Does “distance lend enchantment”? Public attitudes to deepfake technology in the United States

1. Public attitudes toward science and technology

Sociological research on public attitudes toward science and technology suggests, at the very least, that the situation is highly complex. Most national studies that have tried to measure attitudes to science and technology in general (“general attitudes”) have found these to be fairly positive, and there is long-term trend data for the United States that shows stability in such measures over decades. But this apparent simplicity is misleading. For example, even the most apparently stable general attitudes can change as cultures change. Also, it has been known for some time that general attitudes are poorly related to specific attitudes on particular scientific and technological topics. While on some specific issues better informed respondents tend to have more positive attitudes, on others they tend to be more critical. Again, in recent decades, political partisanship has greatly affected different publics’ views of science in the United States. And the situation gets only more complicated when we analyze aggregate attitudinal measures, asking (for example) about the relationship between attitudes and other variables, such as interest in or knowledge of science and technology (Allum et al., 2008; Besley and Hill, 2020; Durant, 2022; Durant and Evans, 1995; Funk, 2020), or about the related concepts of trust in scientists and scientific institutions (Achterberg et al., 2015; Dixson et al, 2022).

On the question of the relationship between attitudes and knowledge, we are faced with two contrasting and seemingly contradictory views. First, there is the familiar view that, “to know science is to love it” (Turney, 1998); that is, the closer to or more familiar with science people are, the more positive will be their attitude toward it. It has been argued that such a view underlies a great deal of practical activity aimed at improving public understanding of science, and the presence of this view among practitioners has been detected in a number of empirical studies (e.g. Lewenstein, 1992). Over against this point of view, however, is a quite different notion coming from sociology of science. On this second view, something close to the opposite idea has been proposed. Here are two sociologists of science setting out their case (Collins and Pinch, 2014: 3):

Another idea introduced in this volume is the notion that in science and technology, as in love, “distance lends enchantment.” That is to say, scientific and technological debates seem to be much more simple and straightforward when viewed from a distance. When we find ourselves separated from our loved-ones we remember only why we love them; the faults are forgotten. In the same way, science and technology, when understood through others’ inevitably simplified accounts, look artless. Closer to the centre of a heated debate, the less pre-determined and more artful do science and technology appear. The irony is that, quite contrary to what common-sense might lead one to expect, it is often that the greater one’s direct experience of a case, the less sure one is about what is right.

Collins and Pinch are offering their view, not as a contribution to attitudinal studies per se, but rather as a claim about the ways in which certainty is established among groups of expert scientists and technologists. This claim has been much developed and modified by other sociologists. For example, in his much admired analysis of the accuracy of strategic ballistic missile guidance system, MacKenzie (1993) has argued for the existence of a “Certainty Trough,” in which expert assessments of the uncertainty attaching to estimates of the accuracy of strategic ballistic missile guidance systems were higher among expert developers of guidance technology and among critical experts who favored alternative approaches, but lower among potential consumers of this technology—those who, as he put it, “believe what the brochures tell them.” In her study of climate modelers, however, Lahsen (2005) qualified this conclusion by questioning that knowledge producers are always the best judges of the accuracy of their own models.

In the present study we are not concerned, as MacKenzie was, with expert perceptions; rather, ours is a study of different publics, none of whom are technological experts. Also, we are not concerned with issues of “certainty” and “uncertainty,” but instead with attitudinal perceptions. None of the participants in our study seriously questioned the certainty or reliability of the technology under consideration. However, they were variously concerned about its personal, social, and even political implications. Taking into account the more general connotations of the term “enchantment,” we ask here: is greater closeness to or familiarity with a specific technology associated with more positive attitudes, or more negative ones? Is it indeed the case that “to know science is to love it,” or rather is it the case that, in this attitudinal sense, “distance lends enchantment”?

