1 Introduction

The acquisition of Twitter by Elon Musk in October 2022 has resulted in a search for alternative platforms among many users. This is the main reason why the decentralized microblogging network Mastodon has recently grown quite rapidly (Zia et al. 2023). While Twitter and Mastodon seem similar in the main functionalities they offer, a key difference lies in how they are organized. Unlike Twitter, Mastodon is not controlled by a single entity, but operates on a federated model (as part of the so-called “Fediverse” that also includes other federated services, such as the YouTube and Facebook alternatives PeerTube and friendica), in which independent Mastodon instances are set up and maintained by groups or individuals but user interaction is possible across the entire Fediverse.

The increasing number of users has also attracted interest in Mastodon among academic researchers studying social media and online communication. In addition to studying Mastodon as a platform of increased interest due to user migration, Mastodon has also become interesting as a data source for research since the Twitter Application Programming Interfaces (APIs) have become largely inaccessible or at least quite costly to use (Braun 2024; La Cava et al. 2022). As of now, Mastodon offers an API that can also be used for data access. The Mastodon API is free-to-use and well-documented.Footnote 1 But while, from a technical perspective, Mastodon provides relatively easy data access, the platform poses new practical as well as ethical challenges to researchers who want to gather data from Mastodon. Specifically, its decentralized structure, the composition of its user community and their usage practices raise novel practical and ethical questions that researchers seeking to study Mastodon need to consider and address, both individually and as a research community. In the context of studying online platforms and their users, the research community has, in the past, engaged with the question to what degree platforms’ Terms of Services (ToS) should strictly be followed during data collection or whether other forms of data access than official APIs, such as web scraping might be a viable alternative (Freelon 2018; Mancosu and Vegetti 2020). Notably, these questions are associated with legal as well as ethical considerations. Generally, there is an inherent risk in relying on APIs offered by commercial platforms as their ToS or functionalities might change, or APIs may be shut down or become subject to (substantial) charges. Researchers have, hence, argued that we may be facing an “APIcalypse” (Bruns 2019) or entering a “post-API age” (Freelon 2018) in social media research.

In contrast to this, the decentralized organization of Mastodon is not dependent on the interests of a commercial platform. From a researcher perspective, this also means that Mastodon does not provide universal ToS, and, thus, also no unified legal framework regulating data access. This specific setting might help to shift the focus away from the platforms’ commercial interests to the platforms’ users and their perceptions of ethical use of their data (Halavais 2019). Anecdotal evidence suggests that Mastodon users are conscious of data privacy and potentially less accepting of their data being collected and used for academic purposes since at least one conference paper has already been retractedFootnote 2 after Mastodon users voiced their disagreement with their data being used for research. The affected Mastodon community argued that the publication presented a danger to vulnerable groups and minorities that are active on the platform as well as a violation to the General Data Protection Regulation (GDPR) by failing to de-identify data.Footnote 3 First studies seem to support the assumption that Mastodon users differ from users of other social media platforms in their motivations, characteristics, and, potentially, also their views on data privacy (Lee and Wang 2023; Nicholson et al. 2023), however, overall, systematic empirical evidence on this is still rather scarce.

With this paper, we want to expand empirical research on Mastodon by studying a platform-specific source of information that can be seen as a manifestation of the views and expectations of users: the Mastodon instance rules. These rules are public and formulated by instances maintainers. It is important to note that these rules are not individual users’ data (and, thus, not personal data), but characteristics or metadata of specific instances. By focusing on the rules when investigating, we are trying to understand how the community wants to be treated by researchers. In the context of Mastodon, each instance may set up its own rules for regulating what is permissible or prohibited. These rules typically address the user community and may, for example, specify their target audience for this instance or regulate practices for what may be posted or what kind of content is prohibited. Additionally, instance rules may potentially also address other actors beyond Mastodon user communities, this may also include addressing researchers and specifying if/how data from this instance may or may not be used for research purposes. If such expressions could be found within Mastodon instance rules, they would constitute a new situation in the context of research ethics, where constantly reflecting on practices and re-evaluating approaches is necessary also due to the constant evolution of social media platforms and the emergence of new research methods (Lukito 2024). Notably, clear best practices for ethical procedures in research with social media data do not exist. Instead, situational judgement is recommended in guides such as the ones by the Association of Internet Researchers (Franzke et al. 2020). In practice, however, many researchers struggle to navigate ethical questions for their own work (Weller and Kinder-Kurlanda 2014). Furthermore, Institutional Review Boards (IRBs) or other institutionalized ethics committees also still arrive at different judgements for fundamental questions, such as whether research based on social media users’ data constitutes human subject research (Lukito 2024) and should, hence, be subject to an ethics review. Traditional approaches to handling human subject research, such as informed consent, are also not easily transferable to users of social media platforms (Breuer et al. 2023). In this overall situation, any additional information on user perspectives on the use of social media data for research should be viewed as a potential complementary piece of information to consider for case-based decisions on ethical research design. Mastodon instances offer such a new starting point for learning about and reflecting on users’ perspectives on general ethical judgements in online communication, and questions of research ethics and data collection in specific.

