Deepfakes and democracy: a case for technological mediation

Introduction ¶

Recent successes in the production of so-called “deep fakes” sparked both the imagination and the fears of many. The word “deepfake” is a contraction of “deep learning” and “fake”, indicating the use of Artificial Intelligence (AI) to synthesize images and videos that are not real, while simultaneously not or barely being recognizable as fabricated. For example, the recently launched website thispersondoesnotexist.com [12] by Philip Wang showcases AI-generated non-existing faces that are extremely realistic. Notably, the underlying neural network technique based on Generative Adversial Networks (GANs) is published [8] and publicly available - including code - to those who are interested in implementing similar applications. Currently, an app called FakeApp is being developed with the goal to make the “technology available to people without a technical background or programming experience.”[2]. At the same time, there are serious concerns that as this technology becomes even better, not only images but also videos can be completely faked. In the current state-of-the-art it is already possible to “face swap” existing faces in videos, allowing for example the face of President Trump to be inserted in an arbitrary video. Despite leading to some very entertaining videos, this technology is simultaneously a next step in the production of fake news and has the potential to thoroughly disrupt democratic discourse.

In this essay I first highlight main threats of deepfakes to democratic discourse. I claim that what these threats have in common is that they result from a deepfake’s potential to mediate what we perceive to be “real”. Secondly, I discuss how awareness of these negative societal consequences elicits different stances towards the underlying AI-technology, in particular concerning the responsibility that developers have in openly publishing (or not) these technologies. Thirdly, I argue that a philosophy of technological mediation is not only an adequate framework for understanding how deepfakes threaten democratic discourse through mediating what is “real”, but also for expressing the full complexity of the question who is responsible for negative societal consequences.

Deepfakes disrupting democratic discourse ¶

Societally undesirable applications of deepfake technology have already emerged, and more negative consequences are anticipated to emerge as the technology matures. One major negative application threatening individuals is the creation of fake porn videos of celebrities, which are now actively being banned from reddit and porn sites as they amount to non-consentual porn [2] [1, p.18]. But on a societal level, there are major concerns that deep fakes might significantly disturb the type of political discourse that is essential for democracy to function. Bobby Chesney and Danielle Citron [1] are the first to extensively explore the relationship between deep fakes and democratic discourse. Deepfakes first of all enlarge threats to democracy that are already present in what some consider to be a “post-truth” era, in which fake news can be as effective for achieving political goals as actual news based on facts. This threatens democratic discourse because, as Chesney et al. adequately express: “One of the prerequisites for democratic discourse is a shared universe of facts and truths supported by empirical evidence. In the absence of an agreed upon reality, efforts to solve national and global problems will become enmeshed in needless firstorder questions like whether climate change is real. The large scale erosion of public faith in data and statistics has led us to a point where the simple introduction of empirical evidence can alienate those who have come to view statistics as elitist.” [1, p.21]. Deepfakes in this sense contribute to what Chesney et al. call intellectual vandalism in the marketplace of ideas [1, p.21]. That development is undesirable for democracy irrespective of its particular form, but is particularly worrisome for those supporting a pluralist or deliberative democracy, as they see opinion forming in a free and open dialogue or debate as essential to democracy [9, 4-5].

But secondly, deep fakes can even more effectively undermine fair and democratic intellectual competition in this marketplace of ideas than “normal” fake news does. Imagine a deepfake video spreading on the evening before elections, showing one of the candidates committing a serious crime. Due to the power of social media such a video can go “viral” and do serious damage to the eligibility of a candidate. In modern media “not guilty until proven otherwise” often hardly holds, and one can be convicted in the public eye for a crime that was not committed, without fair trial. A well timed deep fake can heavily disrupt fair democratic elections in this manner before there is a chance to debunk the deepfake. But even if a deepfake is exposed as false, its disruption of fair elections can still be effective by having set a cognitive bias in the minds of the electorate [1, p.19].

