“It Feels Like Being Locked in A Cage”: Understanding Blind or Low Vision Streamers’ Perceptions of Content Curation Algorithms

Due to the surge of live-streaming, an increasing number of studies investigated users’ live-streaming practices on social media platforms. Prior research indicated that people from a variety of professions actively streamed online, such as professional streamers (Woodcock and Johnson, 2019; Lu et al., 2018) , practitioners safeguarding Intangible Cultural Heritage (ICH) (Lu et al., 2019b) , programmers (Chen et al., 2021b; Hammad et al., 2021) , university instructors (Chen et al., 2021a; Hammad et al., 2021) , students preparing for specific tests (Wang and Zhang, 2021) , community leaders and reporters (Dougherty, 2011) . For instance, Chen et al.(Chen et al., 2021b) indicated that programmers chose live-streaming to share programming knowledge and skills. Some studies also examined a wide range of live-streaming practices, including pan entertainment (Lu et al., 2018) , goods selling (Lu et al., 2018; Hu and Chaudhry, 2020) , online education (Lu et al., 2018; Chen et al., 2021b; Chen et al., 2021a; Fonseca and Barbosa, 2021; Hammad et al., 2021) , personal story sharing (Lu et al., 2018) , intangible culture promotion (Lu et al., 2018, 2019b), eating (Anjani et al., 2020), outdoor activities (e.g. traveling, adventures) (Lu et al., 2019a; Tang et al., 2016), and gaming (Tang et al., 2016; Woodcock and Johnson, 2019). Lu et al.(Lu et al., 2019b) found that streamers who promoted ICH-related content leveraged various streaming structures to showcase ICH practices. For example, ICH streamers held “Questions and Answers” sessions, interviewed experts, asked learners to perform, and delivered tutorials on basic knowledge. Work by Lu et al. (Lu et al., 2019a) analyzed outdoor live-streaming in China.  This work showed that in livestreams, streamers broadcasted outdoor activities or traveled in different environments (e.g., hiking, fishing, and hunting). In addition, they shared spontaneous and unpredictable interactions with strangers (e.g., passersby on streets and Didi (Uber) drivers).

Some prior work presented that people with disabilities actively streamed. Anderson and Johnson (Anderson and Johnson, 2021) studied how people with physical disabilities streamed for changing disabled viewers’ negative mindsets about disabilities. In addition, work by Johnson (Johnson, 2019) presented the enjoyable experiences and challenges of video game streamers with physical disabilities, ADHD (Attention Deficit Hyperactivity Disorder), and dyslexia on Twitch. Moreover, some prior work focused on economic and employment opportunities of live-streaming for disabled people  (Johnson, 2019, 2020; Qu, 2020). Recently, Jun et al. (Jun et al., 2021) explored the motivations, practices, and challenges of streamers with visual impairments on Youtube and Twitch. Although this work uncovered streamers’ motivations (e.g., achieving personal goals, delivering messages, overcoming limitations of recorded videos, getting inspired by other streamers) and challenges (e.g., multitasking, screen reader related issues), it did not investigate BLV streamers’ experiences with content-curation algorithms adopted by the platforms. However, recent research suggested that users from marginalized groups perceived live-streaming platforms may have algorithmic biases against minority communities, such as race and ethnicity minorities and the LGBT community (Karizat et al., 2021). Therefore, in this work, we take a first step to investigate BLV streamers’ experiences with content-curation algorithms adopted by live-streaming platforms.

Content creators of video-sharing platforms were aware that algorithms were a crucial factor, which influenced whether their content could be well-known among passersby  (Bucher, 2017; Trewin et al., 2019; Pedersen, 2019; Wu et al., 2019). Prior research  (Simpson and Semaan, 2021; Lines, 2020; Trewin et al., 2019) has presented how algorithms excluded the videos created by people from marginalized groups, such as people of color and those who came from LGBTQ groups or lower social-economic class. For example, several studies indicated algorithmic bias for sexual orientation or other gender issues. Simpson et al. (Simpson and Semaan, 2021) described that users from the LGBT+ community believed that the Tiktok algorithms censored or removed the content created by the LGBT+ group. Haimson et al. (Haimson et al., 2021) reported that users perceived that the algorithms of Tumblr sometimes oppressed the content from trans through classifying them into adult content. Also, recently, some anecdotes (Lines, 2020; Trewin et al., 2019; Bozdag, 2013; Robertson, 2019; Biddle et al., 2020; Sam Biddle, 2021; Melonio, 2020) revealed that in the process of content evaluation, Tiktok limited the popularity of videos created by people with, for example, disabilities, facial disfigurements, Down syndrome, or autism. Tiktok prevented their content from reaching non-disabled people’s ”For you” feed.