Taking this into account, our study investigates whether and how Mastodon instance rules address the topic of data use for research/scientific purposes and opens the discussion on what implications this has for research ethics. Notably, there are different perspectives on and theoretical frameworks for discussing and investigating research ethics. The two most prominent general theoretical frameworks in the context of (research) ethics are deontology and consequentialism (Salganik 2019). Broadly speaking, a deontological view focuses on following specific norms and adhering to a set of fundamentals, whereas consequentialism emphasizes the evaluation of anticipated outcomes and assessments of their ethical implications. Between these two approaches, our focus is more on a consequentialist perspective to assess trade-offs between potential harms and benefits of using Mastodon data for research. Also, within a deontological or consequentialist approach there can be different perspectives or foci (Salganik 2019; Schlütz and Möhring 2018). For our study, we want to focus particularly on the views and expectations of the people whose data are being collected for research. Given that, in the case of social media research, oftentimes people are not aware that their data are being collected and used for research (as for example shown by Fiesler and Proferes (2018)), using the term (study) participants may not be appropriate in this context (Breuer et al. 2023), especially if data are accessed via APIs. Although, in such scenarios, informed consent as it is normally understood and defined for human-subjects research is impossible or at least very difficult to obtain (Breuer et al. 2023; Stier et al. 2020), we still aim to address questions of consent and platform user views by considering rules of Mastodon instances as expressions of those. Of course, research ethics are not equivalent to considering—let alone following—user views. There are, e.g., also the interests of researchers as well as the value of and public interest in the research itself to take into consideration. However, in this paper we understand Mastodon users mostly as communities whose interests are relevant for ethical considerations, although they may be evaluated differently depending on the specific research context, design, and question. Of particular interest for our research are aspects of informed consent, user privacy perception and risk mitigation. Importantly, this means that we cannot cover the full range of research ethics, but only focus on these specific but important aspects. Based on the general assumption that user perceptions should be part of ethical considerations, with this paper, we aim to provide an empirical basis for understanding instance rules as a potential way of communicating expectations on data usage and research. We first empirically examine the content of Mastodon instance rules with a focus on data collection for research. Second, based on our findings we want to draw conclusions that can inform general guidance for researchers who want to work with Mastodon data and give recommendations that may help their ethical decision-making. The paper, thus, addresses two different research questions:

RQ1

How do Mastodon instances address the collection and use of data for scientific purposes?

RQ2

What recommendations can be given to researchers who plan to use Mastodon data, considering the structure of as well as user expectations in a decentralized social network?


To answer our research questions, we collected all Mastodon instance rules written in English via the Mastodon API (see details on the data collection setup in Sect. 4) and combine both quantitative and qualitative content analysis methods (see Sect. 5). Initially, we use topic modeling to identify common themes across our overall sample of instance rules (see Sect. 5.1). As we are mainly interested in the mention of scientific data use, we further identify rules that deal with (non-)scientific data use with a string detection approach as well as a complementary qualitative content analysis of these rules (see Sect. 5.2). Before describing our research setup and findings, we will first briefly discuss the specifics of Mastodon as an understudied type of social media platform organized as a decentralized social network (Sect. 2) and summarize some current debates around research ethics in the context of social media research (Sect. 3). We then discuss the main findings from our empirical study (Sect. 6.1) and suggest how researchers can use Mastodon in a manner that is ethically sound and meets users’ expectations (Sect. 6.2). We conclude with the limitations of our study and an outlook for future research (Sect. 6.3). Overall, the aim of our study is to contribute to research on ethical consideration in online and social media environments, especially also by focusing on how decentralized networks might differ from other types of platforms.

2 Mastodon and decentralized networks

Mastodon is a decentralized open-source social network (DOSN) that was created as an alternative to centralized social networks like Twitter/X. It was originally developed by Eugen Rochko and launched in 2016. The basic idea behind Mastodon is to create and maintain a “federated” social network where different instances (servers) are interconnected but independently operated (La Cava et al. 2022; Raman et al. 2019).Footnote 4 Key features of Mastodon are essentially equivalent to the key features of Twitter/X: on both platforms, content is organized in posts (formerly known as “toots” on Mastodon and tweets on Twitter/X) that can also be reposted by others via boosts on Mastodon and reposts (formerly retweets) on Twitter/X. Because Mastodon functions as a “persistent channel […] of mass personal communication” with “user interaction” and “user-generated content” (Carr and Hayes 2015, p. 8), it can generally be defined as a social media platform.

