"It's almost like they're trying to hide it": How User-Provided Image Descriptions Have Failed to Make Twitter Accessible
Cole Gleason, Patrick Carrington, Cameron Cassidy, Meredith Ringel Morris, Kris M. Kitani, Jeffrey P. Bigham · 2019 · The World Wide Web Conference · doi:10.1145/3308558.3313605
Summary
This paper presents the first large-scale empirical study of how Twitter's opt-in image description feature (launched in 2016) has affected the accessibility of images on the platform. The researchers analyzed 1.09 million tweets with images collected over five days in June 2018, finding that only 0.1% contained image descriptions — a strikingly low adoption rate two years after the feature launched. The study comprised three components: a quantitative analysis of alt text prevalence across Twitter generally, an examination of the timelines of 94 self-identified blind Twitter users to understand their experienced accessibility, and semi-structured interviews with 20 sighted Twitter users who had written image descriptions to understand their motivations, practices, and barriers. The feature requires users to proactively enable it through Settings > Accessibility, and even then only presents a small "Add description" prompt when composing tweets. The researchers also examined the 50 most popular Twitter accounts and 577 U.S. Congressional accounts, finding that only 3 of the top 50 and 42 of the 577 government accounts had ever used the feature.
Key findings
Of 1.09 million tweets with images, only 0.1% (1,144 tweets) included image descriptions. Among original photo tweets (excluding retweets), the rate was even lower at 0.06%. For blind users' recreated timelines, 18.4% of tweets contained photos and 4.6% of those photos had descriptions — an order of magnitude better than Twitter overall, but still largely inaccessible. Among the 20 interviewed description authors, 12 explicitly intended to describe every image but achieved only about 50% success rate. The most common reasons for not adding descriptions were: forgetting (especially when posting quickly), time constraints, not knowing what to include, and the feature being buried in settings. When human-written descriptions were evaluated (excluding bots), 83% were rated "Good" or "Great" quality — demonstrating that people who do write descriptions generally do them well. Bot-generated descriptions were mostly rated "Irrelevant" or "Somewhat relevant." Participants' top requested changes were: making descriptions visible to sighted users on their own tweets, enabling the feature by default rather than requiring opt-in ("It should be turned on by default. It's almost like they're trying to hide it"), allowing descriptions to be added or edited after posting, and providing guidance on what to include.
Relevance
This paper provides crucial evidence that simply making accessibility features available is insufficient — platform design choices dramatically affect adoption. The finding that only 0.1% of image tweets had descriptions despite the feature existing for two years is a powerful data point for accessibility advocates arguing that platforms must do more than offer opt-in tools. The research directly informed Twitter's subsequent decision to make the image description feature enabled by default and more prominently displayed. For practitioners, the study highlights several actionable insights: accessibility features should be on by default, not hidden in settings; users need guidance and templates for writing good descriptions; descriptions should be editable after posting; and automated tools should supplement (not replace) human descriptions. The interview findings about why well-intentioned users still fail to write descriptions — forgetting, time pressure, uncertainty about what to write — suggest that platforms need prompts, reminders, and AI-assisted pre-fills to bridge the gap between intention and action.
Tags: alternative text · social media accessibility · image accessibility · blindness · screen readers · Twitter · user-generated content
Standards referenced: WCAG 2.0