2. Kitcher’s argument and homosexuality

A first argument for banning research on the causes of homosexuality is analogous to Kitcher’s (Reference Kitcher2001) argument against research on race differences in cognitive ability. Kitcher’s argument goes that one should not pursue a line of research if that line of research is likely to give rise to beliefs that strengthen or stabilize convictions with morally undesirable social consequences. According to Kitcher, this is likely to be the case for research on race differences in cognitive ability regardless of the findings it would yield. Evidence that there are no differences would not harm the underprivileged, but due to existing biases is unlikely to be taken seriously or cause substantial changes. Evidence that the underprivileged have lower cognitive abilities, on the other hand, would strengthen the attitude that the underprivileged group is naturally unsuited for cognitively demanding social roles. Finally, indecisive evidence is likely to be harmful too, because human cognitive biases are such that we tend to interpret indecisive evidence as evidence that supports our prejudices. If society is already biased to believe that there are race differences in cognitive ability, indecisive evidence is likely to be taken as evidence for this hypothesis.

Kitcher’s argument is controversial (Talisse and Aikin Reference Talisse and Aikin2007). However, even if we assume that it holds, it is unclear whether it is applicable to research on the etiology of sexual orientation. Surely there are interesting similarities between such research and what Kitcher considers to be impermissible research. First, homosexuals are still an underprivileged group, especially in countries that continue to criminalize or stigmatize their sexual orientation. Second, harmful beliefs about the inferiority of homosexuals are still legion, for instance that they make bad parents, or that they are attention-seeking narcissists (Margolin Reference Margolin2021). Third, etiological research on homosexuality is such that it will often yield indecisive and hard-to-interpret evidence. For instance, heritability estimates vary dramatically between studies (Långström et al. Reference Långström, Rahman, Carlström and Lichtenstein2010), and it is notoriously difficult to causally interpret genome wide association studies (Tam et al. Reference Tam, Patel, Turcotte, Bossé, Paré and Meyre2019).

However, it is not clear whether another crucial condition is met; namely, the condition that directly links the research findings (or their interpretation) with undesirable attitudes and beliefs. With regard to cognitive differences research, it is obvious that if a person thinks that cognitive differences between ethnic groups exist, and that such differences are hard to alter through policies, this will often lead to the strengthening of that person’s harmful beliefs. However, the link between homonegativity and beliefs about the origin and nature of homosexuality is far less clear. Findings about the genetic or endocrine roots of homosexuality are regularly interpreted as indicating that homosexuals should not be held responsible for their sexual orientation. And indeed, Oldham and Kasser (Reference Oldham and Kasser1999) found that reading a vignette that focuses on the genetic and neuroendocrine causes of homosexuality, led to less homonegativity in students. Other studies, however, did not find this influence of etiological beliefs on homonegativity, and suggest that causality might go in the opposite direction (Bailey and Knobe Reference Bailey and Knobe2021): People often seem to rationalize their pre-existing attitudes by endorsing certain statements about the nature and origin of homosexuality. Moreover, even if it would turn out that findings about the genetic or endocrine roots of homosexuality would be instrumental in combatting homonegativity, this reasoning could itself be branded as a subtle form of homonegativity. After all, it does seem to suggest that we would blame homosexuals if they had chosen their sexual orientation, and that homosexuals would certainly have preferred heterosexuality if they had been given a choice.

Therefore, it is far from obvious whether, from a social justice perspective, it would be better or worse for homosexuals if people believed that their sexual orientation is under strict biological control. In short, Kitcher’s imperative not to pursue certain lines of research may work for cognitive differences research, but not for studies into the etiology of homosexuality.

3. Dangerous applications

Even if research into the etiology of homosexuality is not likely to produce beliefs that harm the underprivileged, it may still harm them in other ways. In particular, such research could form the basis of technology that could be used against the underprivileged. Take, for instance, research on how deep neural networks can be used to detect a person’s sexual orientation. This technology is at the heart of a recent paper by psychologists Wang and Kosinski (Reference Wang and Kosinski2018), published in Journal of Personality and Social Psychology. Wang and Kosinski’s research was inspired by the prenatal hormone theory of sexual orientation. Part of the evidence for this theory is that there is some correlation between sexual orientation and traits that are thought to be determined by intra-uterine hormone levels. For example, the exposure of androgens in utero is believed to have an influence on the so-called digit ratio, i.e., the length of the index finger divided by the length of the ring finger. This would then explain why, on average, male homosexuals tend to have a slightly higher digit ratio than male heterosexuals.