Despite its decentralized architecture, Mastodon facilitates user interactions across instances by supporting key functions like follows, posts, boosts, or favorites across servers/instances using the shared Activity Pub protocol (Zignani et al. 2018). It is important to note here that Mastodon is not merely a replication of Twitter; instead, it offers a distinctive, community-based user experience across various servers (La Cava et al. 2022). These unique features have contributed to the appeal of Mastodon, particularly also among tech-savvy users and communities with shared particular interests. While the general platform functionalities and affordances as well as migration of users suggest comparisons between Twitter and Mastodon, another social media platform that Mastodon shares important similarities with is Reddit. In particular, there are similarities in structure and design, as both platforms are essentially comprised of “autonomous subparts of a larger whole” (Nicholson et al. 2023, p. 86). The communication infrastructure on Reddit is defined by subreddits. These subreddits typically revolve around a common interest, fostering communities based on identity or affinity. While subreddits are subject to Reddit’s overarching platform policies, subreddits often also have their own rules for posting as well as distinct content moderation practices (Fiesler et al. 2018). This is fairly similar to Mastodon which employs a platform governance model of “covenantal federalism” where instances can have their own rules while abiding by a shared ethical code in terms of the structure and maintenance of the overall network (Gehl and Zulli 2022). In addition to requirements regarding technical aspects (e.g., daily backups and other security requirements), the shared principles include active moderation against racism, sexism, homophobia and transphobia, as defined in the “Mastodon Covenant”.Footnote 5 These parallels imply that ethical considerations for research on Mastodon and Reddit should be similar as well, at least to a certain degree (for an in-depth discussion of ethical considerations in Reddit research, see Fiesler et al. 2024). The decentralized structure of Mastodon ensures that there is no central node governing the flow of information, allowing each instance to independently define its own usage rules. This means that, in contrast to centralized platforms that have platform wide ToS and generalized regulations, Mastodon’s instances work independently from one another and largely on their own terms. Mastodon instances should adhere to a general code of conduct defined in the “Mastodon Covenant” but lack a central platform authority to which they are beholden.

Due to the lack of recommendation systems and algorithmically curated feeds, connections in Mastodon are more topic-oriented rather than being popularity-driven. Correspondingly, the platform places greater emphasis on fostering conversations and interactions over favorites and reshares, thereby favoring more community-centered experiences (Jeong et al. 2023; La Cava et al. 2022). Of course, the structure of Mastodon also has some potential downsides or challenges. From a user perspective, the decentralization requires certain additional effort: Since there are many different instances (around 17.000 in 2023) choosing the right instance to join may be a challenge or even a potential barrier for new users. In addition, due to the absence of recommendation systems and algorithmic curation, Mastodon requires users to invest some time into building their own network and feeds. To make it easier to select an instance, new servers are listed if they follow the Mastodon Covenant. One aspect of the covenant is the definition of server rules. These are considered an integral part of allowing users to select an instance. The Mastodon Covenant also suggests that rules should be kept short and concise. The central function of the rules is to inform potential users so that they can make a decision to register on the instance. The rules are publicly accessible. Similar platform governance can also be found on Reddit (Fiesler et al. 2018), although these are subject to the central Reddit platform, while Mastodon lacks a central platform authority.

Finally, as with all social media platforms, one challenge for researchers lies in understanding the user demographics and what they mean for research on the platform. Mastodon may have very specific self-selection biases in its user bases, also compared to other platforms: Besides a certain level of technical expertise and topical interests, users that are drawn to Mastodon may prioritize the protection of their data and a language that is free of hate speech and harassment (Lee and Wang 2023; Nicholson et al. 2023). As we have seen in the case described above, the specific characteristics of the user base can lead to very tangible implications for researchers that intend to work with Mastodon data, when privacy-aware users voice objections towards the use of their data for research purposes. However, while it may (to some degree) be specific to Mastodon, this particular incident should mainly be a reminder, that all social media researchers should carefully reflect on privacy and ethical considerations when working with user data from any platform. And the fact that other users of other platforms have not voiced similar objections publicly does not mean that they are less vulnerable or that they knowingly give their consent to being studied. Hence, it is necessary to consider general discussions of ethics in social media research to be able to better understand and contextualize the specific case of Mastodon.

3 Ethical research with social media data

The collection and usage of data that represent people’s attitudes, beliefs, feelings, and relationships as it is the fact on social media platforms, is naturally associated with ethical questions (Taylor and Pagliari 2018). Hence, since researchers have started working with social media data there has also been a growing need to discuss ethical ramifications of such research, with intensive debate in academia and beyond continuing until today. Particular challenges arise with regard to (the lack of) informed consent and the delineation between public and private domains in communication and interaction. Further, the balancing of tradeoffs between protecting privacy and maintaining transparency and openness in research is an important topic of discussion (Breuer et al. 2023).

