The current study qualitatively analyzed a subset of tweets that were collected in a sentiment analysis and topic modelling study that programmatically explored content and sentiment in tweets mentioning a similar ban on graphic images of self‐harm instituted by Facebook and Instagram [20]. Briefly, the premium Twitter API [22] captured 238,001 original posts, retweets, and quotes from June 1, 2018, to May 31, 2019, containing the terms “self-injurious behavior,” “nonsuicidal self-injury,” or “self-harm.” These terms were selected based on previous studies on adolescent self-harm, and therefore may provide data more representative of academic- or policy-related discussion.

While previous work explored tweets mentioning “Facebook” or “Instagram” [20], the current study identified random samples of tweets that were posted in English, between October 18, 2018, to May 29, 2019. These dates were selected as they were approximately 16 weeks before and after Facebook and Instagram’s graphic self-harm image ban was enacted on February 7, 2019 [18]. The rationale for selecting tweets within this range was to assess and compare the immediate impact of the policy on Twitter posts.

Sample size was calculated using a priori power analysis for a chi-squared test for comparing the rate of certain types of tweets before and after the ban was enacted. It was determined that a sample size of 99 tweets per week would yield adequate power (approximately 0.80) for detecting a small effect (h = 0.1). To account for the fact that many tweets do not have adequate information for coding (e.g., only links or emojis) and that some categories may be sparser than others, a random sample of 126 tweets per week (oversampled by roughly 25%) was taken for a total of 4,032 tweets. All initial data collection protocols were deemed exempt by the Colgate University Institutional Review Board (IRB) according to the Code of Federal Regulations Title 45, Part 46 (Exemption 2: Public data).

Tweets were aggregated and analyzed using an Excel file, with hashtags excluded to focus on the tweet content itself. The qualitative analysis was conducted using multi-stage content analysis methods [23, 24]. Two independent reviewers coded the tweets, and inter-coder reliability was assessed using Cohen’s Kappa. Any discrepancies were resolved by a third reviewer. Tweets were initially classified by their slant—pro, anti, or neutral—to determine whether they conveyed a positive, negative, or neutral attitude toward self-harm or suicide.

Tweets were then classified into one of six content categories adapted from Jimenez‐Sotomayor et al. [25].

(1)

Personal opinions: express the author’s subjective viewpoint, beliefs, or interpretations related to suicide or self-harm, often without citing evidence or sharing lived experience. These may include criticisms, support, or general thoughts on the topic.

Lastly, manifest content analysis was used to identify themes—that is, patterns and meanings—across tweets [26]. This method provides a clear, observable, and literal description of the phenomenon by developing themes that remain close to the text [26, 27]. Initially, two independent coders analyzed approximately 2,500 tweets to create a codebook [28]. After discussion and revision among the research team, the codebook was applied to analyze the entire dataset. The final coding was reviewed by team members, and any discrepancies were resolved through discussion.

Descriptive statistics were used to report each tweet type’s mean, median, standard deviation, and slant. To compare trends before and after the policy enactment, tweet slants and content categories from the period before the ban (October 18, 2018, to February 6, 2019) were compared to those posted after the ban (February 8, 2019, to May 29, 2019).

For the analysis, suicide and NSSI behaviors were collectively reported as self-harm. From the 4,032 tweets, 186 were excluded (n = 143 not in English, n = 13 not relevant to self-harm, and n = 30 were on February 7th—the day the ban was announced). Subsequently, 3,846 tweets were analyzed. Among these tweets, the level of agreement when coding was very high for type (Cohen’s k = 0.9294), slant (Cohen’s k = 0.9004), and the combined coding (Cohen’s k = 0.9018). Both before and after the enactment of the policy, the majority of tweets had an anti-self-harm slant, and only 2.78% of the tweets were pro-self-harm. Most of the tweets were personal opinions (28.19%), followed by personal accounts and experiences (25.20%) and informative posts (24.41%).

The most dominant themes included understanding self-harm (41.11%), political (16.56%), and support (13.36%). While anti-self-harm tweets were less prevalent after the ban, neutral self-harm tweets were more prevalent (Figure 1). Personal accounts and non-categorized tweets reduced after the ban (Figure 2). Conversely, tweets that were personal opinions, informative, and advice-seeking became more prevalent (Figure 2). The sole theme that became more prevalent after the ban was understanding self-harm (Figure 3). A full list of tweet slants, categories, and themes can be found in Table 1.

The qualitative analysis of the tweets revealed eleven distinct themes (Table 2). The majority of these tweets focused on understanding self-harm (41.11%), followed by political tweets that made inferences about self-harm (16.56%), supporting individuals affected by self-harm (13.36%), and calling out wrongdoings (12.69%). Only a small fraction of tweets promoted the acceptance of self-harm (0.16%).

Many tweets provided information to better understand what constitutes self-harm, who it may impact, and how it should be addressed. Most of these tweets were in discussion format (28%; 443/1,581), where different users shared information on topics related to self-harm. Many tweets also supported raising awareness (19.9%; 314/1,581) on self-harm. Information related to self-harm topics was also disseminated through links to news articles and blog posts (17.3%; 273/1,581). Personal experiences of users were sometimes used to learn about self-harm (15.5%; 245/1,581). Different reasons for engaging in self-harm were expressed by users. The most common reason was that the behavior was used as a coping mechanism (26.8%; 45/168). Urges or thoughts were also commonly expressed as a reason to self-harm (18.5%; 31/168). In some users, environmental factors, such as societal triggers and views on self-harm, also increased the risk of self-harm (8.9%; 15/168).

