Expanding the empirical gaze

The debate about the aim of different methodological approaches in empirical bioethics has been intense over the course of the last couple of decades.17 Different methods enable different forms of integration between empirical insight and normative conclusions, depending on whether normative conclusions are to be justified as a matter of consensus or logical argument, whether they depend on processes such as inclusive deliberation or whether such conclusions should be generalisable or limited to a given context of practice.29

Empirical bioethics, however, has only begun to devote specific attention to methods that could be employed to make sense of the way bioethically relevant issues are discursively articulated online.30 The online space is not simply an electronic repository of information. As we mentioned in the introduction, the digital sphere is best understood as a platform of communication and agency existing in a mutually productive relation with analogue cultures and forms of interaction. Over time, this view of the digital as a material culture became prominent in the social sciences. As Horst and Miller10 wrote, digital worlds ‘are neither more nor less material than the worlds that preceded them. […] This concerns humanity’s remarkable capacity to reimpose normativity just as quickly as digital technologies create conditions for change’.

With the rise of the Internet in the mid-1990s, digital networks enabled new forms of communication and interaction. In the span of two decades, the internet evolved from a communication infrastructure to a ‘participatory web’, in which users take an active role creating content and shaping the world wide web, rather than using it as simply an information repository or communication platform.31 32

Missing out on the online space as a site of rigorous empirical research precludes access to influential spheres of discourse and practice that have direct bearing on bioethical analysis and decision making. Digital methods, on the other hand, enable empirical bioethics to document how ethical values and perspectives are articulated online, which principles, interests and intuitions are mobilised and how their relevance fluctuates over time.

However, the value of embracing digital methods in bioethics lies beyond their use in ‘lay of the land’ research. Expanding the reach of bioethics to the online space can foster integrative analysis that fruitfully combine normative and empirical research. More specifically, digital methods can contribute to both normative analysis and the implementation of bioethically relevant norms in practical contexts. As to the former, by looking at how ethically relevant matters are discussed online, digital methods enable probing bioethical arguments based on assumptions about the interests, values and views of different publics. This capacity can also be used to detect areas of overlapping consensus between different stakeholders so as to advance normative conclusions of a pragmatic character.29 33

Moreover, digital methods can detect the emergence of novel normative discourses before they become embedded in mainstream ethical narratives about ongoing controversies. The web is a space of connectedness where bonds are formed and nurtured among people with shared interests and stakes. For this reason, digital methods can be particularly useful to shed light on otherwise disfranchised discursive communities and on the way they spontaneously articulate their moral experience in relation to bioethically relevant issues.3

As we show below, digital methods can reconstruct how moral attitudes of the public or specific stakeholders evolve over time—for instance on social media platforms—enabling the analysis of what triggers people to acquire and manifest publicly their moral attitudes in an issue of bioethical interest. This is possible thanks to the fact that users of social media platforms leave traces of their past interactions that are accessible through computational methods.

Considering integration between empirical and normative bioethics at a more practical level, digital methods can be employed to provide what Mildred Solomon3 has called a ‘bridge between conceiving a moral vision of a better world, and actually enacting it’. In particular, digital methods can reveal how different institutional actors relate to one another as to their online presence enabling interpretive insight into potential impediments due, for instance, to insufficient harmonisation of normative requirements.

We thus suggest that digital methods can be fruitfully employed in empirical bioethics as they shed new light the reality of discursive practices as they take shape in the online space.

In the next section, we explore methods developed by other social science disciplines for the analysis of digital discourse and human interaction online, before turning to illustrate digital methods we have developed and employed specifically in the domain of bioethics.

Online sentiment analysis

In a study conducted in 2018 and 2019, we examined public attitudes towards CRISPR/Cas-9 gene editing through an analysis of social media content posted by Twitter users.51 We acquired a little over 1.5 million tweets containing the keyword CRISPR through GNIP, a Twitter subsidiary and the official Twitter Application Programming Interface (API) for a 6.5-year period. We used a machine learning approach based on NLP to determine the sentiment (positive, neutral and negative) in each of the tweets: a sample of the tweets was labelled through the platform ‘crowdbreaks’ and multiple machine learning models were trained using that sample.52 The type of machine learning model that performed best on a test set was a fairly new type of NLP model called BERT,53 first published in 2018. The trained model was then used to predict the sentiment of the remaining tweets, and a daily sentiment signal with a rolling average of 7 days was calculated, resulting in a sentiment index between −1 (all negative) and 1 (all positive). Using a second classifier trained on the tweet sample, we classified all tweets according to the organism the tweet referred to (e.g. humans, animals, plants etc.) and made a breakdown of the sentiment per organism class over time. The results revealed a generally positive sentiment with a series of negative spikes in the second half of the analysed period and a shift in the general sentiment dependent on the organism class.51 The detection of such discrete changes in ethical sentiment points to a causally relevant role that can be attributed to specific uses of a given technology or to specific advances in basic or applied science. In the case of CRISPR, for instance, the spikes correlate with the announcement, in late 2018, of clinical uses of gene editing to modify the genome of two human embryos eventually resulting in the birth of two baby girls.54 This kind of analysis can help identify precisely which ethical lines the public is uncomfortable with being crossed, which in turn can lead to better calibration of normative arguments and socially more robust decision making around controversial forms of scientific innovation.