In general, the “But-the-data-is-already-public”-attitudes that were particularly prevalent in the early years of social media research (Zimmer 2010) are still very present (Salganik 2019). Advocates argue that if information is openly available, it can be freely used for research without any substantial concerns for data privacy. This argument is often further supported by pointing out the value of such data for research with regard to knowledge generation and innovation (Taylor and Pagliari 2018). From early on, however, there have also been opposing views that stress privacy issues or risks of potential misuse (Zimmer 2010). This perspective emphasizes that publicly available information may be used in ways that undermine users’ privacy and may further contribute to harmful consequences such as doxing, harassment, or surveillance, especially when sensitive information is involved.

Due to absent explicit consent by the studied individuals, researchers need to assess how public or private the information can be regarded. Further, considerations of sensitivity and subject vulnerability are vital (Breuer et al. 2020; McKee and Porter 2009). In this context, it is also important to consider the user perspective. For example, users’ understanding of what constitutes public communication or sensitive content may be highly idiosyncratic. A qualitative study by Williams et al. (2017b), e.g., showed that users are concerned about being identifiable in research and that this could lead to unsolicited attention online and possibly abuse. Other studies have found that many users perceive their data being used for research purposes as “creepy” or “scary” and that they may felt exploited (Golder et al. 2017; Mikal et al. 2016). Overall, many users may not even be aware that social media data is used for research or in what way (Breuer et al. 2023; Fiesler et al. 2018; Quinton and Reynolds 2017).

Uncertainties, however, not only exist on the side of the platform users but also among researchers. While there are several general guidelines available for ethical questions related to social media research (e.g., Franzke et al. 2020; Rau et al. 2021; Williams et al. 2017a), researchers still often feel unsure when making ethical decisions in practice (Taylor and Pagliari 2018). Also, while there is some research on user perceptions of data use for research for other platforms that may guide researchers, especially for Twitter (Fiesler et al. 2018; Mikal et al. 2016), such systematic evidence is not available to the same degree for Mastodon. This also mirrors the general scarcity of empirical research on Mastodon and its users so far.

For the case of Mastodon, the decentralized nature makes ethical decision-making even more complicated, given that user behavior and usage rules vary widely across instances. At the same time the instance rules offer a unique way to study values and expectations of communities on Mastodon. Hence, assessing these rules can help the scientific community in making informed judgements about the benefits and risks of data use. To arrive at a better understanding of the expectations of Mastodon users regarding the use of their data for research purposes, we conducted an empirical assessment of topics covered in Mastodon instance rules. In that, we are specifically interested in whether and how the scientific use of (user) data is mentioned in these rules, and how this information may be used to inform ethical decision making in research with Mastodon data.

4 Methods

In this section, we describe the data collection (Sect. 4.1), processing (Sect. 4.2), and analysis (Sect. 4.3) steps of our study.Footnote 6 It is important to note, that no individual user data was collected for this work. Data was only collected on the level of Mastodon instances’ rules and instance metadata (such as instance name, number of users, activity etc.). Hence, the data is not personal data according to the General Data Protection Regulation (GDPR) that governs the use of data in the EU. We also do not assume any risks to the Mastodon user community resulting from our work. Rather we expect that our work will help in raising the awareness for the values and expectations of Mastodon users and the implications of this for research practices. These aspects have guided our own ethical considerations for this study.

What may be relevant to note in this context is that besides ethical questions, research with Mastodon data also creates specific practical and technical challenges that researchers should be aware of. Many of those are connected to the decentralized nature of the network and an associated potential ephemerality of instances. Since anyone (with the required technical expertise) can run a Mastodon instance to be part of the Fediverse, the network structure is highly dynamic with new instances emerging and old ones disappearing very frequently. As the user numbers per instance can also fluctuate quite substantially, this makes it difficult to arrive at reliable estimates of the number of instances and users across the Fediverse. Likewise, the growth or the half-life of instances is difficult to foresee. The structure of Mastodon also brings along some other methodological challenges. For example, when sending a request to all instances via the Mastodon API, some of the listed instances may not be active anymore leading to a status error.Footnote 7 Hence, monitoring status or error messages is important to identify missing data or potential systematic errors in the data collection.

4.1 Data collection

Regarding the specific procedure for our study, we collected text data from the rules that govern individual Mastodon instances. Using the Mastodon API and the rtoot package (version 0.3.3) (Schoch and Chan 2023) for R (version 4.3.2 via the RStudio IDE (version 2023.12.1)), we accessed all existing Mastodon instances (17,390) as of May 2023.Footnote 8 Importantly, not all instances reported an active status due to error responses or server downtime, which affected our sampling, and were, hence, removed from our data set. After that, we started by collecting rules from the remaining 6838 active instances (39%). We further decided on analyzing only those rules that were written in English. The cld3 (version 1.5.0) language detection package (Ooms 2023) was used to identify rules written in English and removed all instances with non-English rules texts (2467 or 36% of all active instances). Our final dataset used for the analyses comprises 4371 (64% from all active instances and 25% of all instances found in 05/2023) instance rules.

