Competition and cooperation in artificial intelligence standard setting: Explaining emergent patterns

Efforts to set standards for artificial intelligence (AI) reveal striking patterns: technical experts hailing from geopolitical rivals, such as the United States and China, readily collaborate on AI within transnational standard-setting organizations, whereas governments are much less willing global ethical international organizations. Whether competition or cooperation prevails can be explained by three variables: actors that make up membership of organization, issues which organization's efforts focus, “games” play when trying a particular type organization. A preliminary empirical analysis provides support contention actors, issues, games affect prospects standards. It matters because shared vital achieving truly frameworks governance AI. Such frameworks, in turn, lower transaction costs probability world will witness emergence systems threaten human rights fundamental freedoms. Los esfuerzos para establecer estándares la inteligencia (IA) revelan patrones sorprendentes: los expertos técnicos provenientes de rivales geopolíticos, como Estados Unidos y colaboran fácilmente en IA dentro las organizaciones transnacionales establecimiento estándares, mientras que gobiernos están mucho menos dispuestos colaborar. sobre éticos globales internacionales. Que prevalezca competencia o cooperación puede explicarse por tres actores componen membresía organización normas; cuestiones se centran normas organización; “juegos” juegan cuando intentan un tipo organización. Un análisis empírico preliminar respalda afirmación actores, problemas juegos afectan perspectivas IA. Importa porque compartidos son vitales lograr marcos verdaderamente gobernanza Dichos globales, su vez, reducen costos transacción probabilidad el mundo sea testigo del surgimiento sistemas amenazan derechos humanos libertades fundamentales. 为人工智能(AI)制定标准一事所作的努力揭示了惊人的模式:来自美国和中国等地缘政治对手的技术专家很乐意在跨国标准制定组织内就技术AI标准进行合作，而政府则不太愿意在国际组织内就全球AI伦理标准进行协作。竞争占优势还是合作占优势，这可以用三个变量来解释:标准制定组织成员的参与者；组织的标准制定工作所关注的问题；以及当试图在特定类型的组织中设定标准时，行动者所进行的“博弈”。初步的实证分析为一个论点提供支持，即行动者、问题和博弈会影响全球AI标准合作的前景。这很重要，因为共享标准对于实现真正的全球AI治理框架是关键的。这样的全球框架反过来会降低交易成本，并减少世界见证威胁人权和基本自由的AI系统出现的可能性。 Technology is multifaceted, involving many different instruments. In recent decades, standard setting has become one key mode technology governance, evolving an “internal matter individual firms subject among independent players” (Mattli, 2001, p. 328; see also Besen & Farrell, 1994, 117). As Veale Borgesius (2021, 9) point out, standardization often where “real rule-making” occurs, making important endeavor governments, firms, non-profits, multitude other who variously compete cooperate with each advance their diverse business political agendas. At moment, numerous private public vying shape (AI)1—and its governance—by developing area (Cheng Zeng, 2022, 2; Schmitt, pp. 303–314). principle, cover all system lifecycle phases, including data collection, training, continual learning, design, testing, evaluation, use. They take form either specifications guidelines. For example, may lay out how collect so ensure meets certain performance benchmarks (a specification) it stipulate should gender equality (an guideline). specification spell engineer comply An example would describes what engineers need do sure produces unbiased results. Thus, specify methods engineering these meet criteria (which not nature) while moral requirements conform. This means order standard, developers must build way described standard's document—but they goals (such transparency, fairness, explainability, etc.) specified standard. sum, provide instructions systems. Ethical standards, hand, but contain describing Standards explicit norms (Büthe Mattli, 2011) there binding (de jure) non-binding widely used facto) Accordingly, term “AI standard” defined follows this paper: norm takes guideline governs any aspect phases (Figure 1). AI, encompasses both rule-based learning-based et al., 2022), pervasive applications across industries, social relationships, science. could potentially add 13 trillion US dollar economic output 2030, increasing GDP approximately 1.2% per year (Bughin 2018), if potential fully exploited risks adequately mitigated. The wide range “dual use” nature, civilian military ends, increasingly critical infrastructure (Ding Dafoe, 2021), meaning countries develop, adopt, regulate Given high stakes, national regional bodies European