The perceived algorithmic biases and suppression enabled content creators to come up with some coping strategies. For example, Simpson et al. (Simpson and Semaan, 2021) showed that content creators recapped or reposted the content removed by the Tiktok algorithms to push back against the perceived suppression. According to work by Karizat (Karizat et al., 2021), content creators believed that the Tiktok algorithms suppressed the content related to marginalized social identities. Content creators resisted the Tiktok algorithms through individual actions (e.g., sharing the content perceived to be suppressed ), collective actions (e.g., collectively commenting and liking the videos silenced by algorithms), and altering the performance of content.

Despite increasing attention to algorithmic fairness for marginalized groups, whether BLV users perceive their contents are impacted unjustly by the algorithms of live-stream platforms remains unclear, which motivated our work.

Table 1 shows participants’ demographic information, including age, gender, vision condition, and occupation. Eleven (N=11) participants were totally blind, and eight (N=8) were low vision. Seven were female, and twelve were male. Participants were between 18 and 42 years old ($M=31,SD=4$). It was worth noting that most of our participants were not tech-savvy, who mainly utilized the screen readers and Voice-over to access live-streaming platforms.

Our participants were recruited by sending direct messages on Douyin or sending messages to their WeChat posted in their profiles on Douyin or snowball sampling. We searched on Douyin, using BLV-related keywords (e.g., blind, visual impairment, promote accessibility, etc.), and sent messages directly to the eligible streamers. We contacted 82 streamers directly in total, and 6 were willing to accept our interview. Through snowball sampling, we recruited another 13 participants. It was worthy to note that the BLV participants we recruited had diverse educational levels, ranging from high school degree or less to bachelor’s degree.

We watched and engaged in their live-streaming two weeks before the interviews to build rapport and trust with our participants. We also watched their short videos, “like”, and left comments on the videos. At the same time, their live-streamings and short videos were recorded and saved.

We observed their live-streaming practices (e.g., topics chosen, hashtags added, interacting approaches, etc.). Meanwhile, we checked their profiles to gain more useful information the streamers disclosed. This process helped us prepare for semi-structured interviews.

We conducted the interviews between June 22 and August 12 in 2021. All interviews lasted approximately 50 minutes to 3 hours via WeChat voice call, and participants were provided with 100 CNY (about 16 US dollars) after interviewing. All interviews were conducted in the participants’ native language (i.e., Mandarin Chinese) and audio recorded.

We aimed to understand participants’ perceptions of algorithms employed by living-stream platforms. We first their general understanding of algorithms. We then asked them how algorithms worked on living-stream platforms. We further asked about their perceptions of the factors related to the algorithmic evaluation of content, the perceived algorithm-related challenges they experienced, and their coping strategies.

Our data included the audio recordings of the interviews, recordings of their live-streaming, short videos they posted, and research field notes. We first transcribed the audio recordings verbatim. Then two native Mandarin-speaking authors coded the transcripts independently using an open-coding approach (Corbin and Strauss, 2014). In the regular weekly meetings, the two coders discussed the codes (e.g., definitions and example quotes) and resolved conflicts to consolidate the code book. If there were any unresolved conflicts, the rest of the authors joined the discussion to gain a consensus. Afterward, the research team performed affinity diagramming (Hartson and Pyla, 2012) to group the codes into clusters based on their semantic similarity, and we identified the themes of the clusters. These themes and codes, along with representative quotes, form the structure of our findings, which will be reported in the next section.