4.2 Data processing

Data processing involves two steps: First, we applied search terms to the full sample to identify rules that address scientific data use. Second, the rules were pre-processed for further analysis. We use string detection with certain key words as a first step to identify those instances that potentially address the use of data for research. In order to identify relevant keywords, we proceeded iteratively and started with a few general key terms, such as “research”, “science”, or “consent”. Based on this initial research, we analyzed the context of the keywords (keyword-in-context-search) in order to identify further relevant terms. Starting with a broad list of keywords and expanding this list minimizes the risk of overlooking or missing relevant search terms. However, this approach also means that a high proportion of “false positives” is to be expected: the approach was to collect as many as possible instance rules that have may potentially address data collection for research as a topic, in order to miss as little as possible input for the qualitative analysis. On the other hand, this also means that likely many false positives will be included in the data, that match our string search but do not address the topic we are primarily interested in. While this approach is beneficial for the qualitative analysis and for mapping out how often instance rules address data collection from researchers, it also introduces some limitations on the quantitative part of the analysis in the topic modelling section. The found topics are also influenced by the search terms and, hence, the topic model results need to be interpreted accordingly. We have grouped our final list of search terms into three broader categories (see Table 1): These three categories relate to (1) research/science, (2) public data access, and (3) ethics/consent. When applying the search terms, wildcards are used for including various word forms.

Table 1 List of search terms for the string detection
Full size table

The string detection was applied to the entire sample without any data preprocessing. Using the search terms presented in Table 1, we identified a total of 488 (11.2%) instances from our full sample whose rules included one or more of the terms. Overall, 68 (13.9%) instances were found that include our keywords for research/science, 196 (40.1%) included terms from the public data access category, and 297 (60.9%) instance rules contained keywords related to ethics/consent. Notably, these categories are not mutually exclusive, i.e., instances can correspond to more than one category. To prepare for the qualitative analysis we conducted a manual check of the rules that were identified by the string detection based on our keyword categories. In this way, we cross-validate the string detection and, thus, check for false positives (see Sect. 5.2).

With the quantitative text analysis, we wanted to investigate what topics are generally addressed in Mastodon instance rules and whether or where and to what degree the use of data for research is part of those general topics. For thar purpose, we needed to preprocess the instance rule texts. Specifically, our preprocessing for the quantitative text analysis included tokenization, stemming, removing stop words, and counting word frequencies. First, we removed all special characters, numbers, and separators from the data corpus. Within the exploratory text analysis we, then, tokenized the corpus and removed stop words with a predefined stop word list from the quanteda package (version 3.3.1) (Benoit et al. 2018). After an initial analysis we extended the stop word list with additional terms.

4.3 Data analysis

Our analysis relies on a combination of both qualitative and quantitative content analysis methods. While we are primarily interested in how Mastodon instances address the scientific use of data, we also want to contribute to a general understanding of the rules that govern the “Fediverse”.

First, we examine the general content of the rules to get an overview of the topics and restrictions they cover. As there still is a lack of research on the role of Mastodon instances, the nature of these analyses is very exploratory. Specifically, we have employed an LDA topic model with the R‑package topicmodels (version 0.2–14) (Grün and Hornik 2011) to detect and classify the main prevalent themes in all instance rules in our full sample. We chose LDA over other, more advanced topic modeling approaches as it fit our need for simple exploratory approach. While, for example, Structural Topic Modeling (STM) allows for the inclusion of document-level covariates and metadata, we did not classify and collect such information for the instance and had no research question related to instance attributes or potential differences between instance types. Instead, with our exploratory quantitative text analysis we want to find out which topics are mainly covered in rules across instances.

To answer our research questions and in order to get a deeper understanding of how scientific data use is mentioned within the rules, in addition to the exploratory quantitative analyses, we conducted a qualitative analysis of our string detection subset. For this, we used the qualitative data analysis software MAXQDA 2022 (version 22.7.0). We chose a fully inductive approach since it allows flexibility within the analysis process and due to the lack of predetermined codebooks or theories for this new area of research. It also allowed us a more detailed look at the actual content of the rules and patterns in the data. As a first step, we coded the subset for true and false positives, i.e. which rules actually addressed the collection or use of data for scientific purposes, and which were falsely picked up by our string detection. This was done by one coder. However, to assess the reliability of this binary coding, we calculated an intercoder agreement for 30% (148 rules) of the dataset with a second coder. The agreement was 97.3% with a Krippendorff’s alpha (α) = 0.83. Next, we coded the rules inductively for possible emerging themes or overarching patterns. Lastly, we derived categories from the identified patterns in the qualitative analysis. This was done by one coder, while the inductive process and categorization were discussed multiple times and agreed upon within the research team. Our full data collection, processing, and analysis procedure is summarized in Fig. 1.

