Questions for Doug Guilbeault on "Online Images Amplify Gender Bias".

[jamesallenevans]

Please pose thoughtful questions for our speaker by Wednesday midnight, and upvote 5 by Thursday @ 10am, an hour before our session together. Because the paper is under strict embargo from Nature, I will email the draft all students in the program.

[C-y22]

In your study, you highlighted the significant exacerbation of gender bias in online images compared to text, and its impact on societal perceptions and beliefs. Considering the rapid evolution of AI and its increasing use in generating images, how do you foresee AI's role in mitigating or further amplifying these biases? Additionally, what measures do you suggest could be implemented in the development and training of AI models to address and reduce the perpetuation of such biases?

[ymuhannah]

In your study, you discuss the underrepresentation of women in online images, particularly in the context of occupations and social roles. Have you also analyzed the representation of men and women in more negative categories, such as 'criminal,' in Google images? I am curious to know if the underrepresentation of women is limited to positive or neutral categories, or if it is a consistent trend across various types of categories, regardless of their positive or negative connotations. Also, the study includes a comparison of gender bias in images versus text. How did you ensure that the text data was appropriately matched and comparable to the image data for an accurate analysis?

[quanghieu31]

In big tech companies that have access to large image data for training models, they would consider the costs of using unsupervised learning model to label the gender in images rather than using a human-encoded model. However, both of these means can surely have bias as algorithm is based on people and people are biased everywhere and in every culture. Furthermore, these algorithms developed by the big tech companies are considered "black boxes" because we, the public, just could never know what kinds of underlying operations or data that they used. What do you think? How can we do better to avoid gender bias in this particular case?

[HongzhangXie]

Thank you for sharing insights into this intriguing research. In recent years, advancements in data transmission technology have led to a significant increase in the availability of images and videos on the internet. Understanding the potential gender biases inherent in these images is crucial. I am interested in two questions. Firstly, why do images tend to amplify gender biases more than text does? I guess this could be due to the ease with which people access information in images or their stronger retention of visual content. Secondly, I am curious about whether videos, in comparison to images and text, might generate even stronger biases.

[kiddosso]

Thanks for your insightful thoughts on the gender bias in the online age. The first question comes to my mind after reading the bulk of this paper is that did you observe any change of the degree to which online images contribute to gender bias during last 20 to 30 years? In other word, is such an effect a relatively consistent phenomenon ever since the birth of online images or a newly emerging issue in the last couple of years? Although it requires even more data and more complex models to answer this question, I feel this historical aspect is worth of further investigation.

[yunshu3112]

I am very impressed by the methods and the findings of this study. I have also read a paper that addresses the racial stereotypes amplified by the algorithms that displays digital advertisements to the targeted audience. I wonder what do you think is the future direction for the regulation and development of algorithms on social media platforms to reduce discrimination. Thank you!

[lguo7]

Thank you for sharing your research findings. As mentioned, the existence and exacerbation of gender bias in online images is a significant concern. I am curious about how you differentiate between gender bias and stereotypes. For instance, the common portrayal of most nurses as female and engineers as male in images—does this fall under gender stereotypes or align with gender bias, or is there a different interpretation?

[hchen0628]

Thank you so much for sharing your thoughts and research. Your research on images beyond the traditional analysis of text is very enlightening. I wonder what you think about the bias embedded in other information that may be made available to people on the Internet, such as videos and music. And what about the biases embedded in how information is provided on the Internet, such as recommended systems or censorship of information?

[MaoYingrong]

I think this research finding is worth the public's attention since today images indeed play a significant role in social media. I'm wondering whether there is any viable and effective solution that can alleviate this problem. Because gender bias is not about individual behaviors that can be constrained by censorship, gender bias is caused by the collective of the public, and advocation is too indirect and works very slowly.

[yuzhouw313]

Thank you, Professor Gilbeault, for presenting your insightful research. Your findings illuminate the concerning issue of gender bias, which to my surprise is more pronounced and psychologically impactful in image content compared to text. Regarding the ongoing efforts to mitigate such biases through online censorship—whether enforced by platforms' security teams or initiated by users themselves—I am curious about the specific approaches you foresee for image-based content. How might social media platforms effectively deploy algorithms to not only detect but also actively mitigate the presence of gender biases in the images shared within their networks?