Automated topic ranking

In the same study, we analysed the hashtags used in tweets to uncover which topics were most discussed on Twitter. For this method, we used statistics rather than machine learning to see how frequently a hashtag was used per year, which allowed us to see how often the tweets referred to certain topics in each year, that is, the number of tweets containing the corresponding hashtags. The hashtags are particularly useful since they represent labels directly provided by users themselves, for their own statements. Topics were ranked according to their counts, revealing the most discussed topics for each year and how the prominence of a topic changed over time. This topic ranking allowed us to see, for example, the point at which the conversation on the birth of two genetically modified babies in China in late 2018 picked up, and how the prevalence of other topics changed at the same time. Furthermore, linking this topic breakdown to the sentiment of the corresponding tweets allowed us to observe how the sentiment of Twitter users towards that specific topic changed over time.

Sentiment analysis and automated topic ranking employ advanced technical methodologies to large-scale analysis of moral attitudes on divisive public policy issues in bioethics. Moreover, such methods have a remarkable explanatory potential as they enable to observe how ethical views evolve over time and in connection to specific scientific advances or controversial uses of discoveries and technologies. More specifically, analysing hashtags provides a dynamic picture of the relative relevance that the public attributes to different moral concerns. Such a dynamic picture can reveal normative trade-offs between different values and principles, and it can as well drive moral interpretation for instance suggesting ways to balance and specify different ethical principles in specific context.55–57

Thematic probing

Understanding public opinion on gene editing in humans was the focus of multiple studies prior to our study on CRISPR. This offered the opportunity to use digital methods to connect our study with results obtained through conventional surveys and to assess our sentiment analysis in a context beyond Twitter. To that end, we first identified existing studies about public opinion on human gene editing and used an inductive coding approach to identify themes occurring in these studies. The themes were then translated into regular expressions, that is, patterns that describe which character sequences in a text are matched or not,58 which enabled probing of the textual data from the tweets for occurrence of these themes. Thematic probing allowed us to identify which of the themes from the survey results were also present in the conversation on Twitter (ie, the corresponding regular expression was matched in the text of the tweets), what portion of the conversation referred to a specific theme and which sentiments were associated with these themes over the years. This process is similar to a qualitative content analysis with a predefined codebook but allows for probing of a much larger dataset. Moreover, thematic probing enables comparative analysis of online and offline discourses ushering in opportunities for novel insight into how ethical views emerge and diffuses on different communication platforms. Understanding both discrepancies and areas of overlap between online and offline discourses can shed light on different factors that influence moral views. This is useful because it can illuminate the persuasive potential of moral arguments in different contexts. Moreover, as for the previous methods, thematic probing can suggest ways to interpret prima facie norms and balance principles based on their relative weight.

Policy landscape analysis

Another digital method that we developed entailed building research hypotheses to be further investigated with the ultimate aim of generating normative insights for policy and decision making. In a study conducted in 2017,50 we used network analysis to build a hypothesis on existing data sharing policies published online. First, 230 policy documents were identified through a manual online search procedure. We used a digital method to detect the hyperlink structure among the online locations (uniform resource locators or URLs) of sites containing policies or hosting policy documents. We used both a database (moz.org) providing backlinks to policy sites, as well as a web crawler to identify chains of hyperlinks pointing to the policy sites. This procedure resulted in a hyperlink network consisting of connections among the identified policy documents. Using network analysis, we measured the length of the shortest path between policies (ie, the minimum number of steps required to navigate from one policy document to another) and the density of strong network components and found that the our documents set formed only a sparsely connected network. This indicated that organisations working to promote data sharing for research purposes might have developed their guidance in relative isolation from one another. This result suggested that the promotion of data sharing, while widely supported, might suffer from a lack of harmonisation over ethically relevant principles and, on a practical level, inadequate standards to improve data sharing in practice. To test this hypothesis, we performed a conventional qualitative content analysis (with the aid of the NVivo software), resulting in a set of recommendations for how to address bottlenecks hindering efficient data sharing.50

Network analysis enabled a novel understanding of the relationship between different stakeholders (in this case, policy actors and expert bodies) and provided us with insight into the possible explanation for inefficient implementation of ethical norms relative to data sharing practices in the research context. In particular, it enabled us to observe that the actors promoting data sharing guidance form only a loosely connected network online, suggesting a fragmented policy landscape as a possible explanation of why such policies fell short of producing their expected outcome. We further probed this hypothesis by detecting a rather diverse and unfocused set of ethical recommendations in this area. This method focuses primarily on institutional actors and how they relate discursively to one another,25 and it proved useful to drive normative analysis and ethical recommendations.