Fig. 1
figure 1

Summary of the research design

Full size image

5 Results

Depending on the type of analysis, our results are either based on the full sample (n = 4371) or the sub-sample identified via string detection of search terms (n = 488). We will therefore discuss the results of our analysis separately.

5.1 Identifying topics in instance rules

To identify the most common topics addressed in Mastodon instance rules, we have employed a topic modeling approach (Latent Dirichlet Allocation, LDA).Footnote 9 We followed an iterative procedure based on a comparison of the standard LDA fit metrics of the method FindTopicsNumber from the package ldatuning (Nikita and Chaney 2020), and a subsequent manual inspection of the topics and discussions within the research team. The latter enables us to extract interpretable and relevant topics. This resulted in the identification of seven distinct topics, which we grouped into two overarching categories. Figure 2 shows three topics that mainly refer to dealing with the interaction among users together with the most frequent words. This includes targeted attacks against individuals (doxing, dogpiling, etc.), discrimination of groups (racism, transphobia, sexism, etc.), and the overall conversational tone (being kind and showing respect). Another category refers to rules addressing the users’ own sharing and posting behavior, such as the prohibition of expressing and promoting violence, the sharing of misinformation, and posting explicit content or spam. Addressing the conversational tone and targeted attacks against individuals are the most prevalent topics across all instance rules. Both appear in around 20%, while sharing misinformation is mentioned in only 2% of the rules (see Appendix Table 3). One topic could not be grouped into a category and was labeled as “Other”.

Fig. 2
figure 2

Results of the LDA

Full size image

Overall, the results of the topic model suggest that instance rules specifically address issues related to user behavior and interaction. The prevalent topics from the two categories indicate that the instance rules are likely meant to ensure a safe and positive experience for the users. Accordingly, the main addressees of the rules are the instance users and not external actors, such as researchers.

5.2 Qualitative content analysis of instance rules

When looking more closely at the identified rules that matched our keywords from one or more of the three categories, we found that only 31 instances explicitly address the collection or use of data, while the other 457 instances that were found via string detection of our selected keywords were essentially false positives. Table 2 shows in what keyword categories the identified instances fall, including the proportion of false positives with regards to our research question. Within our qualitative analysis of the identified rules, we noticed that the majority of false positives came from the keyword “privacy” in the Ethics/consent category and the word stem “copy*”, mostly referring to “copyright”.

Table 2 Number of true and false positives for the three categories
Full size table

In our in-depth inductive qualitative content analysis of the remaining instance rules, we derived two overarching themes that the rules addressed when regulating the use or collection of platform user data for academic purposes: First, some rules highlighted that data collection through, e.g., scraping or the API is forbidden. Here it also became apparent that a large proportion of the instances use the exact same wording for this.

“[…] No scraping content from web interface or API […]” (Instance 163)

“Toots from or carried by this instance should not be scraped or indexed/the instance’s privacy policy gives details of some of the anti-scraper measures which are and can be used […]” (Instance 330)

In some cases, the rules were also very clear in prohibiting all research practices concerning the instance,Footnote 10 as can be seen in the quote below.

“Don’t index us, don’t research us.” (Instance 723)

Our second category refers to rules pointing towards general requirements regarding research ethics. These instances do not prohibit data collection or use by definition, but rather issue a warning to abide by ethical and data privacy standards.

“If you are doing research […] you are likely performing Human Subjects Research and expected to abide by highest standards of ethics and professional practice. Meeting this expectation is your responsibility. We consider engaging in unethical research […] to be a cause for bans or other sanctions […].” (Instance 395)

Apart from these examples, we also identified a number of instances that, although not addressing research practices in their rules, issued a cautionary note for their users that all content of this instance is public and that it may be seen and used by other parties outside of this instance.

“This forum is NOT confidential. All messages are archived on the web and can be widely read. Do not say anything that would compromise anyone’s confidentiality or violate your professional ethics in any way […]” (Instance 166)

Furthermore, we noted that 18 instances referred to having a more detailed privacy policies or codes of conduct under an external link that either pointed to an instance-related page within Mastodon or to GitHub. We checked these links for additional information on our research subject, however, only one instance that did not previously mention scientific use or collection of data did so under these links. This suggests the rules are typically the place where issues related to the use of an instance are explained. However, in individual incidents, such external links may also include further information that can be relevant for the use of data for research.

6 Discussion and conclusion

6.1 How mastodon instances address the collection and use of data for scientific purposes (RQ1)

In our study, we analyzed the rules of Mastodon instances with regard to whether and how the collection of data for scientific purposes is addressed and regulated. Overall, we found that there are some instances rules that explicitly address this topic, although at the time of our data collection, they only constitute a small minority of all instance rules. From our initial set of 4371 active instances with rules written in English, we were able to identify 31 instances that explicitly and directly address the collection of data for research purposes in their rules. The main focus of instance rules currently appears to lie on prohibiting certain types of content and regulating user behavior and interactions. This often pertains to racism, sexism, harassment, hate speech, or doxing. One likely reason for this is that these issues are also suggested as relevant to address in instance rules in the Mastodon Server Covenant.

