Deepfakes are a type of synthetic or artificial media. By training AI on datasets of images, deepfake technology allows people to create photorealistic fake audiovisual materials. Sometimes these videos depict real people; sometimes they depict artificial people. Deepfakes have provoked philosophical interest in recent times, in part because of the challenge they seem to pose to social epistemic practices, and also because of the significant ethical issues they raise. Is it permissible to create a deepfake video of a real person? If not, why not? If so, under what conditions is it permissible?
These ethical issues are particularly important given the main use of deepfakes. According to one widely-cited Dutch study, 96% of all deepfakes are pornographic in nature. Deepfake technology is commonly used to create videos depicting famous actresses performing sex scenes. Some of the actresses that have been subjected to this include Gal Gadot, Scarlett Johansson and Jennifer Lawrence among many others. Deepfakes are also widely used to create revenge porn depicting ex-partners in sexually compromising positions. Some types of deepfake porn have even been used in political debates and scandals. For example, there have been famous cases of deepfakes (or suspected deepfakes) being used to undermine political opponents in Malaysia and India.
When people hear about these cases, their reaction is often one of intuitive disapproval. They think there is something deeply wrong about the creation of involuntary deepfake porn. If a person has neither consented to have sexual imagery of themselves recorded, nor voluntarily participated in videotaped sexual activity, what could possibly justify or permit their image being involuntarily used to create fake imagery? Surely nothing?
Philosophers, however, sometimes like to ask obvious questions like this to see if they raise any dilemmas in our moral thinking. In his paper ‘Introducing the Pervert’s Dilemma’ (an unfortunate title for reasons I will get to later), Carl Öhman does this for deepfake pornography. In essence, he suggests that if we think the creation of deepfake pornography is morally impermissible, then we should also question the permissibility of generic sexual fantasies. This raises something he calls the ‘Pervert’s dilemma’. Öhman suggests a way out of this dilemma by using the concept of a ‘level of abstraction’ (LoA) to assess different moral phenomena.
Öhman’s paper is a provocative one, and his use of levels of abstraction (LoAs) to resolve the dilemma is certainly worthy of greater consideration, but I’m not sure I agree with key parts of his analysis. I want to explain why in what follows.
1. What is the Dilemma?
The starting point for Öhman’s article is his articulation of the pervert’s dilemma. Here is the basic idea.
Most people have sexual fantasies. In these fantasies, they imagine other people, usually real people that they know or have seen, in sexually explicit scenes. Sometimes they imagine that they are having sex with the object of their fantasy, and sometimes they use these sexual fantasies for self-stimulation. For the most part, we think of these sexual fantasies as a normal and permissible part of human life.
That said, not all sexual fantasies are morally kosher. If a person regularly imagines having sex with a child or fantasises about raping someone, we might call that into question. Such a fantasy might be thought to express a morally problematic desire. This issue is complicated and depends to some extent on what exactly the person is imagining. In any event, even if we agree that it is morally problematic to have such fantasies, it does not follow that we can or should do anything about it. In particular, it does not follow that we should use the law to police people’s thoughts. We might just think it is morally problematic and people that have such fantasies should critically reflect on their moral significance and, if possible, try to change the content of their fantasies. None of this, however, affects the point that generic or ‘plain-vanilla’, sexual fantasies are thought to be morally permissible.
Here then is the dilemma: What if instead of using their imagination to construct a plain-vanilla sexual fantasy, a person creates a deepfake porn video instead? Does the fact that it has been turned into a material representation make a moral difference? Öhman argues that it should not because it is not obvious how materiality can make a significant moral difference. And yet it seems, when pressed, that we would find something morally problematic about creating the video that we wouldn’t find morally problematic about the fantasy. How can we reconcile these differing moral judgments? That’s the essence of the pervert’s dilemma.
Let me say two things about this dilemma before proceeding. First, Öhman’s claim that turning a fantasy into a material representation is not morally significant seems dubious. Surely there is something morally significant about this? When you create the deepfake, you create a file that is easily shareable and, potentially, weaponisable. We see evidence of this all the time. Deepfake porn videos are both (a) widely shared on the internet and (b) often used to humiliate, degrade and intimidate people (typically women). This shareability and weaponisation is part of what makes them morally problematic and, at least in my own case, is part of motivates me to think they are morally impermissible.
Öhman, to be fair, is fully aware of this point. He sharpens the dilemma by asking us to imagine a case involving a purely private, non-shareable, deepfake. So suppose someone creates a deepfake porn video of someone they know for private use only and there is no possibility of it ever being shared or weaponised or, indeed, of the person whom it depicts ever finding out about it. Would we still find it problematic under those conditions? Öhman argues that we would and this means the dilemma is still relevant.