Using deepfakes to disrupt democratic discourse will be even more effective if they target situations that are already extremely tense. Imagine for example a deepfake of “an Israeli official doing or saying something so inflammatory as to cause riots in neighboring countries, potentially disrupting diplomatic ties or sparking a wave of violence.” [1, p.20]. Once such a situation is escalated, despite the cause being “fake news”, it is extremely hard to de-escalate them. In contexts where such distrust is already present, deep fakes can further erode trust in institutions of open democratic discourse. As Chesney et al. point out, in such tense situations the likelihood that opposing camps will believe negative fake news about the other side is higher, and only increases as deepfakes exploit this mechanism to further enlarge social divisions [1, p.23]. Not surprisingly, techniques to detect deepfakes are being developed to counteract these risks, for example by the US military DARPA [4]. But due to the flexibility of GAN neural networks it is likely that whatever technology is developed in detecting fake videos might also be used as a feedback mechanism, ultimately only improving the quality of deepfakes [4]. These examples show that combatting the threats of deepfake technology to democracy cannot be an exclusively technological story. Despite technological counter-measures, deepfakes still threaten democracy by setting cognitive biases and eroding a commonly agreed upon reality that serves as the background for a meaningful democratic dialogue. I argue in this essay that the mentioned threats to democratic discourse are grounded in a deepfake’s potential to mediate what humans perceive to be “real”. Furthermore, through mediating what is “real”, deepfake artefacts can co-determine human praxis. Because of how fundamental this theme is, I think we also need a philosophical story to understand the impact of deepfakes. In the following sections I first explore two diametrically opposed ways of coping with the societal impact of deepfakes. I then show how a theory of technological mediation is an appropriate philosophical framework for understanding this impact, and moreover that it is able to grasp the complexity of the question how to bear responsibility for it.

Stances on technological disclosure ¶

When one develops a technology that has a large societal impact, a quite fundamental ethical question is to what extent the developer is responsible for that impact. Philip Wang of thispersondoesnotexist justifies promoting the GAN technique used for deepfakes in an interview by pointing out that those “who are unaware are most vulnerable to this technology” [7]. This taps into what can be called a deterministic view on technology, which lets societal necessity follow quite automatically from technological potentiality with the motto: “if it can be done, it will be done”. In the field of AI deterministic attitudes are well represented as AI-technology is increasingly changing society. To the deterministic-minded person even those who worry about these societal changes and remind us of the dangers, are nevertheless equally subjected to the great historical impetus of technological progression. And this person then reasons: if the technology will emerge in society at some point in any case, then the best thing we can do now is raise awareness. In this way we, as a society, can adapt to the technology - rather than adapting the technology to human needs.

Other developers of AI-technology share the concerns for its potential negative societal impact, but conceive of their own responsibility differently. For example, the OpenAI research organization, which is dedicated to making sure AI benefits humanity, announced last month that they developed an AI that can write paragraphs of text that “feel close to human quality and show coherence over a page or more of text” [6]. However, contrary to the publications about video deepfakes, the OpenAI organization decided not to release the used datasets, nor the trained model or the used code, due “to concerns about large language models being used to generate deceptive, biased, or abusive language at scale” amongst other “malicious applications of the technology” [6]. They did however release a smaller trained model with less potential for abuse in order to still display the technical innovations that “are core to fundamental artificial intelligence research” [6]. As scientists, they do not want to counteract progression of the field. This experiment in responsible AI disclosure amounts to a more instrumentalist view on technology: its development is controlled by humans, instead of being an autonomous deterministic force to which humans have to adapt.

The primary hope of the decision to withhold the AI is that this will give the AI community as well as governments more time to come up with ways to prevent or penalize malicious use of AI technologies, quite similar to the practice of responsible disclosure in cryptography, where organizations are given time to repair security weaknesses before they are publicized. Interestingly, OpenAI’s explicit concern for the societal impact of their technology is framed in the context of political actors waging “disinformation campaigns” by generating fake content, requiring that “the public at large will need to become more skeptical of text they find online, just as the “deep fakes” phenomenon calls for more skepticism about images” [6]. In their policy OpenAI thus explicitly respond to the media attention surrounding deepfake neural networks that become better at deceiving people and are increasingly publicly available. Although not free of some hint of determinism, the OpenAI initiative exerts a responsibility for actively controlling technological development in AI, to make sure that it brings forth useful instruments that are to the benefit, and not the detriment of humanity.

The contrast in the positions between a) the open publishing of deep fake technology including trained models and code, and b) the controlled disclosure of text-generating networks, again shows that the development of these technologies does not only raise technical issues, but also societal ones. In both cases, the researchers are aware of the societal dangers of their technology, but take responsibility for it in different ways. In a deterministic vein, there is no reason to control disclosure of technology: someone else will publish it anyways, and it is better to inform people as soon as possible. From a more instrumentalist point of view, the act of disclosure is not as neutral: since humans have at least some control over technology, they also share responsibility for possible negative consequences within reasonable limits. After all, the technology itself is just a neutral instrument. Whether it is put to good use depends on humans.

Both views have in common that they conceptualize the human-technological relationship in terms of a subject-object divide, in which subject and object are external to each other, irrespective of whether the subject is human or some technology. But I think that these terms are no longer sufficient for understanding the complexity of deepfakes that heavily blur the demarcation between what is “real” and what is not, and consequently also not sufficient for understanding how this is the foundation of a threat to democracy. Accordingly, if we are to conceptualize the responsibilities of developers of such technologies, we need to take into account how these technologies mediate reality and human praxis.