Participants perceived the algorithms to own a great deal of power, assessing whether their content could pass the “examination” from Douyin. For instance, P17 expressed her perceptions of which factors of a user’s content played essential roles in influencing the content evaluation from the algorithms, which was informed by the several articles on new media and word of mouth from Douyin users.

“There were many metrics in terms of content evaluation, such as the completion rate of your videos, the time users spent on your live room, the number of video views, the ‘likes’ and comments, as well as the virtual gifts you received. It was difficult for us to fit into these rules due to our visual disability.” (P17)

Participants felt that accessibility challenges with Douyin’s user interfaces resulted in disadvantages under algorithmic evaluation. The accessibility challenges were related to difficulties in ensuring visual effect filters, obstacles to keeping active in livestreams, and barriers to video editing.

Participants suffered from two main perceived algorithmic suppression.

Failure in Reaching Sighted Audiences. Participants generally believed that their content was mainly “locked” within the BLV community by the algorithms and was barely amplified to the sighted audience. For example, P7, a blind massage shop owner who wanted to promote the business through Douyin, noticed that most viewers were BLV users, which did not meet his expectations because he looked forward to attracting sighted customers. As he explained, “Whenever I observed the video view counts provided by the system, I found few sighted users while most were BLV users [based on his experience that BLV viewers’ IDs often contain terms relating to ‘A person with visual impairments’ or ‘a blind person’). Also, the comments were mainly from people with visual impairments.” The perception of BLV streamers’ content not reaching sighted users was echoed by P18, who often chatted with sighted streamers at their live-streaming sessions. As she said, “When they first met me (at their live room), most of them felt surprised and said, ‘wow, you are the first BLV user we encountered these years! We never saw any content created by BLV community members before.”’ P18 further noted, “They (sighted streamers) later told me that they eventually saw the BLV-related content only if they intentionally searched the keywords at Douyin, such as blind streamers.”

Participants’ comments spoke to a belief that the algorithms automatically detected their content and account interactions. The algorithms were perceived to mark the account network from followers or viewers who re-posted, liked, or left comments. As a result, they believed that once the algorithms found users with similar networks and behaviors, those users would be confined and categorized as a group with a dedicated user profile. BLV users, under such circumstances, could hardly send the content to massive sighted viewers. For instance, P14, a grocery store owner who wanted to promote goods to sighted people through Douyin while mostly followed by BLV users, elaborated as follows:

“My online business plan failed…one of my videos were viewed and reposted by several BLV users, which induced more BLV users to follow my account. Then the algorithms may consistently amplify my content to BLV users. As a result, this process kept looping, which enhanced the algorithms’ belief that it made the right decision to recommend my content only within the BLV community.”

Participants’ remarks expressed a perception that the more BLV-related information an account had, the more likely the account was identified as and clustered into the BLV user group. Hence, BLV users may find it more challenging to promote content to sighted audiences.

Suppressed Popularity within BLV Community. Participants’ comments highlighted a belief that Douyin’s algorithms limited the popularity of BLV-related content within the BLV community. For example, P16, a public-spirited blind streamer who aimed at popularizing how to manipulate current assistive technology for BLV users through Douyin, perceived that the algorithms suppressed his live-streaming. As he explained, “It was last December, the number of my viewers dropped suddenly.” P19 went on to note that he and other streamers found they could not purchase the Dou+ service, which could help boost the popularity of live-streaming.

“A pop-up window showed that the themes of our content were not suitable to be boosted. You know, once you could not buy such a service, then the popularity of the account was permanently suppressed.”

Other participants expressed that not only their live-streaming but also the short videos they created were not valued by the algorithm. P7, a blind content creator who took videos to promote available assistive technology for people with visual impairments, shared an example of how Douyin held up his video. As he said,

“I had a video about how to quit the screen readers, I tried to upload it to Douyin several times, but the system just told me I failed to upload the video because the platform was overloaded. However, when I uploaded a non-BLV video, it went through”

This perception was echoed by P6, a user with low vision, who observed that compared to sighted users’ videos, the window that showed the “not interested button” popped up at BLV content creators’ videos with higher frequency. As he explained,

“One of my BLV classmates also noticed that such an indicative window seldom showed up in sighted users’ content while it often popped up on BLV users’ videos. The algorithms may not value BLV-related videos and thus provided the (not interested) button to suppress our videos further.”

