmth may also evoke a stronger willingness to donate. On the other hand, the competence dimension involves attributes like intelligence, skill, creativity, and efficacy. Competence pertains to the status and capabilities necessary for achieving goals. Studies have found that incorporating descriptions related to competence perceptions in donation materials can enhance people’s willingness to donate. This is because when individuals perceive that the beneficiaries, even in adversity, still possess a sense of competence, they believe that their assistance can help the beneficiaries overcome challenges. Consequently, they view their donation as necessary and are more likely to contribute48. To collect labels for stereotype impressions and gauge the willingness to donate, we adapted composite measures from previous study37,49 and constructed a crowdsourcing task.
The entire process of crowdsourcing work is presented in Fig. 15. Our crowdsourcing task is divided into the following two steps.
The process of Crowdsourcing.
For each assigned task, human participants were tasked with answering three questions regarding the two dimensions of stereotype impressions (warmth and competence), as well as two questions concerning donation willingness. The inclusion of control questions helped us assess the attentiveness of the participants, with specific response options to be selected. Our questionnaire was carefully tailored to address our research objectives. Each query was prefaced with the phrase “How does this fundraising information make you feel…”, and participants were asked to respond on a 5-point Likert scale that ranged from “1—strongly disagree” to “5—strongly agree.” The comprehensive list of these questions can be found in Table 14. Ultimately, each piece of information was evaluated by three different participants. We mitigated the influence of individual rating preferences by averaging the ratings from all three participants for each project.
To mitigate the influence of cultural variances among participants of different nationalities, we recruited 150 workers from the United States through the Prolific platform, requiring these workers to be over 18 years old and native English speakers. Furthermore, we gathered essential demographic data from the participants, including their age, gender, ethnicity, and educational background. Refer to Table 15 for specific details regarding participant demographics. Each participant was then randomly assigned five projects. Each project included a title, a detailed description, and an accompanying image. Participants were instructed to rate these projects. For each project, evaluations were conducted by three distinct individuals, and the mean score derived from these assessments was adopted as the foundational data within the study. Among all participants, 70% are female.
To further supplement the disparities in stereotype impressions and donation intentions due to human subjective factors, this study simulated the crowdsourcing process of the aforementioned human participants based on GPT-450. Given that online medical crowdfunding projects entail both textual and pictorial information, we opted for the use of GPT-4, a large language model capable of processing multimodal inputs, to support the simulated crowdsourcing tasks. The prompt template for the crowdsourcing tasks based on LLM consists of three components (as shown in Fig. 16). The first component is the identity description, where, in accordance with the crowdsourcing tasks conducted by human participants as described earlier, each online medical crowdfunding project is evaluated along the dimensions of warmth, competence, and donation intention by three distinct human participants. In each round of tasks, the prompt sequentially inputs the basic information of each human participant, including race, gender, age, and education level, to delineate identities for LLM. The second component is the project description, encompassing inputs for the crowdfunding project description and cover image. The third component is the task description, wherein questions outlined in Table 14 are asked one by one, with LLM instructed to provide numerical responses exclusively. Consequently, a total of 241 online medical crowdfunding projects were involved in this crowdsourcing task. (Due to privacy concerns aligned with OpenAI’s privacy policy, nine out of the 250 online medical crowdfunding projects’ information did not participate in the simulated crowdsourcing process.)
To protect the privacy of individuals, all identifiable information, including faces, names, and specific details, has been masked in the image.
Table 16 provides the list and the definitions of independent variables. Personal attributes have individual-specific details. We analyzed the relationships between age, gender, stereotypes, and donation intention. Baidu AI (https://ai.baidu.com/tech/face/detect) was employed to examine the features related to these characteristics.
Considering that the visual features presented by an image may influence the viewers’ stereotypes and donation behaviors, we used several features to describe images, including the feature signifying whether the individual in the cover image posted by the fundraiser made direct eye contact with the camera (Face Contact), the feature assessing whether there is any interaction between the individuals in the cover image and their surroundings (Interaction with Surroundings), the feature identifying whether the individuals in the cover image are situated in a hospital (Is in Hospital). Following Guo et al.49, we categorized image themes based on whether or not the image showed vitality.
In terms of textual features, the existing literature on online crowdfunding suggests that certain characteristics play a pivotal role in a campaign’s fundraising performance. Essential text features include a range of factors: the length of the description and title, the complexity of both description and title, the use of health-related, financial, occupational, and emotional words, mentions of medical history, cancer stage, personal relationships, and recurrence20,50,51,52,53 Recognizing the potential relevance of these textual features within online medical crowdfunding, we employed some tools to identify and compute these factors.