[ecg1331]

Thank you for sharing this interesting research. I am curious about whether you think the amplification of gender biases in online images can be looked at as more of a ‘lag’ in reflecting societal opinion. For example, if you repeated the experiments with data that was only from 2019+ do you think would you get the same results, or would it be closer to the things you observed in data like the 2019 census?

[QIXIN-ACT]

In the study, gender bias is examined in relation to professional roles, which are relatively neutral. I'm interested in exploring whether this bias extends to words with clear emotional connotations, such as positive terms like 'hero' and negative ones like 'criminal.' Does gender bias exist in the context of these emotionally charged words, and if so, is it more pronounced or less compared to neutral terms?

[zhuoqingli526]

Thank you for sharing this enlightening research. It made me realize that images can evoke stronger gender biases compared to text. I noticed that this study analyzed the online sources of images in the last part, and I'm curious to know: Do images originating from influential sources like News and Government websites contribute to reinforcing these biases more significantly?

[natashacarpcast]

Very interesting paper! I have a question regarding the experiment. Since the participants were recruited from a crowdsourcing platform, I assume no minors were included; am I right? If that's the case, I would think it would be very interesting to replicate the experiment, including kids and teenagers, and also separating adults into age bins. Why? I don't know if I'm 100% correct, but I do have the impression that younger generations have been raised very differently, especially in the education they have received regarding inclusion and diversity. What are your thoughts on this? Do you think that there would be differences between age groups in the experiment?

[Hai1218]

I am particularly interested in reading about the experiment which shows that googling for images rather than textual descriptions of occupations amplifies gender bias in participants’ beliefs. Considering images are more potent in priming a gender biased response, do you see a relevant connections between that finding with algorithmic bias (justice)? To put more plainly, was the reason behind images being more pravalent in "boosting" gender bias due to the underlying "biased" algorithm by Google or other services that rank the order of , or pick and choice the images we see?

[AnniiiinnA]

Thank you for sharing this remarkable work with us! I'm wondering how can online platforms and content creators work towards reducing gender bias in visual representation. Also, with the popularity of short video platforms and video browsing, I would like to ask if similar techniques can be applied to the detection of short video content.

[isaduan]

Gender is one dimension of representational harm in online images. There are other dimensions, e.g. class bias, bias against immigrants, etc. I wonder whether you have any thought on how can we possibly expose all these dimensions of biases in auditing online images?

[JerryCG]

Dear Doug, I am looking forward to your thrilling speech. Based on the research findings, do you think a censoring policy can help to reduce the gender bias by eliminating photos which contain harmful elements for female?

[shaangao]

This is a really important finding! With model-editing techniques enabling the removal of biased information representation in generative AI models, would you expect images to contribute to the reduction of gender bias sometime in the future?

[yuanninghuang]

I'm intrigued by your research on the amplification of gender bias in online images. It reminds me of studies that have looked at reducing gender bias in AI language models like GPT. Have you considered similar strategies for platforms like Instagram and TikTok, where visuals are central? I'm curious to hear about any potential approaches or factors you've identified that could help mitigate this bias in such image-dominant environments.

[iefis]

Thank you for presenting this interesting topic! I have a question about the measure of gender bias in online texts as the comparison to images. In the paper, gender bias in texts is measured by leveraging word embedding models. I am wondering how human perception of the gender dimension in texts might be different from the results from natural language processing and how this might bias the measured difference between gender bias in images and texts.

[Brian-W00]

How can computational social science evolve to address and mitigate the amplification of gender biases, especially considering the increasing prevalence of image-centric platforms and AI-generated content in shaping online culture?

[jinyz1220]

Your work is intriguing! I'm wondering whether you have considered using different methods to measure bias. For example, after the treatment, show them a picture of a female doctor vs. a male doctor and ask them how experienced this doctor is. or a picture of a female vs male entrepreneur, and ask them about the income level of the entrepreneur. The effect of the treatment on gender bias might be amplified and it is easier to be detected.

[YutaoHeOVO]

Hi Professor,

Thank you so much for sharing your research. It is really interesting.