The problem with this is that it means we are now debating the moral permissibility of a hypothetical technology. I’m not aware of any deepfakes that do not have the potential to be shared nor ones that can be confirmed to be purely private in nature. Furthermore, if such a technology existed, I’m not sure what my moral intuitions about it would be. If you remove the major potential consequential harms of deepfakes, I think I might find them less problematic than I currently do. Debating hypothetical technologies like this can be fun, and might reveal something important about our moral beliefs and practices, but it has limited relevance for the real-world.
The other point I wanted to make about the dilemma is that I think it is badly named. Öhman is drawing inspiration from Morgan Luck’s ‘Gamer’s Dilemma’ (which I have discussed previously). The gamer’s dilemma is about the permissibility of virtual child sexual abuse vis-a-vis virtual murder. But calling his dilemma the ‘pervert’s dilemma’ seems unfortunate to me since the term ‘pervert’ carries negative moral connotations. Ordinarily, when we use the term ‘pervert’, we use it to describe someone with morally problematic sexual attitudes and practices. The whole point of the dilemma is to highlight a tension between something we deem non-problematic (plain-vanilla sexual fantasies) and something we deem problematic (non-consensual deepfake porn). Using the term ‘pervert’ to describe the dilemma confuses things because it suggests that this is only a dilemma for someone with problematic sexual attitudes. A morally neutral title (e.g. the fantasiser’s dilemma) might have been better. That said, I don’t know what the best alternative title would be.
2. Using Levels of Abstraction to Resolve the Dilemma
Set all that to the side. Suppose we accept Öhman’s analysis and focus on the permissibility of non-shareable private deepfakes vis-a-vis sexual fantasies. Is there a way to resolve the dilemma? Öhman argues that there is if we use the idea of levels of abstraction (LoA).
This is an idea taken from the work of Luciano Floridi, which was itself inspired by formal methods in computer science. In essence, the idea is that different objects and entities can be described at different LoAs. Take two human beings: Dave and Emma. You can describe them at the LoA ‘HUMAN BEING’, in which case you focus on attributes they share qua members of the human species. Or you can describe them at the LoA ‘FAMILY’, in which case you focus on their attributes as members of the same family (or not, as the case may be). The former is a ‘higher’ and hence broader LoA because it ignores many properties (variables, observables) that are distinctive of Dave and Emma in their everyday lives. The latter is a ‘lower’ and hence narrower LoA since it adds in additional properties that are more distinctive and unique to Dave and Emma and not just shared across all humans.
Although this example covers the comparison of individual human beings at different LoAs, it should be noted that the method of LoAs applies more generally to all forms of comparison and analysis. The basic gist of the idea is that, for any given question or inquiry, there are more or less appropriate LoAs at which to try to answer that question.
This has implications for morality. It is common in moral philosophy to focus on the permissibility of actions at high levels of abstraction. Philosophers often discuss the permissibility of our actions qua individual moral agents, abstracted away from our social contexts and roles. Consider, for example, the normative principles underlying utilitarianism or Kantianism. These are highly abstract principles, focusing on our duties as agents and/or sentient beings. The problem with this style of analysis is that when we consider our duties at lower LoAs (qua citizens in the same nation or members of the same family), we might find that different duties apply. Indeed, this is a common critique of highly abstract, impartialist moral principles: that they overlook our strong partial duties given the particular social roles and contexts we occupy.
In addition to this, Öhman points out that actions that seem harmless or permissible at high LoAs might be harmful at lower LoAs. For example, hate crimes might be of this form. Considered as an interaction between two abstracted agents, an incident giving rise to hate crime might not be morally harmful or problematic. The victim of a racial epithet, for instance, might brush it off or ignore it or not find it upsetting (the opposite could also be true). But considered at the LoA ‘ETHNICITY’ (or some equivalent) it might be harmful or problematic. The use of the epithet might harm or oppress all members of a given ethnic group even if it does not harm a specific individual. It is also possible that the individual acts, considered at the most abstract LoA, might combine or emerge into something problematic at a lower level.