Our findings support those from previous research showing that for Mastodon users, the promotion of specific values seems particularly important (Nicholson et al. 2023). The decentralized structure of Mastodon may lead to a selection of a specific type of users who use Mastodon explicitly as an alternative to commercial platforms or, more recently, as a direct alternative for Twitter/X. One reason for this is likely that they can actively contribute to or at least choose instances based on rules that define the conversational culture on and regulate user behavior (Gehl and Zulli 2022; Zia et al. 2023).

The characteristics of the user community may lead to the expectation that users may hold a (more) critical stance towards user data collection on Mastodon. Of course, as stated before, the total number of 31 instance rules explicitly addressing data use for research purposes is quite low. Providing a detailed explanation for why there are not more instance rules reflecting on research data collection is beyond the scope of this paper. However, based on research on the general lack of awareness for the fact that researchers are using data collected from social media platforms (Breuer et al. 2023; Quinton and Reynolds 2017), we assume that this is also true for a substantial share (if not the majority) of Mastodon users. Accordingly, it is likely that instance administrators are not aware that (scientific) data collection is something that can or should be regulated. There does, however, seem to be a general awareness regarding the accessibility of data as we found several instance rules pointing out the public availability of all data posted from Mastodon.

There might be further reasons for low quantity of mentions in the rules. Although the instance rules are the main site for this, regulations may, in individual cases, also be addressed elsewhere (e.g., in extended external links to terms or policies). Finally, with increasing attention directed towards Mastodon in general as well as in academic publications in particular, it is likely that the awareness of data being used for research and, thus, also the quantity of instances regulating this may increase over time.

A key finding from our analysis, however, is that, while their overall number is small, there are instance rules that explicitly address data use for research and some of those explicitly object to it. Importantly, we would stress that the connection to ethical considerations is not only dependent on the prevalence of instance rules explicitly addressing this. In other words, this does not mean that cases in which data use for research are not explicitly mentioned imply that anything goes. Instead, in such cases, researchers need to engage in a general assessment of risks and benefits and use general sources, such as the AOIR internet research ethics, for guidance. This observation leads to broader implications for the research community and to the question of what can be recommended to researchers planning to work with Mastodon data.

6.2 Recommendations how researchers can use mastodon data in a manner that is ethically sound and compliant with user expectations in a decentralized social network (RQ2)

To answer our second research question, based on our general considerations regarding the structure of Mastodon and research ethics when working with social media data in general as well as the findings of our empirical analysis in particular, we want to present some recommendations for researchers who (intend to) collect Mastodon data. In addition to the general recommendations by Fiesler et al. (2024) for Reddit research that also apply to research with Mastodon data, we want to put forth the following recommendations:

  1. 1.

    Be aware that Mastodon users may consider their activities to be less public than, e.g., users of Twitter/X and may be opposed to the use of their data for research purposes. Given the structure of Mastodon, the platform has a pronounced appeal for specific groups of users. People with high technical expertise and/or interest in particular topics, including uncommon or less socially accepted ones, are likely to be more privacy-aware.

  2. 2.

    Pay attention to instance rules and take them into account when deciding (a) what data to collect and (b) how to handle the data (with regards to analyzing and potentially also sharing them). As the variation in the instance rules is quite large, this most likely requires case-by-case decisions per instance. If researchers want to exclude instances whose rules forbid data collection for research, the practical solution could be as follows: The instances can be identified by their handle and users belonging to them are removed from the data set immediately. Researchers should also keep in mind that the instance rules may change over time. It may also be helpful to document relevant rules texts for reflections on and discussions of research ethics for/in publications based on Mastodon data.

  3. 3.

    Turn the decentralized structure of the platform into an advantage. The decentralized nature of Mastodon provides researchers with the option of involving instance administrators in the data collection process. This involvement can help in obtaining informed consent, increasing transparency, and allowing users to make informed decisions about the use of their data. Administrators have the possibility of posting server-wide announcements that all users can see. This can, e.g., be used to announce a data collection and give users an option to opt-out. For centralized platforms, the implementation of such opt-out options is much more difficult, although some researchers have attempted to develop solutions for this (Zong and Matias 2022).

  4. 4.

    Carefully consider the balance between benefit and harm before collecting Mastodon data. As with all online platform data, reflecting on potential harm on the user community is crucial for research ethics. This should include questions like: What is the societal benefit of the research you seek to conduct? Can the studied communities also directly benefit from this? What are the potential harms for the studied communities and individuals? What measures can be taken to mitigate those? In this context, researchers also need to reflect that the absence of statements about data collection from any Mastodon instance does not automatically represent consent and that collecting data from instances that explicitly object to data collection might still be justified for some research scenarios. For example, for the need to study communities that would never give consent to being studied but pose a threat to the online community or to society as a whole, such as far right actors or other extremists (Fuchs 2018).