Participants’ comments about the “not interested button” expressed a perception that even for BLV audiences, Douyin’s content promotion algorithms may consider the BLV-related videos they viewed to be of low quality and not worthy of being promoted within the BLV community. It is worth noting that participants also perceived that sighted people most likely encountered such a “not interested button” when they saw BLV-related content on their For You Page. Participants believed that if sighted viewers clicked the button, the algorithm would likely not recommend their live-streams/videos to sighted populations (as Section 4.1.4 reported).

Our findings show that BLV participants took various actions to mitigate the perceived algorithmic suppression. Participants adopted the following strategies related to their BLV status: Hiding BLV Identity, Creating Counter-intuitive Content for Sighted Viewers,  Leveraging Occupational Convenience to Attract Sighted Followers, and Collectively Ignoring The “Not Interested” Button on BLV-related Videos.

One common recurring difficulty was adjusting the shooting angles properly to ensure that a BLV streamer’s face and body were within the camera frame when streaming due to visual impairment. Current face-tracking technology shows the promise to address this challenge, which could detect and locate objects or faces in nearly real-time (Wagner et al., 2009; Lei et al., 2011; Naik and Rathna, 2015; Ataer-Cansizoglu and Taguchi, 2016) on low-end devices like mobile phones (Chen et al., 2009). However, BLV content creators’ streaming behaviors complicated the utility of face-tracking technology. Some BLV streamers did not want to show their eyes and thus always wore sunglasses, which made it challenging to track their face since current face-tracking technologies rely on capturing the whole parts of a human face  (Anwarul and Dahiya, 2020; Kaur et al., 2020). Also, some BLV participants streamed in nearly total darkness because they were not sensitive to the indoor light conditions due to visual impairments, which deteriorated the accuracy of face-tracking technologies  (Rahman et al., 2020). We suggest that algorithm designers may train the face recognition model with as many BLV streamers’ videos as possible. Additionally, when applying face-recognition technology in BLV users’ live-streaming, it is essential to take into consideration how to avoid leaking private information. For example, the faces of the BLV streamer’s family members or the photos of faces (Gurari et al., 2019) in the streamer’s room might be accidentally captured. Prior research indicated that privacy was a core value in system design (Friedman et al., 2013) and people with visual impairments were concerned about their privacy when using technologies (Hayes et al., 2019; Dosono et al., 2015; Briotto Faustino and Girouard, 2018). To protect users’ privacy, Zhou et al. developed a technology to blur irrelevant people’s faces during live-streaming (Zhou and Pun, 2021). However, many questions remain unresolved. For example, what visual content, in addition to faces, do BLV streamers perceive as private? Would failures of blurring different privacy-sensitive content cause equal concerns to BLV streamers?

Some participants with eyeball atrophy or who were not good at dressing up also mentioned that they encountered barriers in image management. They worried that their appearance would leave a negative impression on the audience. Hence, some of them tended to wear sunglasses and even turned off the camera directly during streaming. This raises an interesting question of how to help BLV streamers better manage images without compromising viewers’ experiences? A possible solution could be to explore the use of virtual idols on short-video platforms that support live-streaming (Lu et al., 2021; Black, 2012; Freeman et al., 2020; Yin, 2018). A virtual idol is a 2D or 3D animated virtual avatar with the voice of a human, which could be operated by individuals or agencies. An increasing number of streamers already started to use virtual avatars in their live-streams, such as voice actors, amateur and professional musicians, and artists (Lu et al., 2021). One streamer named CodeMico on Twitch used virtual avatars to achieve visual effects, such as blowing herself up and flying into space in the live-streaming, that would otherwise be impossible to perform by humans (真探, 2022). Moreover, live-streams with virtual avatars had a broad audience in China (时代周报猛犸工作室, 2022; Lu et al., 2021). By June 2021, over 30,000 active streamers conducted live-streams with virtual avatars on Bilibili and gained over 560 million Danmu (i.e., scrolling comments) (时代周报猛犸工作室, 2022). Recently, some content creators with disabilities also began to use virtual avatars. For example, a content creator with paralysis recorded his videos with virtual avatars rather than his real appearance. He commented that without showing his appearance, he could avoid being questioned for sad-fishing with his disability (干物哥の小宇, 2021). Thus, we believe that if streamers or video creators with disabilities do not want to show their faces, they could use virtual avatars. However, it remains an open challenge of how to inform BLV streamers about the visual appearances of, and more importantly, the audience’s reactions to, their virtual avatars. Towards this direction, designers should consider providing explicit audio descriptions of the virtual avatars’ features such as gender, skin tone, hairstyles, makeup, accessories, clothing, and ability status (Bennett et al., 2021) for BLV streamers. It is also essential to inform BLV streamers promptly about their viewers’ reactions and preferences of their avatars.