Our study was designed to explore this question empirically. We have chosen one specific area of science and technology, namely recent developments in Artificial Intelligence (AI) that are associated with so-called deepfake technology (see Section 2). We have been enabled to explore the relationship between “closeness” to this topic and feelings about it by taking advantage of the fact that the MIT Museum—where we were both located at the time this research was undertaken—offered highly relevant exhibits and public programs in the period 2022–2023. These resources allowed us to convene three focus groups (FGs) comprising individuals with very different degrees of “closeness” to or familiarity with deepfake technology. While the results reported here should not be generalized uncritically to other areas and issues, they do provide in-principle testable insights into the relationship between public knowledge of and attitudes toward technology.

2. Deepfake

The manipulation of still and moving images has a history almost as long as the histories of photography and film-making. However, in recent decades, AI has been widely exploited in the increasingly sophisticated manipulation of visual and audio content in ways that are potentially deceptive and often—though not always—controversial. ¹ The term “deepfake” appears to have originated in 2017 from a Reddit user named “deepfakes” (Cole, 2018). From around this time, videos started to appear online in which politicians’ and celebrities’ faces were superimposed on the bodies of other people (including porn stars), and these and other applications provoked an increasing amount of critical attention. As one commentator has observed (Wong, 2022),

It was 2018, and the world as we knew it—or rather, how we knew it—teetered on a precipice. Against a rising drone of misinformation, The New York Times , the BBC, Good Morning America , and just about everyone else sounded the alarm over a new strain of fake but highly realistic videos. Using artificial intelligence, bad actors could manipulate someone’s voice and face in recorded footage almost like a virtual puppet and pass the product off as real. In a famous example engineered by BuzzFeed, Barack Obama seemed to say, “President Trump is a total and complete dipshit.”

Deepfake is not the only application area that has recently focused public attention on the wider ethical and social implications of AI. In particular, large language model-based chatbots such as the Generative Pre-trained Transformer (GPT) series have attracted enormous public interest, especially since the launch of ChatGPT in November 2022. In December 2022, tech entrepreneur Elon Musk announced that AI was at risk of being “woke”; in January 2023, the Microsoft Corporation announced an investment of over $10bn in OpenAI; in February 2023, New York Times columnist Kevin Roose published an extraordinary transcript of his own conversation with Bing AI, in the course of which the AI was reported to have assured him, “I Want to Be Alive.”; in March 2023, Musk and tech entrepreneur Steve Wozniak co-signed a call for a 6-month moratorium on the training of AI systems more powerful that GPT-4; and in May 2023 AI pioneer Geoffrey Hinton resigned from his position at Google in order to be able speak freely about what he saw as the dangers of a technology in whose development he had played a major part. By the time that the US Congress came to conduct its first hearing on oversight of AI, also in May 2023, ChatGPT had become one of the most widely talked-about technologies of recent times (Bartholomew and Mehta, 2023).

The MIT Museum contained some exhibits intended to engage audiences with some of these developments. An exhibition entitled AI: Mind the Gap announced that it “compels us to question [AI’s] potential and recognize its risks.” This exhibition included an exhibit on generative AI called Let’s Write a Poem Together, which allowed visitors to create a poem collaboratively with GPT-3 (at that time ChatGPT had not yet been released), and another on deepfake, called True or False. True or False had two sections: an art installation, In Event of Moon Disaster; and an interactive video-wall, Detect a Fake. In Event of Moon Disaster was a museum implementation of an immersive multi-platform art work using deepfake technology to produce media that would have existed in the event that the 1969 Apollo 11 mission had failed (Panetta and Burgund, 2019). Detect a Fake was an array of videos of people talking, in which visitors were invited to guess which videos were real and which were fake.

