Value alignment, human enhancement, and moral revolutions

Abstract

ABSTRACTHuman beings are internally inconsistent in various ways. One way to develop this thought involves using the language of value alignment: the values we hold are not always aligned with our behavior and are not always aligned with each other. Because of this self-misalignment, there is room for potential projects of human enhancement that involve achieving a greater degree of value alignment than we presently have. Relatedly, discussions of AI ethics sometimes focus on what is known as the value alignment problem, the challenge of how to build AI that acts in accordance with our human values. We argue that there is an especially close connection between solving the value alignment problem in AI ethics and using AI to pursue certain forms of human enhancement. But in addition, we also argue that there are important limits to what kinds of human enhancement can be pursued in this way, because some forms of human enhancement—namely moral revolutions—involve a kind of value misalignment rather than alignment.KEYWORDS: Artificial intelligencehuman enhancementmoral revolutions AcknowledgementsBoth authors would like to thank the National Endowment for the Humanities for support for their work, the University of Puget Sound and the John Lantz Senior Fellowship for Research or Advanced Study, and the participants at the Philosophy, AI, and Society Workshop at Stanford University. Ariela Tubert would like to thank the audience at the Ethics and Broader Implications of Technology Conference at the University of Nebraska at Lincoln.Disclosure statementNo potential conflict of interest was reported by the author(s).Notes1 See for instance Russell (Citation2019), Christian (Citation2020), Gabriel (Citation2020), Wallach and Vallor (Citation2020).2 Appiah (Citation2010). See also Baker (Citation2019).3 Russell and Norvig (Citation2010).4 Gershman (Citation2021, 156) makes this point while arguing that the ‘folklore’ about how machine learning has its origins in neuroscience overstates the level of influence neuroscience has actually had.5 See for instance Kahneman, Slovic, and Tversky (Citation1982), Kahneman and Tversky (Citation2000), Kahneman (Citation2011).6 Lieder et al. (Citation2019).7 Lieder et al. (Citation2019, 1096).8 Lieder and Griffiths (Citation2019). The notion of ‘rational analysis’ is drawn from Anderson (Citation1990).9 This is a point of focus in Griffiths (Citation2020).10 Lieder et al. (Citation2019, 1096).11 Lieder et al. (Citation2019, 1096). On gamification and AI more generally, see Deterding et al. (Citation2011).12 Chasse (Citation2021).13 Lieder et al. (Citation2019).14 Sinnott-Armstrong (Citation2008).15 Tversky and Kahneman (Citation1981).16 As Kühberger (Citation2017, 79) notes, the effect is robust and has been replicated across hundreds of papers.17 Kahneman and Tversky (Citation1979).18 Sometimes this point is used as part of an argument that we should be skeptical of moral facts at all, but this move requires a further inference. For influential discussions of some of the issues involved, see Street (Citation2006), Joyce (Citation2007).19 Singer (Citation2005), Greene (Citation2007).20 See for instance Haidt (Citation2012).21 Kass (Citation1997).22 Nussbaum (Citation2004). Kelly (Citation2011) offers an extended discussion of the moral significance of disgust.23 The notion of an expanding circle of moral concern comes from Singer (Citation2011).24 On Tay, see Victor (Citation2016). On the Turkish translation case, see Olson (Citation2018).25 On search engines, see Noble (Citation2018). On facial recognition systems, Buolamwini and Gebru (Citation2018). On hiring decisions, Dastin (Citation2018). On loan and credit card applications, Angwin et al. (Citation2016). On predictive policing, O’Neil (Citation2016). On sentencing and parole decisions, Eubanks (Citation2018).26 See for instance Kleinberg et al. (Citation2018), Kleinberg et al. (Citation2020).27 See for example Dovidio and Gaertner (Citation2000), Amodio and Devine (Citation2006), Gendler (Citation2011), Levy (Citation2017). For a critical assessment of work on implicit bias, though, see Machery (Citation2022).28 Wallach and Allen (Citation2009). We note though that they frame their discussion in terms of building moral machines rather than in terms of value alignment. For Wallach’s thoughts about value alignment, see Wallach and Vallor (Citation2020).29 Mill (Citation1861/1998). Discussions of a utilitarian-oriented AI include Gips (Citation1994), Grau (Citation2011), and Russell (Citation2019).30 Kant (Citation1785/2012). Thomas Powers’ (Citation2006) ‘Prospects for a Kantian Machine’ connects the view to AI.31 Asimov (Citation1950).32 Each of these examples is mentioned by Wallach and Allen (Citation2009, 79).33 Shortliffe and Buchanan (Citation1975).34 Savulescu and Maslen (Citation2015), Giubilini and Savulescu (Citation2018). For critical discussion of the proposal that is still sympathetic to the idea of pursuing AI-based human moral enhancement, see Lara and Decker (Citation2020).35 Deterding (Citation2014) discusses moral gamification, defending a ‘eudaimonic design’ approach.36 Millar (Citation2015) and Contissa, Lagioia, and Sartor (Citation2017) argue in favor of user control over the ethical settings on autonomous cars, while Lin (Citation2014) and Gogoll and Müller (Citation2017) argue against the idea.37 Santurkar et al. (Citation2023). See also Rozado (Citation2023).38 Thompson, Hsu, and Myers (Citation2023).39 See Narayanan and Kapoor (Citation2023) for a critical discussion of Santurkar et al. (Citation2023).40 OpenAI (Citation2023).41 Steinberg (Citation2023).42 Walker (Citation2023).43 Marcus (Citation2023).44 Appiah (Citation2010). Klenk et al. (Citation2022) provides a survey of recent work on moral revolutions.45 Appiah (Citation2010:, 8), Kuhn (Citation1962). Klenk et al. (Citation2022) emphasize how this connection to Kuhn is common also in other authors discussing moral revolutions.46 Wallach and Allen (Citation2009, 79).47 LeCun, Bengio, and Hinton (Citation2015), Bengio, LeCun, and Hinton (Citation2021).48 Ensmenger (Citation2012).49 Holodny (Citation2017).50 Metz (Citation2016).51 Knight (Citation2017).52 Strogatz (Citation2018).53 Rini (Citation2017) also uses AlphaGo’s Move 37 as an analogy for a radically new AI moral view.54 Appiah (Citation2010:, 66), Klenk et al. (Citation2022).55 See discussions of what is needed for significant society-wide moral progress: Moody-Adams (Citation2017), Rorty (Citation2006), Nussbaum (Citation2007).56 Appiah (Citation2010).57 On AI and the risk of value lock-in, see for instance Ord (Citation2020: Chapter 5), MacAskill (Citation2022: Chapter 4).58 Kenward and Sinclair (Citation2021).