In addition, when it comes to enabling BLV streamers to create and craft their virtual avatars, designing BLV user-friendly audio instructions and avatar building interfaces remains unclear. Thus, it is imperative for designers to consider lowering the burden of designing avatars while ensuring design quality. Furthermore, given that some BLV content creators preferred not to show their BLV identities in live-streaming or videos, a potential tension between BLV identity and avatar’s appearance may emerge, which calls into questions: Is the BLV streamer willing to show his or her BLV identity through a virtual avatar? What are the viewers’ reactions to the avatar that discloses BLV identity? How should BLV streamers balance maintaining disability identity and keeping the avatar attractive? It is worth investigating different stakeholders’ perceptions of the avatar design of BLV streamers.

Accessible Filters Editing Experience.  While prior work (Bennett et al., 2018) showed that Instagram users with low vision could apply filters to photos to achieve a specific aesthetic effect, it did not mention the usage of visual filters in live-streaming among BLV streamers. Our work filled this gap and revealed that BLV streamers, especially those who were blind, were unable to apply visual filters to short videos and live-streaming. For instance, some functional buttons of filters were not labeled, such as the percentage of luminance. Participants also emphasized that they could not picture the visual effects the filters may achieve. Not only is it essential to ensure all UI buttons have been labeled correctly, but it is also important to provide audio instructions to help BLV streamers develop a prior idea of the functioning of the filters during live streaming. However, as people’s perceptions of the aesthetic of visual content are personal and subjective (Zhao et al., 2017), it can be challenging for BLV streamers to judge whether a visual filter would meet their expectations. Thus, it is important for screen reader designers to investigate ways to make computer-generated descriptions of filters match their expectations of such filters. Additionally, designers of screen readers need to personalize audio descriptions to different kinds of visually impaired users in order to improve intelligibility. For instance, an audio description comprehensible to people with non-congenital visual impairments may be incomprehensible to people with congenital visual impairments. Similarly, designers also need to personalize assistive editing features for streamers with differing levels of vision impairment. Unlike low-vision people, participants with total blindness noted that navigating the filter interfaces and locating the buttons were challenging. One possible solution is to use voice commands or gesture inputs (e.g., tapping, swiping, or even eyelid gestures) for navigation (Li et al., 2020; Vtyurina et al., 2019; Fan et al., 2020a; Li et al., 2020; Fan et al., 2021; Li et al., 2017).

Accessible Video Editing Experience. Another challenge for assistive editing design is to provide audio descriptions on mobile editing tools (e.g., Jianying, a software that allows users to edit videos on mobile phones and imports videos to Douyin). Although previous works introduced algorithms to generate video descriptions (Johnson et al., 2016; Krishna et al., 2017; Wang et al., 2021), such algorithms were designed or trained to recognize videos taken by sighted people. The quality and content of videos taken by people with visual impairments, however, may differ from those taken by sighted users. For example, a video shot by a BLV streamer may result in inaccurate or even absurd descriptions for current algorithms due to factors, such as jitters or being out of focus. Therefore, common video quality issues should be carefully considered when designing algorithms to generate audio descriptions of videos shot by BLV streamers. Designers should consider collecting diverse sets of videos shot by BLV streamers in different contexts and training such algorithms on those videos.