Union, EU), minilateral organizations Organization Economic Cooperation Development, OECD), traditional intergovernmental (IOs, Nations, UN, World Trade Organization, WTO), (SSOs, International Standardization, ISO, Electrotechnical Commission, IEC), host multi-stakeholder Partnership PAI) AI—and own solutions intended What explains emergent patterns intelligence? My answer question focuses specifically SSOs IOs, types have long been fora 2011 , 39). Before proceeding, note definitions order, crisp facilitate analytical clarity regarding factors examined article: When I speak “transnational SSOs,” refer whose express purpose voting consists (who vote directly through committees). Examples IEC, Institute Electrical Electronics Engineers (IEEE). “IOs,” states covers at least 80% world's sovereign, recognized (i.e., 156 countries). Public than countries, OECD Security Europe (OSCE), “minilateral bodies.” While employing detailed might seem unnecessary, allows clearer IOs SSOs. argue Section “Three variables explaining setting” two (with having foci facing opportunities challenges related non-governmental actors), focus (in particular, whether concentrate drafting guidelines, turn impacts play), goods coordination games). establish usefulness explanatory model “Emergent elucidate why guidelines draw limited, further work needs done substantiate my analysis. Nonetheless, theoretical framework, show matter. divergent impact rivalries domestic constraints governmental compared meant address, payoff structures contribute addition, interactions between understood well. popular narrative years rivalry China negatively impacting (see, e.g., Cheng 2022; Gualtiero Blancato, 2019), dampening Yet, achieved time again, ISO/IEC JTC 1 SC 42 (the subcommittee AI) published 14 since inception 2017 (ISO, n.d.). Generalizing results areas setting, put forth part analysis, attempting attempted focuses, carefully fact, case “easy” those hold undermines governance. After all, well known increase influence against backdrop Sino-US trade frictions Arcesati, 2021, 1; Cordell, 2020; Kamensky, 2020), American decision-makers attempted—both publicly behind scenes—to prevent Chinese aim 2022). Despite level high. matters, precondition article concludes “Conclusion” brief appeal view serious negative consequences freedoms split into technological blocs. There number works examine (aspiring) setters type, (Gilmore 2006), (den Uijl, 2015; Schilling, 2002; Suarez, 2004; van Kaa 2011; den Ende 2012) (Jain, 2012; Leiponen, 2008; Mattli Büthe, 2003). Focusing setters, Büthe (2011) conducted in-depth study Accounting Board (IASB) determine wins loses organizations—and why. handful studies describe interests, fierce involved (e.g., Cihon, 2019; Cihon Jobin Kerry 2021). However, descriptive accounts develop explain it. Previous moreover only, setters—a task regard timely given stakes associated setting. al. (2022, 1738–1739) pointed “research promise limits governance—and conditions under achieved—is greatly needed.” contributes fulfilling need. Understanding account levels insights thus, ultimately, development globally framework Without common rules, blocs employ jeopardizing respects freedom. complements another paper special issue: Chou Gomes press) explore AI-powered algorithms deployed digital platforms freelancers livelihoods depend platforms. concrete effects abundantly clear uphold crucial values privacy (see Taeihagh submit mostly (i) organization; (ii) organizations; (iii) focus. relationships well: kinds address and, vice versa, attract process. being addressed, affects kind game playing. external environment, differing already large amount relevant facto proposed important, too. one, more detail sections below, players' behaviors games—in words, “the context individuals face dilemmas” (Ostrom, 2010, 160) matters. Additionally, trust therefore influences sorts topics brought fore processes extent feel get processes. Game theory valuable interplay competitive cooperative dynamics member state incentive try freeride created others thereby exploit gain. dominant strategy game, due rather non-excludability non-rivalry characteristics goods. Competition does, however, chances overcoming dilemma. Geopolitical creates distrust makes hesitant engage mutual assistance. behave poorly, prompting them forego even doing carries risk bad reputation over consequently becoming stuck uncooperative relationships. harder overcome incentives non-cooperation