The prenatal hormone theory has also produced hypotheses that correlate sexual orientation with visible facial features. After all, if prenatal exposure to hormones influences both sexual orientation and facial morphology, it seems reasonable to expect some kind of correlation between the two. Building on this idea, Wang and Kosinski used deep neural networks, i.e., artificial neural networks able to learn to detect patterns often missed by human observers, to predict one’s sexual orientation solely on the basis of one’s facial features. Their database consisted of tens of thousands of pictures found on dating websites and on Facebook, and they established the sexual orientation of the individuals in these pictures by means of their profile or their having liked pages such as “I Love Being Gay” and “Gay and Fabulous.” According to Wang and Kosinski, the neural network they developed was able to accurately predict one’s sexual orientation based on facial features alone. Moreover, the accuracy of their “artificial gaydar” was much higher than the accuracy of human judges, who did barely better than chance.

Wang and Kosinski’s research ran into fierce opposition. First, there were concerns about methodology and data interpretation. Critics objected to the authors’ reliance on pictures from dating sites and from openly gay Facebook users because these pictures probably contain hints about the person’s sexual orientation (Miller Reference Miller2018). A replication study confirmed that hunch: the success of Wang and Kosinski’s gaydar may have been not so much based on its interpretation of facial features as on its interpretation of other aspects of the pictures, such as colors and lighting (Leuner Reference Leuner2019). Other critics pointed out that the authors overstated both the robustness of the evidence for the prenatal hormone theory and the accuracy of their neural network (Gelman, Marrson, and Simpson Reference Gelman, Marrson and Simpson2018).

Secondly, and more importantly for our purposes, there were moral concerns. Some of these concerns related to privacy and consent. Are Facebook pictures public data, for example, to be used at will by scientists (Miller Reference Miller2018)? The main moral objections, however, concern the potential abuse of Wang and Kosinski’s neural network. Critics feared that homonegative individuals, organizations, and governments could use it to identify and trace homosexuals, with the express purpose of harming them. This criticism echoes earlier criticisms of biological research on homosexuality because it could lead to screening procedures that would allow prospective parents to select for heterosexual children (Schüklenk et al. Reference Schüklenk, Stein, Kerin and Byne1997).

There is an asymmetry at work in these research projects similar to the one Kitcher points to in his argument: Depending on the outcome of the research it may be neutral or harmful to homosexuals, but it is unlikely to ever benefit them. Unlike Kitcher’s argument discussed in the previous section, this seems to provide a good reason not to allow such research. However, this argument only targets applied research on the etiology of homosexuality, and leaves more basic research on the etiology of homosexuality untouched. There are three reasons for this.

First, it may be that imposed restrictions on basic research are counterproductive. When Kitcher argued for moral constraints on the freedom of scientific inquiry, he explicitly warned against censoring scientific research by governments, funding agencies, or academic journals. Such censorship, he says, will often backfire: “In a world where (for example) research into race differences in IQ is banned, the residues of the belief in the inferiority of the members of certain races are reinforced by the idea that official ideology has stepped in to conceal an uncomfortable truth” (Kitcher Reference Kitcher2001, 105).

If research on the etiology of homosexuality is forbidden, chances are that anti-gay voices will spin this as an attempt to hide “inconvenient truths” about homosexuality.

Second, the ship sailed on this a long time ago. There is already such an enormous amount of published research on the etiology of homosexuality that a ban would be pointless. And if we were to restrict all new research on this topic, we would make it much more difficult to criticize the extant body of pretty-robust-but-still-fallible knowledge on the etiology of sexual orientation. The current state of this field would then be the final state. Of course, this may make it more difficult to design well-functioning applications than would be the case if we had a better theoretical grip on the origins of homosexuality. But applications that are functionally suboptimal need not be less harmful than applications that function better because they are based on better basic research.