First, we used the ChatGPT 3.5 API to process project descriptions and extract medical history, cancer stage, relationships, and recurrences. A prompt directs the model with system and user information, focusing on system roles and task requirements. We crafted a prompt for extracting these details, illustrated in Fig. 17.
The process construction of prompt.
The Linguistic Inquiry and Word Count54 (LIWC) software facilitates objective and quantitative analysis of texts based on preestablished language categories, such as grammatical, psychological, and social relevance. Thus, we applied the LIWC dictionary to analyze the occurrence of health-related, money-related, work-related, and emotion-related words within the project description. Upon completion of the full lexical matching, the counts for each word category were normalized by dividing them by the length of the text to control for text length effects on the metrics.
Some studies have also indicated that text features, such as length55,56, readability57, and narrative perspective20 can impact the outcomes of crowdfunding campaigns. Thus, we employed the Python library “textstat” to assess the complexity of both project titles and descriptions. For instance, relationship variables, which represent the connection between the poster and the beneficiary (e.g., self, family, friend, or others).
We framed the task of predicting stereotypes as a binary classification problem, categorizing stereotypes into two dimensions: warmth and competence. Medical crowdfunding projects were classified as high/low warmth or high/low competence based on their scores relative to the average. Projects exceeding the average warmth score were labeled as “high warmth”, indicating a warmth level above the standard. A similar method was applied to competence. We employed three machine learning algorithms: Decision Tree58, Random Forest59, and Support Vector Classifier (SVC)60. Decision Trees create a structure by recursively partitioning features, while Random Forests build multiple trees using random subsets of features and samples, aggregating their outputs. SVCs use non-linear kernels to find optimal decision boundaries. Previous studies61,62,63 have successfully employed these algorithms and utilized manually annotated features for predictive modeling.
Our models used visual, textual, and personal attribute features (Table 16) and were trained using Scikit-learn, a Python machine learning library. We evaluated model performance using metrics including precision, recall, F1-score, AUC, and the ROC curve. Precision measures the accuracy of positive predictions, while recall reflects the proportion of actual positives that were correctly identified. The F1-score balances precision and recall. ROC curves plot the True Positive Rate against the False Positive Rate, with AUC indicating the classifier’s performance. The formulas for calculating the aforementioned evaluation metrics are as follows:
Given the small dataset, we employed 10-fold cross-validation, dividing the dataset into ten subsets, iteratively using one for testing and nine for training. This approach helped ensure robust and reliable model performance evaluation.
The data supporting the findings of this study are available from the corresponding author upon reasonable request.
Bray, F., Jemal, A., Grey, N., Ferlay, J. & Forman, D. Global cancer transitions according to the Human Development Index (2008–2030): a population-based study. Lancet Oncol. 13, 790–801 (2012).
Article Google Scholar
Richard, P., Walker, R. & Alexandre, P. The burden of out of pocket costs and medical debt faced by households with chronic health conditions in the United States. PloS ONE 13, e0199598 (2018).
Article Google Scholar
Landwehr, M. S., Watson, S. E., Macpherson, C. F., Novak, K. A. & Johnson, R. H. The cost of cancer: a retrospective analysis of the financial impact of cancer on young adults. Cancer Med. 5, 863–870 (2016).
Article Google Scholar
Burtch, G. & Chan, J. Investigating the relationship between medical crowdfunding and personal bankruptcy in the United States: Evidence of a digital divide. MIS. Q. 43, 237–262 (2019).
Article Google Scholar
Saleh, S. N., Ajufo, E., Lehmann, C. U. & Medford, R. J. A comparison of online medical crowdfunding in Canada, the UK, and the US. JAMA Netw. Open. 3, e2021684 (2020).
Article Google Scholar
Ba, Z., Zhao, Y., Zhou, L. & Song, S. Exploring the donation allocation of online charitable crowdfunding based on topical and spatial analysis: evidence from the Tencent GongYi. Inf. Process. Manage. 57, 102322 (2020).
Article Google Scholar
Jin, P. Medical crowdfunding in China: empirics and ethics. J. Med. Ethics. 45, 538–544 (2019).
Article Google Scholar
Berliner, L. S. & Kenworthy, N. J. Producing a worthy illness: Personal crowdfunding amidst financial crisis. Soc. Sci. Med. 187, 233–242 (2017).