Overall, our findings underline that taking into consideration the structure of a platform like Mastodon, the characteristics of its user base and instance rules in particular can be one way to address issues associated with what Puschmann (2019) has called “the wild west of social media research”. Making use the of platform features, e.g., to collaborate with instance administrators to announce data collections, gather informed consent or offer opt-out options can be an implementation of what Halavais (2019) has described as “ethical distributed research”. While not exactly the same, such engagement and collaboration with the community is similar to the idea behind data donations (Boeschoten et al. 2022) as it grants independence from commercial platforms and their interests and can facilitate meeting ethical standards, such as obtaining informed consent.

6.3 Limitations

Notably, as all empirical studies, ours has several limitations that need to be taken into account: First, given the technical issues mentioned before, our collection of instance rules may be incomplete. Due to the nature of the federated network, downtime and short half-lives of instances should be considered when accessing the Mastodon-API. Second, we only analyzed instance rules written in English, which limits the generalization of our results. For example, one of the largest instance after mastodon.social is pawoo.net with almost one million users. As the instance is mainly used by Japanese users and its rules are, hence, written in Japanese, we excluded that instance from our analysis. Third, our analysis can only serve as a starting point for further investigations. The methodological triangulation of quantitative and qualitative approaches has shown that our string detection yielded a high proportion of false positives. This was expected, as we aimed at high recall rather than high precision, in order to reduce the risk of missing any potentially relevant cases. This improved the data for the qualitative part of the analysis. However, at the same time it also affected the quantitative part of the study. Notably, while the search terms were the results of extended discussions among the authors of this paper and their knowledge of the literature on social media research (and related ethical questions), the list has not been systematically validated. However, the high number of false positives we detected in our screening of the rules for our qualitative analysis indicates that our search terms seem to have been broad enough to reach high recall.

Our quantitative analysis approach also had additional limitations. To begin with, despite the manual inspection of fit indices and topic composition, our exploratory approach using a simple LDA contains some subjectivity, e.g., related the selection of number of topics. To arrive at more robust insights, a more systematic validation of the topics is necessary. Ideally, this would entail the repeated collection of instance rules to (a) assess stability or changes over time, and (b) also take into account further languages as well as relevant (structural) differences between instances (such as user numbers or topical foci). Hence, in future research, it may be worthwhile to also consider and include attributes of instances that can be used to compare different types, e.g., using a structural topic modelling (STM) approach. Furthermore, the results from topic modeling are highly sensitive to the preprocessing steps that were applied to the text data beforehand. These could be systematically varied as well in further studies. Overall, our analyses are rather explorative. Besides extending the scope of the analysis to other languages and potentially also other documents (e.g., the “About” sections of instances or external sources linked in some instance rules), one worthwhile extension of this work could also be going beyond a bag-of-words approach for the quantitative analyses.

Going beyond the specific conceptual and methodological limitations of our study, Mastodon’s future relevance must be critically considered for research on this platform. For 2022, we observed an increase in instance numbers, but a downward trend can already be assumed for 2023. This could be related to a stagnation or decline in the number of individual users. The individual effort involved in selecting and registering for an instance may be too much for users looking for a Twitter/X alternative. At the same time, at least two other Twitter/X alternatives have recently been launched, namely Bluesky and Threads. Jack Dorsey, former founder of Twitter, has been involved in the launch of Bluesky, while Threads is part of the Meta Group. Compared to Mastodon, a stronger focus on the user experience and algorithmically sorted feeds are likely to play a greater role for these platforms.

7 Conclusion

Despite the limitations outlined above, we believe that our study provides valuable insights into the self-governance and rules of Mastodon and its instances that, together with the recommendations we have formulated, can inform research with Mastodon data, and ensure that ethical questions in this process can be addressed appropriately.

For future efforts, we particularly also see the exchange between researchers and instance administrators as well as users as a valuable approach to create a better understanding and awareness for intentions and ideas of all parties involved. This could, for example, include interviews with different actors (including researchers as well as Mastodon users and instance administrators) to discuss the general ideal precedence of researchers who want to work with Mastodon data or how to best opt-out as a user. Ideally, such qualitative approaches focusing on instance administrators, or a small group of users should be complemented with further quantitative research on the general attitudes and awareness of Mastodon users regarding the use and collection of digital trace data for research purposes.

Overall, we find that Mastodon and decentralized networks in general require special attention and ethical consideration from researchers who want to work with data of said platforms. They face challenges that have not been encountered with platforms that are based on a centralized structure. It is advisable for researchers to be mindful of technical and especially also ethical challenges in designing their studies and take into consideration the rules of all instances that are included in the data collections. We hope that our empirical insights and our recommendations in this paper can serve as guidance for ethical decision-making for researchers interested in studying and gathering data from Mastodon and other similar platforms.