Further, as automatic generated audio descriptions may still contain errors, post-verification is needed. For example, previous research mentioned that people with visual impairments tend to ask sighted friends or family members to verify the accuracy of computer-generated descriptions of photo, but this also imposes a burden on those sighted people (Wu et al., 2017; Zhao et al., 2017). This issue may be exacerbated because of the extra efforts required to evaluate video descriptions due to longer length. Therefore, future work should investigate more effective way to validate the quality of video descriptions.

Our finding regarding BLV content creators’ perceived algorithmic barriers mirrored Karizat et al.’s identity strainer theory (Karizat et al., 2021), which expressed that the algorithms acted as a strainer to marginalize a variety of social identities, including ability status. However, in  (Karizat et al., 2021), there was only one participant who mentioned his or her observation that Tiktok valued the content of able-bodied streamers but oppressed the content created by people with disabilities. We extended Karizat et al. ’s work (Karizat et al., 2021) by providing an in-depth investigation of BLV content creators’ perceived challenges from Douyin’s algorithms. Participants believed that they were vulnerable under the evaluation metric of the algorithms due to the harmful trolling behaviors and the lack of visual ability to ensure content’s visual appearance, timely interaction with audiences during streaming, and quality of video editing. Participants perceived that content created by BLV content creators was seldom amplified to sighted audiences. Instead, they believed that algorithms limited the popularity of their content within the BLV community.

Participants adopted BLV status-related strategies to mitigate perceived algorithmic biases, including 1) hiding BLV identity; 2) creating counter-intuitive content; 3) leveraging occupational convenience in massage shops to attract sighted followers; ; and 4) collectively ignoring the “Not Interested” button on BLV-related videos. Among these strategies, “hiding BLV identity” may lead to conflicts with BLV identities. For instance, in order to “trick” Douyin’s algorithms to promote their streaming contents to more sighted viewers, participants avoided disclosing any BLV-related information in their accounts or shied away from any engagement with any BLV-related content and community. Indeed, this tactic may help them reach a wider audience of sighted people, but it would unfortunately force them to hide themselves in non-BLV algorithmic identities (Cheney-Lippold, 2011). Such a strategy is contrary to those in previous works  (Karizat et al., 2021; Simpson and Semaan, 2021), in which marginalized users tamed the algorithms to make their algorithmic identities (Cheney-Lippold, 2011) and their self-identities further closer together. Future research may explore whether content creators from other marginalized groups may use their own status-related mitigation strategies. It may also be fruitful to explore whether there exist tensions when trying to preserve identities and tame algorithms in order reach a certain targeted audience simultaneously. Researchers could also explore how people with visual impairments may resolve these tensions?

Participants also adopted strategies that may not uniquely belong to the BLV community to mitigate perceived algorithmic biases, including: 1) tagging locations at popular places; 2) directing more comments from other platforms; 3) actively interacting with target users; 4)peer support; 5) sticking to trending topics and 6) negotiating with the Douyin authority. Since these mitigation strategies are non-BLV status-related (e.g., tagging location and direct comments), they may be applicable to other marginalized content creators (e.g., people of color, people of LGBTQ community, people with other types of disability, or those who are socioeconomically disadvantaged) who want to promote their contents to a broader community but perceive their contents being “locked” within a confined group.

Next, we present design suggestions for creating inclusive algorithmic experiences for BLV content creators.

Accessible Explanation Interface. Prior research (Alvarado and Waern, 2018) defined algorithmic experience (AX) as “an analytic tool for approaching a user-centered perspective on algorithms, how users perceive them, and how to design better experiences with them”, which included five themes for improving AX in the context of social media: algorithmic profiling transparency, algorithmic profiling management, algorithmic awareness, algorithmic user-control, and selective algorithmic remembering. We suggest that algorithmic profiling management of AX may be applied to create better AX for BLV content creators, which enables them to modify and tune their profile created by algorithms (i.e., how algorithms think about them). For example, BLV content creators, who want to promote their live-streams or videos to more sighted audiences, could adjust the factors contributing to failure in reaching sighted audiences in their algorithmic profiles, making such profiles aligned with their expectations. However, this raises an open question of how to effectively and sufficiently communicate complex internal operations of algorithms to people with visual impairments. Current research on Human-Centered AI leveraged interactive visualizations to allow stakeholders to explore the data and audit the algorithms synchronously and asynchronously (Cheng et al., 2019, 2021; Fan et al., 2020b, 2022). However, such visually-demanding interfaces are inaccessible to people with visual impairments. Therefore, it is worth exploring ways to design accessible explainable AI interfaces for people with visual impairments, such as combining the potentials of accessible visualization (Lundgard et al., 2019) and dynamic shape-shifting user interfaces (Siu et al., 2018).