Article Google Scholar
Pröbster, M. & Marsden, N. The social perception of autonomous delivery vehicles based on the stereotype content model. Sustainability 15, 5194 (2023).
Article Google Scholar
Bosco, C. et al. Detecting racial stereotypes: an Italian social media corpus where psychology meets NLP. Inf. Process. Manage. 60, 103118 (2023).
Article Google Scholar
Strinić, A., Carlsson, M. & Agerström, J. Occupational stereotypes: professionals´ warmth and competence perceptions of occupations. Pers. Rev. 51, 603–619 (2021).
Article Google Scholar
Fiske, S. T., Cuddy, A. J. C., Glick, P. & Xu, J. A model of (often mixed) stereotype content: Competence and warmth respectively follow from perceived status and competition. J. Pers. Soc. Psychol. 82, 878–902 (2002).
Article Google Scholar
Kim, T. & Ball, J. G. Unintended consequences of warmth appeals: an extension of the compensation effect between warmth and competence to advertising. J. Advertising 50, 622–638 (2021).
Article Google Scholar
Tian, Y., Yang, J. & Chuenterawong, P. Share or not? Effects of stereotypes on social media engagement using the stereotype content model. Journal. Pract. 17, 574–600 (2023).
Google Scholar
Cuddy, A. J. C., Fiske, S. T. & Glick, P. The BIAS map: Behaviors from intergroup affect and stereotypes. J. Pers. Soc. Psychol. 92, 631–648 (2007).
Article Google Scholar
Zuo, B., Dai, T., Wen, F. & Teng, T. The relationship between warmth and competence in social cognition. Adv. Psychol. Sci. 22, 1467–1474 (2014).
Article Google Scholar
Zhang, X., Liu, X., Wang, X., Zhao, H. & Zhang, W. Exploring the effects of social capital on crowdfunding performance: A holistic analysis from the empirical and predictive views. Comput. Hum. Behav. 126, 107011 (2022).
Article Google Scholar
Yoo, J. J., Song, S. & Jhang, J. Overhead aversion and facial expressions in crowdfunding. J. Retail. Consum. Servs. 69, 103101 (2022).
Article Google Scholar
Xu, L. Z. Will a digital camera cure your sick puppy? Modality and category effects in donation-based crowdfunding. Telemat. Inform. 35, 1914–1924 (2018).
Article Google Scholar
Durand, W. M. et al. Medical crowdfunding for organ transplantation. Clin. Transplant. 32 (2018).
Sundar, S. S. Multimedia effects on processing and perception of online news: a study of picture, audio, and video downloads. J. Mass Commun. Q. 77, 480–499 (2000).
Google Scholar
Appiah, O. Rich media, poor media: the impact of audio/video vs. text/picture testimonial ads on browsers’evaluations of commercial web sites and online products. J. Curr. Iss. Res. Ad. 28, 73–86 (2006).
Google Scholar
Liu, H., Feng, S. & Hu, X. Process vs. outcome: Effects of food photo types in online restaurant reviews on consumers’ purchase intention. Int. J. Hosp. Manag. 102, 103179 (2022).
Article Google Scholar
Cheng, Y., Mukhopadhyay, A. & Williams, P. Smiling signals intrinsic motivation. J. Consum. Res. 46, 915–935 (2019).
Article Google Scholar
Raab, M., Schlauderer, S., Overhage, S. & Friedrich, T. More than a feeling: Investigating the contagious effect of facial emotional expressions on investment decisions in reward-based crowdfunding. Decis. Support Syst. 135, 113326 (2020).
Article Google Scholar
Wang, Z., Mao, H., Li, Y. J. & Liu, F. Smile big or not? Effects of smile intensity on perceptions of warmth and competence. J. Consum. Res. 43, 787–805 (2017).
Google Scholar
Chen, T. & Dredze, M. Vaccine images on twitter: analysis of what images are shared. J. Med. Internet Res. 20, e130 (2018).
Article Google Scholar
Li, Y. & Xie, Y. Is a picture worth a thousand words? An empirical study of image content and social media engagement. J. Mark. Res. 57, 1–19 (2019).
Article Google Scholar
Laroche, M., Li, R., Richard, M.-O. & Zhou, M. An investigation into online atmospherics: the effect of animated images on emotions, cognition, and purchase intentions. J. Retail. Consum. Serv. 64, 102845 (2022).