inherent structure games. SSOs, members advocating proposals. good reasons cooperate. very helpful foundation thinking about tells us outcomes we expect based payoffs. obviously simplified reality (otherwise never able observe real world), allow deconstruct complex, messy constitute first explanation providing under-provision markets (Abbott Snidal, 358ff; Kindleberger, 1983; 2003; Sykes, 1995), strict sense non-excludable non-rivalrous (Conybeare, 1984, 6).2 Achieving game.3 creating promoting goods, is, benefits extend globally, just nationally regionally—for security (political stability), stability, environment (Stiglitz, 1995). fact definition contributors cannot non-contributors enjoying benefits, enabling freeride. Non-rivalry refers availability affected consumption, existence freeriding does result getting quality give immediate cease good). games, substantial will, aggregate, under-supplied: beneficiaries becomes larger, beneficiary's optimal contribution most instances) tends towards zero declining marginal benefit, problem prisoner's dilemma 6). players akin no cooperation. like lead suboptimal leave worse off collectively had cooperated 7; Ostrom, 155). “assist” entering commitment individually rational choice “mutual assistance” (Snidal, 1985, 927) player larger gains acting independently would. choice, reduces likelihood actually agree mutually assist other. Abbott Snidal (2001, 358) out: “PD [prisoner's dilemma] collective action models suggest agreements difficult achieve.” To illustrate difficulty let me game: Assume IO reach agreement limit use power-intensive graphics processing units machine learning training reduce carbon emissions improve environment. cooperate, sign keep using cooperating exclude improved stem reduction emissions. non-cooperating state's enjoyment moreover, benefit reduction, giving itself stop good. capabilities benefitting cleaner air “greener” world. logic same state, want indeed achieve. are, scenarios. These include temporal, repeated dimension attendant reputational concerns effective enforcement mechanisms 1985). repeated-game interacted before know likely again (given stable, established unlikely dissolve suddenly). unlike classic one-shot possible come rewards stable enduring nature behavior now prompt retaliation later. “shadow future” (Powell, 1991, 1306) engenders reputation. non-cooperator stuck, consequence, unable achieve cooperation, early on. Ostrom (2010, 162) points players, investment trustworthy reputations, reciprocity successful action. temporal increased comes strong provisions. provisions later complying defection, otherwise signing place. name implies, exist standards—and trade.4 So, create strengthen goods: Global excludable country barred trading world, punishment starting war) rivalrous (if fulfills B's foreign services, then A's exports B expense C, can—at most—fulfill whatever remaining services has). Since fulfill criteria, 1984). 1995)—but so. worsen contagion financial crisis, destabilizing were network (Kali Reyes, 2010). Some themselves 1983), Kaul Mendoza, 2003) believe case, standards—in product standards—can certainly made excludable, property standard-essential patents. engaged games.5 Rather, modeled SSO discussing tricky questions main (as discussed section below). Technical effects, users has, users. convergence (Spruyt, 2001 374–375)—but select. considering various specifications, choose best without knowledge choose, obvious settle (since interests prefer specifications). reap approve representation situation policy knowing intending do, coordinate 1985 932). everyone better off. example: Chinese, proposing testing robustness neural networks, unclear outset settle. place, specification. Once was inefficiencies multiple disappear. dynamic right (Nativi De Nigris, relative dearth movers, few invested enough costly standards,6 return economies scale spring standardization. distributional conflicts low settling promote standardization, challenge lying picking situation, “cooperative preferences” prevail “competitive dynamics” 375) endorsed. useful tool analyzing equilibria line varying characteristics, pursue Their calculations desirability agreeing different. Setting rules confer upon home votes satisfied citizens. government representatives pressure industries 376). weapons producers lobby forbids production lethal autonomous (LAWS). Governments playing “two-level games” (Putnam, 1988) quite complex groups, citizens, industry,