Third, the link between basic research and the misuse of technology resulting from such research is obviously more indirect than between the development of this technology and its misuse. This distance is relevant because it entails that the danger of misuse is not so imminent for basic research. Furthermore, the greater the distance is between research and its potential misuse, the more potentially useful research lines we’d also have to limit. There are usually many more research lines that have an indirect link with a misused technology than there are research lines directly linked to such misuse. For instance, if it is decided to forbid research on the etiology of homosexuality because the results can be used for developing “gaydar” technology, one could argue that any research on the intra-uterine effects of androgens (or of all hormones) should be banned as well.

Thus, the applications-argument for limiting research into the etiology of homosexuality only works for applied research. Still, it establishes a substantial and particularly high profile-group of research projects that should be subject to limitations. Three years after its publication, Wang and Kosinksi’s paper has already been cited more than five hundred times. Vox, Quartz, the New York Times, and the Financial Times quickly mediatized the findings, and the Guardian published a long interview with Kosinski on the study. Because the topic is salient and controversial—and thus likely to be highly cited and published in high-profile journals—scientists are incentivized to engage in such research. Therefore it is important to know which research should be limited, and what these limitations should be. We turn to these questions in the next sections.

4. Limitations to research or limitations to dissemination?

Nonconsensual medical experimentation on a (vulnerable) population, as happened in the infamous Tuskegee syphilis study, is widely agreed to be unethical, regardless of its actual or expected results. Moreover, the scientific community is well aware that applied research with potentially harmful uses should sometimes be subject to limitations. Indeed, most funding applications require that researchers flag the dual-use potential of their research. However, because of the asymmetry discussed above, applied research on the etiology of homosexuality is not a typical dual-use case: While the research may be harmful, it is unlikely to have any direct benefits. The main reason to permit such research, then, is to enable us to mitigate the potential harms that could ensue if someone with bad intentions does the same research. Indeed, Wang and Kosinski claim that this was the main motivation for their study: “[s]ome people may wonder if such findings should be made public lest they inspire the very application that we are warning against. We share this concern. However, as the governments and companies seem to be already deploying face-based classifiers aimed at detecting intimate traits […], there is an urgent need for making policymakers, the general public, and gay communities aware of the risks that they might be facing already. Delaying or abandoning the publication of these findings could deprive individuals of the chance to take preventive measures and policymakers the ability to introduce legislation to protect people” (Wang and Kosinski Reference Wang and Kosinski2018, 255).

Research like Wang and Kosinski’s should be permitted if the potential benefits of taking preventive measures and developing legislations outweigh the potential harms of providing ill-intentioned people and governments with these technologies. However, this harm/benefits calculus is particularly hard to make in these asymmetrical dual-use cases, as it is often impossible to predict how soon ill-intentioned people will come up with a technology that doesn’t yet exist, how much we speed up that process by doing the research for them, and how much we gain from taking preventive measures. Unlike in traditional dual-use cases, where there are often direct benefits to the research that are reasonably easy to estimate, the benefits of asymmetrical cases like Wang and Kosinski’s are nearly impossible to judge.

Fortunately, such a harms/benefits calculus is not necessary here. This is because the potential harm (ill-intentioned people using the technology) can be minimized while still enjoying the potential benefits (taking preventive measures). This is because instead of limiting research into the etiology of homosexuality, we can limit the dissemination of this research: if the research-outcomes are only communicated to the groups of people that are in a position to take preventive measures, then no harm ensues and future harm can be mitigated. We therefore propose that limitations to applied research on homosexuality should take the shape of responsible disclosure procedures.

Limitations on the dissemination of research findings are already common in various fields. In cyber security, vulnerabilities in security systems are studied and documented, but the researchers initially only share the results with the developer, the vendor, or the consumer, so that these stakeholders are given the opportunity to find countermeasures first. Once the safety issues have been sufficiently resolved or mitigated, vulnerabilities can be shared with a broader community of stakeholders to avoid their future reoccurrence (Householder et al. Reference Householder, Wassermann, Manion and King2017). Similarly, restrictions on the dissemination of research are already broadly accepted in biotechnology and the biomedical sciences, and one of the main recommendations of the influential Fink report was to review research at the publication stage (NRC 2004). For example, there is a lot of research on the genomics of deadly viruses, but the viruses’ genomes are not made publicly available because there is a real danger that terrorists would use this information to develop a very effective bioweapon (Kourany Reference Kourany2016).