Deepfakes and Technological Mediation ¶

In this section I argue that the philosophy of technological mediation as put forward by Verbeek [11][10] is appropriate for conceptualizing the threat of deepfakes to democracy in terms of their mediation of human praxis. Technological mediation “concerns the role of technology in human action (conceived as the ways in which human beings are present in their world) and human experience (conceived as the ways in which their world is present to them)” [10, p.363]. That technological artefacts mediate means that they “are not neutral intermediaries but actively coshape people’s being in the world”, and that they do so in two directions: they mediate how the world appears to humans (perception) and how humans give shape to their own reality by acting in the world through the use of technological artefacts (praxis) [10, p.364]. The mediation of deepfakes can be shown in both directions, and I will indicate how they are interrelated in the example of democracy.

First of all, what the name “deepfake” expresses is that a given image or video is perceived to be “real”, while what is represented does not exist in the represented capacity: i.e. it is “fake”. I chose this specific formulation because a deep fake of Trump does not necessarily mean that Trump does not exist, but merely that he did not say or do what is represented in the deep fake video or image.

Now imagine a video of a man committing a serious crime, with the face of Trump swapped in. In case of a successful deepfake, we do not see a man with Trump’s face superimposed. Instead we perceive this man as Trump. The “as” in that sentence indicates an important insight from hermeneutic philosophy: the beings in our world always already appear to us as meaningful in a quite practical sense. The stereotypical example, based on Heidegger’s early philosophy, is that we see a hammer not as a composite object with one wood handle and one metal head, but intuitively and immediately take it as something we can hit nails with [10, cf. p.364]. Philosophical hermeneutics regards this as an act of interpretation that is not some scholarly exercise, but one that quite fundamentally determines how beings become present to us in the context of a world [cf. 5]. The particularity of deepfakes is that their technology mediates this process by making us pre-reflectively take something “fake” as something “real”. What is important is that, against instrumentalism, a deepfake’s deceiving character is not simply due to the bad intention of its designer. The technology itself is not a completely neutral tool in the theory of technological mediation. As it helps to shape what counts as “real”, this technology quite fundamentally sets a horizon for human moral and/or political action. Instead, mixing up fiction and reality is a core feature of the GAN technology that actively influences the relationship between a human and its world. A deepfake can thus be said to have its own “technological intentionality” [10 p.456] that affords (not causes!) the interpretation of “fake” as “real”.

But against determinism, this technological intentionality does not imply that the technological artefact autonomously decides our social realities, as if the technological artefact takes care of its own interpretation. As Verbeek makes clear, following Don Ihde, this technological intentionality only takes form in the interaction with humans [10 p.456]. Stating that technological intentionality does not coincide with human intentionality is analogue to the hermeneutic insight that the meaning of a text is not equal to the intention of its author. Despite this independence from the author’s intention however, it is equally naive in hermeneutics to say that the meaning of a text resides solely in the text itself as some pure ideal content, which would then be the same and equally complete even if nobody ever read it. Instead, and herein lies the analogue, a text’s meaning unfolds in the interaction with a reader. With respect to deepfake technology, this also means that its effects cannot be fully predicted independent of any real world interaction of humans with deepfake artefacts. I argue that in this manner a deepfake mediates how we perceive beings in the world by affording an interpretation of the fake as the real. If effective, a deepfake is not seen as just a video, but as representing an event in the world as we perceive it around us. But this interpretative step is everything but neutral. If we revisit the example of a deepfake of Trump performing a criminal act, we can see that this does not only imply we perceive the criminal as Trump, but that it also implies we now might perceive Trump as a criminal. We can then see how the hermeneutical effect of deepfakes underlies its effects in praxis:

If the fake is interpreted as real, then the real is reinterpreted in terms of the fake.

So if a candidate for a democratic election is shown in a deepfake to perform e.g. criminal acts (something fake is interpreted as real), then this candidate is potentially reinterpreted and reassessed by citizens as if he were a criminal (the real interpreted in terms of the fake). The aforementioned cognitive bias could also be interpreted along these lines: it is a re-valuation of something in the world because the deepfake artefact meddled with the interpretative process by which we take something as something.

Deepfakes thus contribute to the further blurring of the demarcation between real and fake news. As a result, even real and genuine discourse can become suspect, as it is now fair game to the question “fake or real?” But can we then still establish what we said was necessary for democracy? Can we in the future still have the certainty of an agreed upon reality, on the basis of which we can have a meaningful dialogue in the marketplace of ideas within a democracy?