True or False was the immediate inspiration for the research reported here. In fall 2022, we co-taught an MIT undergraduate class (STS.036: “Science in American Life”), in the course of which our students undertook a pilot evaluation of this exhibit. In February 2023, the MIT Museum offered a 3-session evening class, entitled “Make a Fake,” in which MIT Media Lab graduate student Matt Groh—himself a co-author of Detect a Fake—taught a group of adult students the basic principles of generative AI, guiding them in both how to make and how to detect simpler deepfakes. Following the final session of this class, all of the students participated in a moderated discussion of deepfake, chaired by one of us. This recorded conversation became the first Focus Group (“FG#1”) of the present study. It was then followed up by two further FGs, one involving participants recruited from among general visitors to the MIT Museum, and another recruited from members of the general public visiting a large shopping mall in downtown Boston. The rationale behind these choices is outlined in Section 3.

3. Methodology

For the present study we organized three FGs, each designed to bring together groups calculated to be at different social distances from deepfake technology. ² FG#1 comprised the students in the short course Make a Fake, which was held in the MIT Museum on 13th, 20th, and 27th February 2023. We term the graduates of this short course a group of the “engaged public” for deepfake technology. FG#2 comprised recruits from among the general visitors to the MIT Museum. We have known for some time that on average such visitors tend to be highly interested in and attuned to science and technology, and so we term this a group of the “interested public.” Finally, FG#3 comprised recruits from among general visitors to downtown Boston, many of whom were shoppers visiting the Prudential Center Mall. We term this a group of the “general public.”

There are several limitations to this study. The sample size (3 FGs) is admittedly small, though we believe it is justified by the ability to perform a comparative analysis of three significantly different publics. Again, the study is purely qualitative—no quantitative estimates can be obtained from it. This is partly because of the small sample size, and partly because we have chosen qualitative methods for the analysis. We agree with Lunt and Livingstone that the FG is a valid method of social research in its own right, and not (as Merton and others argued) a mere adjunct to other, primarily quantitative methods. Furthermore, we agree with these authors that the FG should be regarded as, “a simulation of various aspects of social communication” (Lunt and Livingstone, 1996). We see the FG as an admittedly contrived conversation that can reveal much about the ways in which particular social groups deliberate about particular topics. Finally, we should note that the three publics sampled here are not even approximately of the same size. Those who are “engaged” (FG#1) or merely “interested” (FG#2) in science are minorities; most members of the public would fall into our “general” (FG#3) category. This should be borne in mind in judging the significance of our findings.

In summary, the FGs were recruited as follows:

The self-stated personal and/or professional backgrounds of FG participants are listed in Supplemental Appendix A. As will be seen, the composition of the three FGs was markedly different: many members of FG#1 had clear professional reasons for wanting to know more about how to make deepfakes; many members of FG#2 had substantial educational experience and/or expertise in science and technology; and most members of FG#3 had no special relationship to deepfake technology in particular, or science and technology in general. None of the participants in FGs #2 and #3 knew about the topic of the conversation in advance. Care was taken to ensure that in each case the FG samples were diverse in gender and age (all participants were over 21 years of age).

Each FG conversation lasted 1 hour, and each was recorded on different devices in order to capture all voices across the rooms. The data were anonymized and transcribed under the review of both authors. The final transcripts of the three conversations can be consulted in the open access EU repository Zenodo (Denia and Durant, 2024), operated by the European Organization for Nuclear Research (CERN). A standard question protocol was followed in each group, with modifications only where these were required by the immediate context of recruitment (e.g. we began FG#1 by asking participants about their experience in the Make a Fake course). The FG Protocols are provided in Supplemental Appendix B. In the cases of FGs #2 and #3 only, a brief and neutral definition of deepfake technology was provided: “Deepfakes are digitally manipulated videos that show people saying or doing things they never really did.”

The transcripts were analyzed qualitatively, with a view to identifying the distinctive analytical categories (point of view, attitude, tone, media associations, and frames; see Section 2) present in each conversation. No attempt was made to quantify the results in a coding frame, as the sample size was considered too small to make this worthwhile. Instead, each author undertook close readings of the transcripts independently in a search for characteristic modes of discourse; and the results were then shared between us and found to be broadly similar. This resulted in the identification of a number of analytical categories that together served to highlight the most striking features of each conversation. In the results section, each distinctive analytical category is exemplified with multiple quotations from the relevant transcripts. Individual FG participants are referred to for purposes of quotation using the following convention: number of FG (1, 2, or 3), followed by participant number (see Supplemental Appendix A). Thus, for example, the third participant in FG#2 would be referred to as: P2.3, and so on.