Öhman thinks that this points the way to a solution to the pervert’s dilemma. In brief, he argues that when considered at a very high LoA, the creation of non-shareable deepfakes might not be morally problematic: the moral agent that created the deepfake (A) might not harm or injure the person represented in the video (B). This makes it, effectively, the equivalent to a private sexual fantasy. But considered at a lower LoA, it is problematic because the creation of the video is made possible by, and feeds into, a system of gender-based inequality and oppression. I’ll quote from Öhman’s article on this point (please note that this quote has been modified to remove some technical terminology that Öhman uses in his article but that I have not explained in this article):
For by abstracting the Deepfake phenomenon into a matter of “A” and “B”, one also subtracts from it the very thing that gives it its ethical significance namely its role in the social system of gender oppression. In one sentence, you cannot take gender out of pornography, and you cannot take society out of gender. As a societal phenomenon, Deepfakes are arguably enabled by a [combination] of male consumers, producers, technology, and misogyny. Moreover, it arguably plays a role in the machinery which systematically reduces women (as a collective identity) to sexual objects, even if none of the individual instances can be held to cause this. So it should be fair to say that the phenomenon is highly gendered (indeed, one need not spend much time on one of the forums or websites devoted to Deepfakes to realise this). While each isolated video may not affect the women it stars as individuals, the phenomenon as such…is, in its current form, inseparable from the systematic degrading of women as a collective identity.
Contrariwise, he claims, sexual fantasies are not like this. While the content of specific sexual fantasies might be inspired by a system of gendered inequality, the general phenomenon of sexual fantasies is not part and parcel of such a system. It is, rather, a common and normal part of human experience.
3. Is this a plausible resolution of the dilemma?
Öhman anticipates and responds to two critiques of this line of reasoning. One of them is that his analysis is just a sophisticated way of saying that the ethical permissibility of non-shareable deepfakes depends a lot on social context and meaning of deepfakes, and hence doesn’t say much of use at all. Öhman is, I think, rightfully dismissive of this critique. His point in using LoAs is that the moral assessment of actions requires some preceding sociology of those actions. This is a fair point and has been made by other philosophers. Purely abstract moral reasoning often doesn’t tell us much.
The other critique is more interesting. It is that his analysis of the impermissibility of deepfakes could apply equally well to other forms of media representing sexual fantasies. He gives the example of a man that draws crude pictures of sexual fantasies instead of using photorealistic deepfakes. Is this morally impermissible? Öhman says that it will depend on the social meaning of such images. He thinks it is unlikely that they have the same meaning as deepfakes:
I believe the answer to this question must be sought in the cultural role of the phenomenon of drawing pornographic images of women one has met. To my knowledge, this is not a common practice used in gender oppression in today’s society, but in a hypothetical society, it certainly could be.
I think this is a bit too quick. When we peel back the sophisticated layers of analysis, it seems to me that Öhman is effectively repeating a very common (radical) feminist critique of all pornographic media, namely: that even if the production of particular pornographic media does not harm anyone involved in its production (and possibly may even benefit them), the media contributes to (or maybe even constitutes) a form of gender-based oppression. In this regard it is telling that Öhman cites the well-known feminist critic of pornography — Gail Dines — to support his analysis of the social meaning deepfakes. So I think his critique could easily apply to other forms of media, possibly including drawings or sketches of sexual fantasies since the very reason why a man might draw such images could be, if we follow the critique, because of some toxic, partriarchal ideology he has imbibed..
What should we make of this? Well, I’ve written a lot about the ethics of pornography on this website before, assessing this common critique from multiple angles. In brief, I would say that the problem with this critique is twofold: (i) it often requires us to overlook, ignore, or explain away examples of pornography that genuinely do not seem morally problematic or harmful to either producers and consumers (and thus it generates a tension between two beliefs about these practices) and (ii) it ignores the fact that the social meaning or significance of pornographic production is not fixed or set in stone. It can change and, indeed, many sex positive producers of ‘feminist pornography’ claim that it may already be changing. Furthermore, when it comes to the social meaning of pornography, it is difficult to completely divorce this from its potential social consequences. If pornography is really harmful to women, and clearly contributes to their ongoing oppression and inequality, then the social meaning of any particular instance of pornographic media is more likely to be negative. If pornography is beneficial or if its effects are more doubtful or unknown, then the picture is less clear.
What does this mean for deepfakes? I think it means two things. First, the analysis that Öhman offers of non-shareable deepfakes is quite similar to, and probably subject to similar critiques as, more general critiques of pornography. Second, it might be harder to defend this analysis if we limit ourselves to the purely hypothetical example of non-shareable deepfakes: whether or not deepfakes are part of a system of gender oppression and inequality depends on how they are shared and weaponised.
This brings us back to the point I made earlier in this article: that debating the merits of a hypothetical form of deepfake porn may have little practical utility.