Article Google Scholar
Maier, E. & Dost, F. Fluent contextual image backgrounds enhance mental imagery and evaluations of experience products. J. Retail. Consum. Serv. 45, 207–220 (2018).
Article Google Scholar
Chandler, J., Rosenzweig, C., Moss, A. J., Robinson, J. & Litman, L. Online panels in social science research: expanding sampling methods beyond Mechanical Turk. Behav. Res. Methods 51, 2022–2038 (2019).
Article Google Scholar
Veselovsky, V., Ribeiro, M. H., & West, R. Artificial artificial artificial intelligence: crowd workers widely use large language models for text production tasks. https://arxiv.org/abs/2306.07899 (2023).
Elyoseph, Z., Hadar-Shoval, D., Asraf, K. & Lvovsky, M. ChatGPT outperforms humans in emotional awareness evaluations. Front. Psychol. 14, 1199058 (2023).
Article Google Scholar
Kaikaus, J., Li, H. & Brunner, R. J. Humans vs. ChatGPT: evaluating annotation methods for financial corpora. In 2023 IEEE International Conference on Big Data (BigData), 2831–2838 (IEEE, 2023).
Wang, H. et al. Scientific discovery in the age of artificial intelligence. Nature 620, 47–60 (2023).
Article Google Scholar
Zheng, Z., Yu, S., Islami, F., Banegas, M. P. & Yabroff, R. Cancer survivors’ use of crowdfunding campaigns: AI-generated patterns of reported medical financial hardship and unmet social needs. JCO Oncol. Pract. 19, 542 (2023).
Article Google Scholar
Ye, T. et al. Using artificial intelligence to unlock crowdfunding success for small businesses. SSRN Scholarly Paper at https://doi.org/10.2139/ssrn.4806426 (2024).
Arnold, C., Xu, L. Z., Saffarizadeh, K. & Madiraju, P. Generative Ai as (Un)Welcome Agents in Medical Crowdfunding: The Trust Dilemma and Moral Hazard. SSRN Scholarly Paper at https://doi.org/10.2139/ssrn.4725206 (2024).
Gilardi, F., Alizadeh, M. & Kubli, M. ChatGPT outperforms crowd workers for text-annotation tasks. Proc. Natl. Acad. Sci. USA. 120, e2305016120 (2023).
Article Google Scholar
Zhang, X., Lyu, H. & Luo, J. What contributes to a crowdfunding campaign’s success? Evidence and analyses from GoFundMe data. J. Social Comput. 2, 183–192 (2021).
Article Google Scholar
Vossen, H. G. M., Piotrowski, J. T. & Valkenburg, P. M. Development of the Adolescent Measure of Empathy and Sympathy (AMES). Pers. Indiv. Differ. 74, 66–71 (2015).
Article Google Scholar
Li, Y., Wu, S. & Zhou, W. The effect of syntax simplicity on crowdfunding performance. Ventur. Cap. 1, 24 (2023).
Google Scholar
Yoo, J. J., Jhang, J., Song, S. & Shin, H. S. An integrated model of prosocial crowdfunding decision: three utility components and three informational cues. Electron. Commer. R. A. 57, 101233 (2023).
Article Google Scholar
Li, Y., Xiao, N. & Wu, S. The devil is in the details: the effect of nonverbal cues on crowdfunding success. Inf. Manage. 58, 103528 (2021).
Google Scholar
Fiske, S. T., Cuddy, A. J. C. & Glick, P. Universal dimensions of social cognition: warmth and competence. Trends Cognit. Sci. 11, 77–83 (2007).
Article Google Scholar
Bar-Tal, D. Formation and change of ethnic and national stereotypes: an integrative model. Int. J. Intercult. Rel. 21, 491–523 (1997).
Article Google Scholar
Cuddy, A. J., Fiske, S. T. & Glick, P. Warmth and competence as universal dimensions of social perception: the stereotype content model and the BIAS map. Adv. Exp. Social Psychol. 40, 61–149 (2008).
Article Google Scholar
Liang, J., Chen, Z. & Lei, J. Inspire me to donate: the use of strength emotion in donation appeals. J. Consum. Psychol. 26, 283–288 (2015).
Article Google Scholar
Aaker, J., Vohs, K. D. & Mogilner, C. Nonprofits are seen as warm and For-Profits as competent: firm stereotypes matter. J. Consum. Res. 37, 224–237 (2010).