4. Results

The three FGs succeeded in capturing three coherent conversations about deepfake technology. Judging from the transcripts, at the time that we conducted this research AI was very present in the US public mind. This being said, we wish to capture the extent to which these three conversations were both similar and different. By similarity, we have in mind for example that all three groups:

None of our FGs bought into the idea that deepfake technology was either wholly good or wholly bad.

On item 1, general awareness of AI: P1.5 said of ChatGPT that it “Sounds like a high school senior or a very beginning college level”; P2.5 said, “we can see that the ChatGPT can now answer all of our questions”; and P3.8 said, “ChatGPT. It’s everywhere today.” On item 2, the display of ambivalent attitudes: P1.9 said, “I can [detect a deepfake] until I can’t … and that’s the part that’s a little concerning. I mean, right now I think I do, but I don’t think that’s going to stay that way”; P2.9 said,

we can try to use it [AI] in healthcare, we can try to use it in finance and this and that, but we have this tool and it’s exciting and it’s got so many users, but it wasn’t driven by a problem looking for a solution. It was driven by “we have this technology now and we’re able to utilize this technology to predict things. So let’s try to figure out what we can do with it.”

On item 3, algorithmic bias, P1.12 said, “When we teach kids to look for bias, we usually focus on articles, but this opens a whole new venue [sic]”; P2.4 said, “.. .they [people] don’t realize that the data is trained on biased things”; and P3.10 said, “… humans might impose their bias on the AI system.”

So much for similarity between the focus groups. Table 1 captures some of the key differences between the three conversations.

OPEN IN VIEWER

Table 1.

Differences between the three focus group conversations.

Analytical category	Engaged publicFG#1	Interested publicFG#2	General publicFG#3
(i) Point of view	1st person (“Me”/“We”/“Us”)	3rd person (“They,” as in the sense of other less well-informed people)	3rd person (“They,” as in the sense of the authors and operators of AI)
(ii) Attitude	Generally positive	Generally ambivalent	Generally negative
(iii) Critical Evaluation and Tone	Highly nuanced: the vast majority express both positives and negatives, using emotionally restrained and moderately technical language	Moderately nuanced: half express a weighted opinion, in language that is less restrained, moderately emotional and quite technical	Not nuanced: the vast majority express purely negative views and a few purely positive ones, in language that is relatively unrestrained, often quite emotional and not very technical
(iv) Cultural/Media associations	Lower presence of cultural and mass media images	Moderate presence of cultural and mass media images	High presence of cultural and mass media images
(v) Frames	Social Progress	Deficit Model	Pandora’s Box

By “Point of View,” we mean the preferred pronouns (first person, second person, third person) used in FG conversations, as markers of overall stance of FG participants toward AI/deepfake. By “Attitude,” we mean the general way of thinking or feeling about AI that was expressed by the FGs. By “Critical Evaluation,” we mean the extent to which the FGs revealed a willingness on the part of participants to evaluate multiple perspectives on AI, and by “Tone” we mean the form of conversation characteristic of each FG—more or less restrained, more or less emotional, more or less technical, and so on. By “Cultural/Media Associations,” we mean the extent to which FGs reached for examples of AI from wider cultural sources (mass media, social media, movies, etc.). Finally, by “Frames” we mean prevailing perspectives that serve to identify what really matters in ways that resonate with core values and assumptions. Framing theory was originally developed in the social sciences (e.g. Entman, 1993; Scheufele, 1999), but it has been applied with some success to science and technology (e.g. Nisbet, 2009; Nisbet and Mooney, 2007). In what follows, we present results from the FGs under each of these analytical categories.