Accessible Participatory algorithms Design. Our study unpacked insights into BLV streamers’ perceived algorithmic biases against their BLV identity, which can be difficult to be effectively incorporated into algorithmic design (Zhu et al., 2018). Towards this challenge, participatory algorithm design may be a promising direction to explore (Lee et al., 2019). Participatory algorithm design involves stakeholders building the computational models that represent their motivations, values, and goals. However, it remains an open question of how to leverage such a participatory algorithm design. For example, how to ensure the whole model training process accessible to BLV participants? How to deal with the different values and opinions of participants with different levels of visual impairments (e.g., blind and low-vision participants)? How to take the social-economical conditions of BLV participants into consideration to improve the diversity of the design of the participatory algorithms?

Social Awareness. While it is necessary to address the algorithmic challenges by leveraging technical solutions, socio-technical solutions are also crucial. For example, the misperceptions and trolling behaviors from sighted audiences may negatively influence the input into the video-sharing platform’s algorithms feedback loop, enabling algorithms to oppress content created by people with visual impairments. Policymakers should make continued effort to mitigate the general public’s stigma towards people with disabilities  (Li et al., 2021), addressing the knowledge gap  (Li et al., 2021) about disability from mainstream society, and strengthening the inclusive awareness of the public. It is also imperative for video-sharing platforms (e.g., Douyin) to create an inclusive online community for content creators with BLV or other disabilities. Platforms may help promote the content regarding mitigating misperceptions about people with visual impairments or promoting accessibility knowledge to a broader audience, outlining respect to people with different ability levels as the community rules, combating the toxic trolling behaviors towards BLV streamers. In addition, participants noted a lack of understanding about how algorithms work to influence content within the less-educated BLV content creators or those located in rural areas. Informed by (Alvarado and Waern, 2018) ’s framework for AX, we suggest that in the future, video-sharing platforms (e.g., Douyin) should strive for improving the algorithmic awareness of BLV content creators who are vulnerable in education or social-economic status, telling them explicitly how algorithms operate and what sorts of user behaviors affect the feedback loop of the algorithms. Furthermore, it is also necessary for algorithm designers to consider and respect the diverse needs of people with different abilities instead of just focusing on the algorithmic performance.

Algorithmic Identity. The perceived suppression from the algorithms shaped participants’ behavior to break the feedback loop (Rader and Gray, 2015) to promote their content to their target audiences, for example, by manipulating the input to the Douyin’s algorithms individually or collectively. While prior work (Karizat et al., 2021; Wu et al., 2019) noted that influencing the input of the feedback loop had the potential to tilt the algorithms, we found that a tension emerged when the BLV users wanted to promote their content to both the sighted and BVI community because the strategies they employed for reaching these two communities were paradoxical at times. For example, the strategies that participants utilized to promote content to sighted people, such as hiding BLV identity, increasing view and like count of sighted people, attracting sighted followers, resulting in continuously influencing the input into algorithms, which might gradually enhance the algorithms’ belief that their content should be recommended to more sighted audiences. This strategy worked at the expense of abandoning the opportunities to make their content show at BLV users’ FYP (for you page). This tension was echoed by P9, noting “You could only choose one side (sighted or BLV community) under the current rules of the algorithms. ” The aspect of algorithmic user-control in Alvarado and Waern’s AX framework (Alvarado and Waern, 2018) showed the potential to address this tension, which empowered users to control the algorithms directly. For instance, algorithm designers can design a feature that enables BLV content creators to craft and switch between different algorithmic profiles (i.e., promoting content to different categories of viewers ) to click a switch button on their profile pages directly and conveniently.