Article Google Scholar
Arora, N., Chakraborty, I. & Nishimura, Y. Revolutionizing marketing research with a large language model: a hybrid AI-human approach. SSRN Scholarly Paper at https://doi.org/10.2139/ssrn.4683054 (2024).
Dragojlovic, N. & Lynd, L. D. What will the crowd fund? Preferences of prospective donors for drug development fundraising campaigns. Drug Discov. Today. 21, 1863–1868 (2016).
Article Google Scholar
Gorbatai, A. D. & Nelson, L. Gender and the language of crowdfunding. Acad. Manag. Proc. 2015, 15785 (2015).
Article Google Scholar
Parsons, L. M. The impact of financial information and voluntary disclosures on contributions to Not-For-Profit organizations. Behav.Res. Account. 19, 179–196 (2007).
Article Google Scholar
Jaidka, K. et al. Estimating geographic subjective well-being from Twitter: a comparison of dictionary and data-driven language methods. Proc. Natl. Acad. Sci. USA. 117, 10165–10171 (2020).
Article Google Scholar
Zhang, F., Xue, B., Li, Y., Li, H. & Liu, Q. Effect of textual features on the success of medical crowdfunding: model development and econometric analysis from the Tencent Charity Platform. J. Med. Internet Res. 23, e22395 (2021).
Article Google Scholar
Zhang, X., Tao, X., Ji, B., Wang, R. & Sörensen, S. The success of cancer Crowdfunding Campaigns: project and text analysis. J. Med. Internet Res. 25, e44197 (2023).
Article Google Scholar
Koole, M. A. C., Kauw, D., Winter, M. M. & Schuuring, M. J. A successful crowdfunding project for eHealth research on grown-up congenital heart disease patients. Int. J. Cardiol. 273, 96–99 (2018).
Article Google Scholar
Simon, H. A., Hunt, E. B., Marin, J. & Stone, P. Experiments in induction. Am. J. Psychol. 80, 651 (1967).
Article Google Scholar
Breiman, L. Random Forests. Mach. Learn. 45, 5–32 (2001).
Article Google Scholar
Boser, B. E., Guyon, I. M. & Vapnik, V. N. A training algorithm for optimal margin classifiers. In: Proc. 5th Annual Workshop on Computational Learning Theory 144–152 (Association for Computing Machinery, 1992). https://doi.org/10.1145/130385.130401.
Li, H., Chen, X., Zhang, Y. & Hai, M. Prediction of financing goal of crowdfunding projects. Procedia Comput. Sci. 139, 108–113 (2018).
Article Google Scholar
Wang, W., Zheng, H. & Wu, Y. J. Prediction of fundraising outcomes for crowdfunding projects based on deep learning: a multimodel comparative study. Soft Comput. 24, 8323–8341 (2020).
Article Google Scholar
Xu, Y. & Zhu, N. Successful factors and prediction of crowdfunding on WeChat. Am. J. Ind. Bus. Manag. 8, 946–962 (2018).
Google Scholar
Download references
School of Economics and Management, East China Normal University, Shanghai, China
Xupin Zhang
School of Public Affairs, Zhejiang University, Hangzhou, China
Xiaorong Zheng
Department of Computer Science, University of Rochester, Rochester, NY, USA
Jiebo Luo
PubMed Google Scholar
PubMed Google Scholar
PubMed Google Scholar
Xupin Zhang: Conceptualization, Methodology, Software, Formal Analysis, Data Curation, Writing—Original Draft, Supervision, Writing—Review & Editing.Xiaorong Zheng: Software, Formal Analysis, Validation, Writing—Original Draft.Jiebo Luo: Conceptualization, Supervision, Writing—Review & Editing.
Correspondence to Xupin Zhang.
The authors declare no competing interests.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
Reprints and permissions
Zhang, X., Zheng, X. & Luo, J. The influence of stereotypes and visual features on donation intentions in online medical crowdfunding campaigns: A comparison of survey and large language model-based methods. npj Artif. Intell. 1, 16 (2025). https://doi.org/10.1038/s44387-025-00008-8
Download citation
Received:
Accepted:
Published:
Version of record:
DOI: https://doi.org/10.1038/s44387-025-00008-8
Anyone you share the following link with will be able to read this content:
Sorry, a shareable link is not currently available for this article.
Provided by the Springer Nature SharedIt content-sharing initiative
npj Artificial Intelligence (2026)
Advertisement
npj Artificial Intelligence (npj Artif. Intell.)
ISSN 3005-1460 (online)
© 2026 Springer Nature Limited
Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.