5. Discussion

Although there are some basic similarities between our three FGs, it is clear that their varied social distances from deepfake technology are associated with some systematic differences in the ways in which FG participants thought about this technology. To summarize these differences we offer the following brief caricatures of the three separate focus group conversations.

FG#1 was a conversation among novice deepfake practitioners. As budding practitioners will, the discussants behaved for the most part like insiders: they spoke about the technology in the first person; they spoke about it with a certain amount of confidence; they were alive to many potential benefits, but aware of the potential for misuse by bad actors; and they were also concerned about what others (by implication, outsiders) might think of the technology. But on the whole, this group was positive about the technology’s implications, and concerned more than anything about its own ability to keep up with the rapidity of technological developments in the field. The prevailing frame of reference for this exchange was “Social Progress.”

FG#2 was a conversation among learners who were keen to know more about this and other technologies. As learners sometimes are, they were aspirant insiders, aware of their own—and, especially, others’—shortcomings, and concerned above all about how such shortcomings on the part of others might lead to misapprehensions and misunderstandings that would stand in the way of progress. Theirs was not just talk about technological and social progress, it was also talk about the ways in which alarmist and ill-informed speculation among people less knowledgeable about AI/deepfake than themselves might stand in the way of the potential benefits that this technology might bring. For these reasons, we identify the predominant frame of reference for this conversation as “Deficit Model.”

FG#3 was a conversation among people who, for the most part, felt themselves to be outsiders. Far and away the most negative of our three samples, this group worried a lot about the question: what might “they” do next with a technology over which “they,” and they alone, had any kind of control? And as the group pondered this question, so it conjured multiple, more-or-less familiar and more-or-less negative images from mass media and movie representations—images of politically compromising deepfakes of former UK Prime Minister Boris Johnson, images of mischievous deepfake videos circulating on TikTok, and images of movie representations of technologically inspired disasters, such as Jurassic Park. Here, indeed, the predominant frame of reference was what is known in the literature as “Pandora’s Box.”

It will be worth paying more attention to what we have termed “Point of View” than is customary in research on public attitudes to science. Choices made among personal pronouns are revealing of individuals’ and groups’ assumptions about their relationships to the worlds of science and technology. Do individuals perceive themselves as active members of these worlds (first person), aspirant or would-be members of these worlds (third person, with the object of attention being other, less well-informed members of the public), or passive spectators or even victims of these worlds (third person, with the focus on science and technology as remote sources of authority)? Our results here appear to connect with wider sociological literatures on “self-efficacy,” which have emphasized the importance of subjective feelings of competence, effectiveness, and causal agency for many life outcomes (e.g. Gecas, 1989). It will surely be worth further research to determine whether, as we suspect, the very language in which members of the public capture their thoughts and feelings about technological issues is itself revealing of different levels of self-efficacy with respect to science and technology.

Our deployment of framing analysis suggests that members of the public who are at different conceptual distances from science and technology may frame these subjects in quite different ways. “Social Progress” and “Pandora’s Box” are alternate frames that have relatively widespread currency within contemporary American culture, and it is not surprising that they apply to some of the FG conversations described. Hitherto, however, it has been widely assumed that the so-called “Deficit Model” is characteristic of special groups, including many science communicators. The results reported here suggest that this model may apply to a much larger section of the public, namely what we have termed the interested public for science and technology. The implications of this finding for science communicators whose work may be targeted principally at precisely such interested publics are surely worthy of further exploration.

Our FG participants viewed AI/deepfake from three very different conceptual distances, and as they did so they saw three significantly different things. For one thing, those closer to the technologies appear to have had more nuanced assessments of the impact of these technologies on wider society. Also, there is evidence that those closest to AI/deepfake (FG#1) preferred their own personal—and generally positive—experiences over the less personal and generally more negative social images that were available to them in mass media and the movies; whereas those farthest away from AI/deepfake (FG#3), having far fewer personal experiences on which to call, relied much more heavily on the social images of AI with which they were at least generally familiar. Since these social images were prevailingly negative (not least, in our participants’ perception of them), the overall stance that FG#3 participants took toward deepfake and related technologies was correspondingly negative.