The most common comparisons of self-harm were different lifestyle choices (26.1%; 40/153), such as staying up too late or traveling to certain regions. Consuming certain types of foods, particularly fast foods, was also seen as self-harm (13.1%; 20/153). Viewing certain TV shows and video games was also described as a form of self-harm (10.5%; 16/153). Much of the commentary about the media centered on self-harm portrayed in TV or film (39.1%; 52/133), social media (30.1%; 40/133), and books (14.3%; 19/133). Many of the tweets supported or challenged censorship of self-harm and highlighted the importance of discussing self-harm in these media formats.

Many users also expressed what they perceived to be wrongdoings. The majority of these tweets focused on social media (15%; 73/488), with some users concerned about the lack of attention that social media companies pay to self-harm content and users reporting this content. Some users (13.9%; 68/488) also discussed the quality of care that made it difficult for individuals to seek help for self-harm. Other concerns that were expressed pertained to reducing bullying that may result in self-harm (12.7%; 62/488), trivializing or minimizing self-harm behaviors (9.8%; 48/488), and glorifying self-harm (7.8%; 38/488). Most political tweets considered certain policies and laws as being a form of self-harm. More than half of the political tweets centered on either the United Kingdom (18.1%; 115/637) or, more specifically, Brexit (51.2%; 326/637), referring to it as an act of self-harm.

Tweets pertaining to support involved users being proud of themselves (39.7%; 204/514) and others (31.7%; 163/514) for coping with self-harm and their recovery journey. Several encouraging tweets mentioned the duration of time that individuals refrained from engaging in self-harm. Some users also expressed the need to get help from others (14%; 72/514). Links to resources for individuals seeking help from self-harm were predominantly treatments or services (34.9%; 37/106) or hotlines (35.8%; 38/106) provided by organizations focused on self-harm. Links to mobile applications to manage self-harm or related mental health concerns, such as anxiety, panic attacks, and depression, were also observed in the corpus of tweets (11.3%; 12/106).

With respect to the minor themes, the majority of tweets accepting self-harm promoted the behavior and glorified self-harm as a positive activity that should be engaged in. Tweets pertaining to self-harm terms contained either a single word or multiple words related to self-harm, and without a coherent sentence structure.

With respect to tweet slants (Figure 4), most of the anti-self-harm tweets (40.4%; 1,055/2,611) fell under the theme of understanding self-harm. The same theme also had a large proportion of the total neutral slanted self-harm tweets (43.4%; 490/1,128). While there were very few pro-self-harm tweets, most fell under the theme of understanding self-harm (33.6%; 36/107) and reasons for self-harm (29.9%; 32/107).

Within almost all the themes, anti-self-harm tweets were the most predominant. Tweets categorized under political (44.6%; 284/637), comparisons (44.4%; 68/153), and self-harm terms (81.3%; 39/48) were mostly neutral slanted. All themes under accepting self-harm were pro-self-harm tweets (n = 6). Conversely, none of the tweets that fell under the themes of self-harm help resource (n = 106), self-harm terms (n = 48), and repercussions of self-harm (n = 12) contained pro-self-harm tweets.

Regarding tweet categories (Figure 5), most personal opinions fell under the theme of understanding self-harm (38.5%; 417/1,084) or calling out wrongdoings (25.4%; 275/1,084). More than half of informative tweets were also categorized as understanding self-harm (77.6%; 729/939). Half of the tweets that were jokes or ridicule were comparisons (50%; 36/72). Personal accounts and experiences were mostly coded under the themes of support (35.2%; 341/969) or understanding self-harm (32.5%; 315/969). Tweets under advice-seeking mostly represented the theme of understanding self-harm (46.7%; 56/120) or support (29.2%; 35/120). Tweets that did not fall under any type (other) were mostly political (79.6%; 527/662).

A notable limitation of the study pertained to the self-harm language that was captured. Although the dissemination of self-harm content is a pervasive problem on social media, only about 3% of the total tweets analyzed were pro-self-harm. This number is substantially lower than reports of 9–66% pro-self-harm content within other social media platforms [36] but may also reflect the impacts of the earlier ban on Twitter. In contrast, a recent report indicated a 500% increase in self-harm-related hashtags since October 2022 [16]. Specifically, they noted that the "#shtwt" (self-harm TWiTter) hashtag, grew from 4k to nearly 39k tweets in less than a year despite moderation policies. Further, they describe the development of ban-evading coded language (e.g., “cat scratches” instead of “cuts”) that enables users to sidestep moderation. See Atauri-Mezquida et al. [37] for more discussion about these workarounds and further investigations into the (lack of) effectiveness of the ban and Twitter’s moderation policy implementations

The low number of pro-self-harm tweets in the current study may also indicate that the ambiguous language of self-harm within social media platforms evolves in ways to bypass self-harm policies [38] (e.g., #shtwt and “cat scratches”). Given the dynamic and ever-changing nature of this language, it may be challenging to delineate unique terms and examples over the long term. Although identifying specific strategies for tracking emergent self-harm language was beyond the scope of this study, future research should consider integrating machine learning, user-informed keyword development, or collaboration with people with lived experience to better detect and interpret evolving terminology.

The current study demonstrated observable changes in online dialogue surrounding self-harm, offering insight into how public discourse may be shaped by platform-level policies aimed at reducing exposure to potentially harmful content. While the implementation of Instagram and Facebook’s ban on graphic self-harm imagery appeared to coincide with a shift toward conversations focused on understanding self-harm, it also highlighted the challenge of distinguishing between factual information and personal narratives. Importantly, personal reflections and firsthand accounts may serve a meaningful peer-support function, particularly for users seeking connection, validation, or a sense of community. Although this study was not designed to evaluate regulatory or governmental interventions, the findings may inform platform governance strategies and contribute to broader public health conversations. Future research may benefit from interdisciplinary partnerships with mental health organizations, legal scholars, and user communities to guide the development of user-centered, evidence-informed content moderation practices.