Yet there might be some questions or extensions related to this. Maybe in English, images might induce more gender bias. But what about in languages where there are genders for words? (For instance, French.) Can this result be extended to a broader context about the comparison between text and image, or it is limited in the realm of English?

And another thing is that maybe pictures comparing with words can convey more biological features themselves. (Michal Kosinski has already published a lot of papers on using facial images to detect sexual orientation, it will be no surprise that it can directly tell us something about a person's gender.) Moreover, the gender bias that images illustrate might not be bias but also part of people's mechanism of identifying gender. Is it possible that it is the natural result of people relying on 'gender bias' to identify or define gender?

[bairr1208]

Thank you for sharing this intriguing study with us! I am wondering how the results of the study, showing a stronger prevalence of gender bias in online images compared to text, reflect on the current content moderation policies of major online platforms. And what do you think are some applicable changes that might be necessary to address this imbalance in the representation of genders?

[nourabdelbaki]

I found this paper a very interesting read. I wonder if the gender biases persist across cultures and languages, which may be difficult to judge because of sampling (limited posting of pictures but also possible 'cross-contamination' if it is user generated and they are using ready-available pictures found on Google/the internet), I would think the biases persist, but it would be interesting to explore it in light of specific cultural context and semantic characteristics of languages (like Arabic and French that has an associated/assumed base gender to each word), do images in these context still carry a stronger prevalence of bias or is it more consistent bias in both images and text?

Also, I see a lot of comments/questions about possible censorship or regulating mechanisms to improve this bias, but would it actually help improve the bias or is it just camouflaging the presence of these biases?

[kunkunz111]

Thanks for sharing! In your study on gender bias in online images, how can AI be utilized to mitigate these biases, and what measures should be implemented in AI model development to ensure a more balanced gender representation?

[HamsterradYC]

Thanks for sharing your research with us , the result showed that online images amplify gender bias more significantly than text. Considering the increasing use of AI in generating images and the potential biases in these AI models, how do you foresee the impact of AI-generated images on societal perceptions of gender roles?

[ethanjkoz]

This was an interesting read! I appreciate the attention paid to articulating that using MTurk workers to classify gender was meant not as a way to measure self identification but rather gender perceptions. As you mention in the article, more work ought to be done to elucidate other social biases present in image related data. Pertaining to race, what would expect to find regarding under and over representation of specific races/ethnicities by occupations in images?

[erikaz1]

1. To what extent may bias in the interpretation of images reflect a kind of collective consciousness/collective memory or expectation of gender roles (especially when subsequently prompted to judge whether the image leans more male or female)? (Yutao's last sentence above rewords the essence of this Q in a better way.)

2. I was also wondering: does the wording of the surveys and questions regarding gender affect how we perceive the concept and gender bias? For instance, in the Testing Psychological Effects section, it was stated that "each participant was asked to rate which gender they most associate with the occupation being described". What would have happened if we had added an additional option not on the male-female scale, such as "I do not immediately associate this occupation with a particular gender"? If the scale mentioned in the text had been a sliding scale from -1 to 1, in which case 0 (no skew) would have been an option, what then would a fractional value between 0-1 or -1-0 mean?

[Daniela-miaut]

Is it possible that the Internet is prone to amplify bias in general because those biased content requires less cognitive efforts as they comply with people's stereotype? I am also interested in the findings that images amplify people's implicit bias. Could it possibly be explained by cognitive efforts? Should we raise consciousness in preventing the pitfall of cognitive ease in order to combat the exacerbated bias on the Internet?

[lim1an]

How might the findings of this study on gender bias in online images influence the development of content moderation policies and the integration of AI algorithms in major online platforms to address and mitigate gender biases effectively?

[Caojie2001]

It's an interesting research about gender bias in online images. In the conclusion part, you mentioned that a direction for future research is to investigate the social and algorithmic processes contributing to biz in online images. I wonder if you have any suggestion for dealing with this bias through computational methods.

[YucanLei]

The phenomena mentioned is quite surprising, but also make sense. It is no surprise that the biases from texts could easily extend to images and biases caused by images are much more influential. Yet how could we be certain about this? How could we be certain that biases have became more prevalent? I look forward to this presentation to see how the idea is developed and extended.