Added value

Empirical methods of the kind we have proposed here allow research to gain access to novel sites in which different stakeholders (both individual and institutional) express their moral ideals, declare their interests and share their views as to the norms and principles that should guide action in bioethically controversial matters. This is consistent with a common integrative understanding of empirical bioethics and cognate disciplines (see, for instance, refs 25 and 29). However, at the same time, digital methods hold promise to expand the explanatory power of empirical bioethics. In particular, methods to trace the evolution of moral views over time (like sentiment analysis and automated topic ranking) enable a fine-grained evaluation of what triggers certain moral reactions in the public. This can have direct implications for normative analysis, for example, by challenging value judgements about what is—in a given bioethical controversy—that should be deemed worth of ethical consideration.

The relationship between online and offline worlds is complex and mutually productive, but a lot of work is still needed to clearly understand how they influence each other, under which circumstances and to which extent. Methods like thematic probing offer explanatory insight into such dynamics. Finally, a method like the analysis of policy landscapes offers the opportunity to explain how different institutional actors, such as policy-making bodies determine the implementation (or lack thereof) of ethical norms in a given domain of practice, with implications on how best to address potential bottlenecks from a normative point of view.

The methods we have introduced seem particularly useful in the context of bioethical issues reaching the status of public controversies or topics that lend themselves to specific forms of decision making by institutional bodies and actors. Nevertheless, it can be expected that, in the future, computational methods will be employed also in the context of medical ethics, where ethical decision making is of a more particular and context-specific character.

Technical challenges

The successful integration of our new digital bioethics methods into empirical and—in turn—normative bioethics depends on a variety of technical factors and factors concerning the disciplinary constitution of the different fields.

Technically, we ourselves encountered several obstacles that could hinder other researchers employing similar methods. In our analysis of the data sharing policy landscape, for example, different URLs can lead to the same object. While, for example, different character encodings (ie, the technical representation in 0s and 1s, which can be done in different ways) can be detected rather easily, assessment of the equality of two targets might only be possible by comparing the contents of the two webpages. A further challenge resulted from the sheer number of webpages and the even greater number of hyperlinks between them. Finding the shortest path from one webpage to another using manual methods is virtually impossible, and even in employing a web crawler, the time needed to explore several steps of hyperlinks is significant and grows exponentially with the number of steps and corresponding number of websites. We used a backlink service (moz.org) that provided the URLs of webpages that point to the policy webpages, in order to reduce the number of websites from which we needed to extract hyperlinks. However, we could only make limited use of this service due to rate limits.

Cost of data

An additional challenge is presented by the cost of access to data, an issue that we encountered for instance in the Twitter study on CRISPR. While Twitter offers a cost-free API, access to historic data through this API is limited. For our study, we needed to acquire a significant amount of data through the Twitter subsidiary Gnip, which sells historic Twitter data. The cost of these data is significant for a research project, not only in the case of Twitter, but for other data providers as well. Therefore, funding for digital bioethics research must also cover data acquisition costs, for purchasing both archived data and API subscriptions. Alternatively, the scientific community could campaign for Internet data to be freely available for not-for-profit research purposes.

Standards

The variety of accepted methodologies in empirical bioethics has fostered the distinct interdisciplinary nature of this field of research.29 Accordingly, many scholars have called for a more rigorous definition of methodological standards to support the growth of the discipline. Samia Hurst, for instance, points out that bioethics has imported methods from empirical disciplines without absorbing the ‘standards to which researchers in these disciplines are held’.14 This concern spans choice of appropriate methodology, data collection practices, data presentation, consistency of empirical conclusions and understanding of study limitations. Ives and colleagues60 identified a series of advantages that could be derived from widespread adoption of methodological standards of practice in empirical bioethics research. In particular, such standards would enhance consistency in the assessment of grant proposals and the evaluation of peer-reviewed articles, making such processes more predictable for researchers and more manageable for funders and editors.

With the development of novel digital methods, questions arise over standards and data quality as well.34 While social scientists can resort to standards of conventional empirical methods, the question of how these standards can and should be adopted for digital methods is a matter of debate.49 This challenge is further pronounced in the case of digital bioethics methods, due to the ongoing discussion surrounding empirical bioethics standards in general.14 60 By contrast, many of the technical components used for digital methods (such as machine learning) have established standards and best practices of their own. Therefore, we can and should adhere to these standards when such computational methods are used. However, rapid improvements in fields like machine learning may quickly render present standards obsolete.

Training and skills

Beyond technical and structural limitations, the use of digital bioethics methods requires a substantial amount of technical skills and expertise. The background of bioethics scholars and researchers does not typically include extensive technical training, and this can limit the choice of digital methods for bioethics research groups. In this respect, the potential of digital methods for empirical bioethics research depends on the willingness of research institutions to further support the kind of interdisciplinarity that has already characterised bioethics over the past two decades. This consideration applies to universities, funding agencies and editorial boards alike. As the importance of digital methods in the social sciences increases, bioethics journals and reviewers must be open to manuscripts employing these methods. Initially, the use of digital methods will not be supported by commonly accepted quality standards, which will arguably complicate the assessment of manuscripts. However, this problem was also present in the early days of empirical bioethics and did not prevent the growth of the field.14 60