We should be careful not to allow these characterizations to pass into judgments—as, for example, that attitudes in FG#1 were more “discerning” or more “rational” than those in FG#3. Given the general levels of social anxiety about the wider ethical and social impacts of Artificial Intelligence in the 2020s, it could as well be argued that the more positive attitudes of FG#1 were overly influenced by narrower personal and professional interests, while the more negative attitudes of FG#3 represented a wider view of larger concerns. The point here is not to judge the various attitudes displayed from some supposedly independent perspective, but rather to understand what these different attitudes may tell us about the relationship between social distance from and attitudes toward technology among different segments of the public.

The general verdict of this study is that in the case of deepfake technology greater social distance did not lend greater enchantment. On the contrary, in the United States in 2023 social distance from AI/deepfake appears to have served to multiply and reinforce various forms of disenchantment. One form of disenchantment was an increasing level of distrust in the motives and intentions of the authors of these technologies. But above all, greater social distance allowed prevailing social images and representations to shape our participants’ responses to a technological area of which they had little firsthand experience, in which they were not substantially invested and over which they felt they had no meaningful control. Conversely, Focus Group #1 comprised people who felt themselves to be participants in deepfake technology, albeit of a rather junior sort; and with this sense of participation came corresponding senses of confidence and optimism about the potential for the constructive use of this technology. In this particular instance, it would appear that to “know” the technology was indeed to “love” it.

If there is a general conclusion to be drawn from this study, it is not that social distance lends disenchantment with science and technology; rather, it is that social distance decreases the influence of personal experience and increases the influence of social images and representations in the formation of attitudes to science and technology. This idea is consistent with the finding from survey studies that the relationship between knowledge and attitudes in relation to specific science- and technology-related issues is complex and issue-dependent (see Section 1); also, it is an idea that is open in principle to further empirical exploration. For example, we have dealt here with the relationship between social distance and attitudes in relation to a specific area of science and technology where the social images available to our focus group participants were mainly negative. To put it bluntly, media coverage of AI in the United States in the first half of 2023 was extraordinarily negative—at times, even apocalyptic—and it is not surprising, perhaps, that members of the public for whom these were the chief sources of knowledge would have generally negative attitudes.

What, though, if we were to repeat this study in a very different scientific or technological issue area, where prevailing social images were generally positive? Here, our general conclusion would predict that social distance would indeed lend enchantment, as the influence of prevailingly positive social images and representations increased with increasing social distance from the relevant issue area. This is an empirical study that could in principle be undertaken. However, the challenges facing researchers in such a study should not be underestimated. In order to conduct this kind of study, it will be necessary first to select an appropriate scientific or technological topic. To be suitable for purpose, this topic will need to be: (1) socially salient (so that most members of the public are at least generally familiar with it); (2) socially accessible (so that some members of the public have a much closer relationship to it than others); and (3) the subject of generally positive social images and representations.

Identifying a topic that fulfills all of these criteria will be no easy task. In a previous era in the United States, one or more aspects of space exploration might have been suitable. Today, perhaps some areas of astrophysics—research on black holes, for example, or the search for life elsewhere in the universe—might fit the bill, though there are questions to be asked about quite how socially salient such research really is. Alternatively, we might look for a topic from within the general area of medical science. (Were it not for the fact that it has become increasingly politicized in the United States, an obvious candidate topic here might have been the recent development of COVID-19 vaccines.) We would welcome other candidate topics. Whatever the topic chosen, further research must be timely, in relation to scientific developments and public representations, and well-targeted, in relation to focus group composition. We hope that such research will be undertaken.

Supplemental Material