[schen115]

Thank you for your sharing about this insightful findings! I was wondering if there is any technical strategies or interventions could be implemented to mitigate this gender bias in online visual content?

[oliang2000]

This is a fascinating paper:) I'm curious about the impact of online search recommendation algorithms on the observed phenomenon. The conclusion hints at this; how could we test it?

[ana-yurt]

Thanks for sharing your research! In addition to priming gender bias, I wonder if there are any findings on the specific mechanisms through which online images 'gather' more bias.

[Ry-Wu]

Thank you so much for sharing the interesting insights! I'm wondering if there are some social implications your study brings, such as how to reduce the gender bias in online images. Looking forward to your speech.

[xinyi030]

Given the significant findings of your study regarding gender bias in online images, what are your thoughts on the role of education and awareness programs in addressing this issue? Do you believe that incorporating digital literacy and critical media consumption skills into educational curricula could help in mitigating the impact of gender biases perpetuated through online images? How might such programs be designed to effectively educate both younger and older internet users about recognizing and challenging these biases?

[mingxuan-he]

Thank you for the great presentation! Based on your findings, could you comment on how can existing technologies be adapted or improved to more effectively distinguish between inherent bias in the data versus e.g. culturally and contextually relevant representations?

[bhavyapan]

Very interesting paper! I echo some of the cultural and language-related factors that could have an impact on and would be worth exploring within the context of the paper's line of reasoning. What are some of the ways in which researchers and practitioners are working on these potential biases?

[secorey]

Hi Professor Guilbeault, Thank you for presenting your research. I'm wondering about some of the implications you draw to social media, and how your findings point to a potential danger in how people are being exposed to gender bias. While I think this is an understandable concern, I think there must be further research examining images in social media specifically. With Google images, you normally get a "stock image" when searching for broad categories like "doctor." Many of those images may not contain a real doctor at all. Furthermore, I'm not sure how often people look to Google images to get information on these concepts. Rather, many of these influences come from social media platforms where concepts are communicated in a much more natural context.

[Jessieliao2001]

Thank you professor Guilbeault, my question is: How can digital platforms and content creators effectively mitigate the amplification of gender bias in online images, considering the findings that gender bias is more prevalent in images than in text and that this bias is influencing public perception of various social categories?

[Yuxin-Ji]

Hi Professor Guilbeault, Thanks for sharing your research! I am curious about your data collection method, where you retrieve the top 100 images in each category. Though this composes of a large dataset, it is far from the total number of images online. I wonder how truthfully do these data represent the real online dataset and thus how well could the result represent the online dynamics?

[yiang-li]

I really learnt a lot on the methods and outcomes of this research. Considering the findings of this paper, how do you think the increasing use of AI-generated images, which often inherit biases from training data like public images from platforms like Google and Wikipedia, will impact the future of gender representation and bias in online content?

[lbitsiko]

You mention that findings are robust to certain linguistic features, I was wondering if you could further expand on this topic.

[Zhuojun1]

Thank you for sharing! In conducting laboratory experiments, how do you ensure the reliability and comparability of participants' perceptions and reactions to images and text obtained through search engines in the real world? Have you considered the impact of the experimental environment on participants and the potential differences in how they interact with online images and text in their daily lives?

[PaulaTepkham]

Thank you for those four papers that will be shared in this workshop. I am every interesting about the gender difference in academia, and I found that the "The Gender Gap in Scholarly Self-Promotion on Social Media" paper is pin point one dimension of gender-based in academia. Since self-promotion seems to be a typical practice for well-known research, but this practice seems to be occupied less for female researchers. What would be the reason/ mechanism of this? How can we help the female researcher to be able to do self-promotion better in the academia?

[KekunH]

Hi Professor Guilbeault, How can we address gender bias in the development of image-labeling models used by big tech companies, considering the inherent biases in both unsupervised learning and human-encoded models? Additionally, do you believe that integrating digital literacy and critical media consumption skills into educational curricula could effectively mitigate the impact of gender biases in online images, and how should such programs be designed to educate both younger and older internet users in recognizing and challenging these biases?

[cty20010831]

After reading the abstract, one question I have is that how did the authors quantify gender bias here? Specifically, how is gender bias quantified in images versus text, and what metrics are most effective for this comparison?