Conclusion ¶

I have argued that the threat of deepfakes to democracy can be framed in terms of technological mediation, as we have regarded serious threats to democracy as a result of interpreting something fake as real. That means that deepfake technological artefacts can mediate both the (hermeneutic) experience of the surrounding world, and the actions humans take in it. But the perspective of technological mediation only makes the question who is responsible for (unintended) negative consequences more complex. One the one hand, developers of these technologies cannot be held fully responsible for negative consequences of technology, because they cannot fully predict how the interaction with users works out. But neither can developers realistically waive all responsibility by claiming that the development of AI is a historical movement shaping our social realities, independent of human interaction. Instead, when AI increasingly changes our social and political reality in unexpected ways, the more accurate position is admitting that somehow responsibility is distributed between developers, the technology itself, and its users. And especially if AI systems take on more autonomy in the future, the question of sharing responsibility with moral machines becomes increasingly urgent and intriguing.

Although such an open conclusion is not satisfying, it is the more honest position. When it comes to the moral responsibility (rather than a more limited legalistic story), issues around deepfakes can join the ranks of complicated ongoing debates about ethical responsibility in accidents with self-driving cars, or killer drones. The unresolved paradox is that unforeseen negative consequences may occur due to the learning capacity of AI, whereas at the same time, this flexibility is intended and exactly the main innovation of state-of-the-art AIs. And yet, we can reasonably ask of developers to foresee certain undesirable applications of their technologies. From the viewpoint of technological mediation both the stances of Philip Wang and of the OpenAI foundation have their own place. The decision of OpenAI to withhold their AI technology results from a reasonable anticipation of negative consequences, awaiting further democratic discussion before full disclosure. At the same time, this attitude should not tip the balance towards censorship. Withholding a technology from society in order to protect democracy seems paradoxically undemocratic and patronizing if not based on a sustained debate. Informing the general population about the threats of a technology is also desirable, but should not depart from a deterministic motivation. It is good, not because we have to learn to adapt to an uncompromising technology, but to spark a democratic debate with all involved stakeholders about how to design a better interaction with the technology [cf. 10].

References ¶

[1]: Robert Chesney and Danielle Keats Citron. Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security. SSRN Electronic Journal, 2018. doi: 10.2139/ssrn.3213954.

[2] Samantha Cole. We are truly fucked: Everyone is making ai-generated fake porn now, 2018. URL https://motherboard.vice.com/en_us/article/bjye8a/reddit-fake-porn-app-daisy-ridley. (accessed: 2019- 03-21).

[3] Maarten Franssen, Gert-Jan Lokhorst, and Ibo van de Poel. Philosophy of technology. In Edward N. Zalta, editor, The Stanford Encyclopedia of Philosophy. Metaphysics Research Lab, Stanford University, fall 2018 edition, 2018. URL https://plato.stanford.edu/archives/fall2018/entries/technology/. (accessed: 2019-03-27).

[4] Will Knight. The defense department has produced the first tools for catching deepfakes, 2019. URL https://www.technologyreview.com/s/611726/the-defense-department-has-produced-the-first-tools-for-catching-deepfakes/. (accessed: 2019-03-23).

[5] C. Mantzavinos. Hermeneutics. In Edward N. Zalta, editor, The Stanford Encyclopedia of Philosophy. Metaphysics Research Lab, Stanford University, winter 2016 edition, 2016. URL https://plato.stanford.edu/archives/win2016/entries/hermeneutics/. (accessed: 2019-03-27).

[6] OpenAI. Better language models and their implications, 2019. URL https://openai.com/blog/better-language-models.

[7] Danny Paez. 'this person does not exist' creator reveals his site's creepy origin story, 2019. URL https://www.inverse.com/article/53414-this-person-does-not-exist-creator-interview. (accessed: 2019-03-21).

[8] Timo Aila Tero Karras, Samuli Laine. A style-based generator architecture for generative adversarial networks, 2019. URL https://arxiv.org/abs/1812.04948. (accessed: 2019-03-21).

[9] Jan A. G. M. Van Dijk. Digital democracy: Vision and reality. Innovation and the Public Sector, 19:49-62, 2012. doi: 10.3233/978-1-61499-137-3-49.

[10] Peter-Paul Verbeek. Materializing morality: Design ethics and technological mediation. Science, Technology, & Human Values, 31(3):361-380, 2006. doi: 10.1177/0162243905285847. URL https://doi.org/10.1177/ 0162243905285847.

[11] Peter-Paul Verbeek. Mediation theory. 2019. URL https://ppverbeek.wordpress.com/mediation-theory/. (accessed: 2019-03-23).

[12] Philip Wang. Thispersondoesnotexist, 2019. URL https://thispersondoesnotexist.com/. (accessed: 2019- 03-21).