Of course, individual social scientists that are aware of the risks of their research may already be careful about how they disseminate their work. For example, Wang and Kosinski did not publish the full source code and the learned data structures, and collaborated with stakeholders such as the American Civil Liberties Union in an early stage of their research. Still, a formal and institutionalized framework that oversees such responsible dissemination is lacking in the social sciences (and much of academia). Spelling out such a framework requires answering a range of difficult practical questions and is beyond the scope of this paper. Instead, the remainder of this paper considers two crucial questions that have to be answered for responsible dissemination: Who should regulate dissemination, and what limits to dissemination should we consider? We offer tentative answers to these questions, based on best practices in fields that already have mature procedures to deal with the responsible disclosure of research that poses potential risks (see also Brundage et al. Reference Brundage, Shahar Avin, Helen Toner and Ben Garfinkel2018).

5. Objections

These brief remarks obviously fall short of fully developed procedures for responsible disclosure that can deal with complex cases like Wang and Kosinski’s research. However, we hope they are sufficient to illustrate the complexity of the decisions at hand, and the need to have a formal framework in place to deal with them. While Wang and Kosinski may not have made the best decision, at least they were aware of the problem and took some mitigating measures. It would be unwise to assume that all researchers in the future will be equally (but probably still insufficiently) careful.

We are aware that the limits to free publication of research that we propose here are not uncontroversial, and subject to various problems and objections. We finish this paper by briefly considering some of these.

First, one may argue that our proposal is at odds with the rapidly growing open science movement. And just like openly publishing research like an “AI gaydar” may entail risks, there are risks to limiting open access to research findings. For one, it could decrease inclusivity and lead to a concentration of power in certain research groups (Whittlestone and Ovadya Reference Whittlestone and Ovadya2020). This objection shows that the decisions at stake here are value-laden decisions, trading off openness, scientific progress, and potential harms to minorities. Rather than showing that there should be no limits to free publication, we think this objection strengthens our proposal to take this decision out of individual researchers’ hands, and develop procedures that are supported and evaluated by a broader community.

Second, one may worry that review procedures are slow and, if some research gets censored or communicated to a limited audience, may sometimes be at odds with the professional ambition of researchers (Ovadya and Whittlestone Reference Ovadya and Whittlestone2019). Thus, there would be strong incentives for researchers to defect, and use different channels of publication to benefit from the attention that spectacular and controversial research is likely to get. While the risk of defectors is real, it may be possible to minimize it if procedures of responsible disclosure are implemented and supported across the relevant field of research. Even if researchers get their research out in ways that avoid the proposed pre-publication review, they could be punished by not citing such research or not accepting such research at conferences.

Finally, one could object that for almost all research in the social sciences, the risks are not immediately visible. Instead, research sometimes only entails risks in combination with other, potentially future research that in itself also does not entail risks either. Alternatively, it may be that research that is harmless now could have safety risks at some point in the future when it becomes more refined. This objection ties back into our remarks about fundamental research, which we think should not be subject to these measures of responsible disclosure. However, even if we assume that there should be no restrictions to fundamental research, difficult questions remain. For example, there is not always a clear difference between fundamental and applied research, and it remains important to be mindful of research with future potential of misuse: it may be much harder to mitigate risks if one waits too long. For example, Kourany (Reference Kourany2016) argues that research on cognitive differences between genders or ethnic groups has done much harm in the past, and that the harms of the continuation of this research tradition are foreseeable and should trump concerns about the freedom of research. Others are less sure about how foreseeable scientific results and the associated harms are (Carrier Reference Carrier2021). However, the difficulty of these decisions is no reason not to try and tackle them. Instead, it highlights the need for strong and clear procedures to deal with them.