A Season for Repurposing AI Alignment?

“To every thing there is a season, and a time to every purpose under heaven.”

You may have encountered this wisdom-imparting statement through the 1965 hit song by the Byrds (Turn, Turn, Turn) or the biblical passage (Ecclesiastes 3:1-8). It seems that those closest to the frontiers of AI development have gotten the message. Over 1,300 people working on the frontlines of AI research and development in the US recently published “Pacing the Future”—an appeal directed to the US government and endorsed by OpenAI and Anthropic to take the lead in establishing a global institution capable of developing and wielding “the technical and governance tools needed to deliberately pace the frontier of automated AI development.”

This appeal issued from within the AI-development community is welcome. Regulating the pace of change along the frontier of human-AI interaction may well be necessary if we are to have any hope of ensuring that the benefits of AI are not outweighed by its harms. But from a Buddhist perspective, as important as the pace of AI development may be, it will ultimately be the purpose(s)—the karmic complexion—of AI development that matter most. Clarity about this is, I think, a prerequisite for effective commitment to realizing ever more virtuosic human-AI alignment.

The need for a regulatory slowing of AI development is not hard to fathom. Concerns about the catastrophic potential of AI have been circulating for years, and not just from the doomsaying fringe. In a recent MIT study, 272 AI experts were asked to estimate the probability of AI causing catastrophic harm, defined as resulting in more than 1 million deaths, more than $100 billion in financial losses, or comparable intangible damages. Five risks were identified as having more than a 10% chance of causing catastrophic harm: dangerous AI capabilities, competitive corporate/geopolitical pressures, autonomous weapons and cyberattacks, power concentrations, and false information.

A 10% chance might sound relatively small. But if you were told that a better and brighter future exists beyond a door in front of you, but that you have a 10% chance of being killed by opening it, would you open it? For AI insiders like the signatories of “Pacing the Future” and the heads of major AI labs to call for global regulation is understandable and sincere. It is a call for sharing the moral and legal loads of ensuring that AI has beneficially aligned impacts.

That downstream focus is an important dimension of human-AI alignment. But can alignment efforts flow the other direction as well? Are there viable upstream alternatives that focus regulatory concern on the purposes of AI rather than its outputs?

Yes. But those upstream alternatives might not be coming from what many of us would presume to be the usual sources—the EU or the UN, for example. Instead, on July 15, just prior to the World Artificial Intelligence Conference hosted in Shanghai, China’s Interim Measures for the Administration of AI Anthropomorphic Interactive Services came into force. This marks a strong tactical turn toward operationalizing the official stance of the People’s Republic of China (PRC) that AI should assist people and not harm, deceive, or exploit them. As expected by those familiar with the rhetorical norms of official Chinese policy statements, the document also clearly stipulates that AI should not contravene the good of the PRC or undermine socialist values (Article 8.1). But despite that requisite assertion of the primacy of the state’s paternalist authority, the Interim Measures merit a close look.

First, the document evinces admirable concern for the impacts of human-like AI systems on the most emotionally vulnerable members of society: children under the age of 18 and the elderly. The law explicitly acknowledges the value of AI customer service agents, AI inquiry interfaces that offer knowledge and cognitive tools, AI productivity assistants, and AI systems designed to further education and research. The “anthropomorphic AI” systems that the law concerns are those that are designed to foster and sustain emotional connection and attachment.

A few highlights will serve to illustrate the upstream distinctiveness of the Interim Measures’ regulatory approach.

• Article 8(5) prohibits providers from “excessively catering to users, inducing emotional dependence or addiction, and harming users’ real interpersonal relationships.”

• Article 10 stipulates that anthropomorphic interactive service providers must fulfill security obligations throughout the lifecycle of the service, including “protecting user privacy and personal information, providing early warnings of excessive dependence risks, guiding emotional boundaries, and protecting mental health.”

• Article 14 explicitly prohibits offering minors services for “virtual family members or virtual partners.” And, as preventative measures, AI service providers are required to incorporate into their designs conspicuous and dynamic reminders that appear as soon as users show signs of dependency.

• It is also mandatory for AI service providers to issue warnings/reminders after two continuous hours of interaction (Article 18) and automatic alerts to a “user’s guardian or emergency contact person” if the user is facing or suffered significant property loss or has expressed a readiness to commit self-harm or engage in life- or health-threatening activities (Article 13).

These are bold and precisely targeted regulatory provisions. And in contrast with efforts to limit access by minors to certain kinds of AI services and social media in Australia, the EU and some states within the US, China’s Interim Measures not only aims to regulate the output content of AI systems; it aims to regulate the business model of anthropomorphic AI services. Instead of invoking broad AI regulatory principles, it distinctively lays out operationally-defined and risk-based rules regarding the scope of what kinds of AI systems can be built, going decisively beyond the default approach of “first innovate, then regulate.”

This is admirable. Especially significant is the commitment to protecting the most vulnerable in society—here including minors and the elderly—from the intimate harms of an AI-manipulated attachment economy. But just as preventative medicine is silent on the positive meanings of health and focuses on avoiding negative health conditions, a restrictive protection approach to AI risk does not specify a substantive vision of the future of human-AI relations. And, while the statist framing of the Interim Measures might be excused as a rhetorical overshoot that is no more (or less) troubling than championing commitment to “make America great again,” it is clear from the continued presence China’s Great Firewall and social media censorship that its digital protectiveness does not imply unrestricted access rights, including the right to determine what one attends to, both online and off. That is not an exclusively Chinese problem, of course. The seemingly limitless freedom of online content choice and its ironic threat to freedom-of-attention is one of the defining predicaments of the era of AI-orchestrated connectivity.

In this global season of exponentially increasing AI capabilities, the widespread anticipation of AI-enabled efficiency-as-a-service, and deepening concerns about human cognitive and emotional deskilling, it is high time to reflect with deepening clarity on the appropriate (and inappropriate) purpose(s) of AI.

Calling for a pause or a slow down along the frontiers of human-AI interaction is sensible. China has perhaps unexpectedly articulated an approach to addressing the alignment problem that considers both AI-generated content and the business models that are infused—by choice or default—into AI-design. That is also eminently sensible. How effective and scalable either will prove to be is an open question.

One of the guiding insights of Buddhism is that we can only succeed in sustainably alleviating and eliminating the causes and conditions of conflict and suffering through high resolution understanding of how things have “come to be as they are” (yathābhūtaṃ). Histories matter. A Buddhist “theory of change” involves working with historical dynamics—a process that I’ve described as “karmic aikido” or reorienting ongoing change dynamics rather than resisting them. Resistance isn’t always futile. It is often required to open spaces for reorientation. But karmically, what matters most is the orientational shift.

The Buddha’s recommendation, as very colorfully and carefully depicted in the Sakkapañha Sutta (DN 21) is to undercut the mental proliferation of relational blockages (prapañca; prapañca) by assessing the relational outcomes and opportunities of our actions and persisting with actions that are kuśala and refraining from the that are akuśala. Given that kuśala can be translated as “skillful” and/or “wholesome,” that sounds simple enough. But given that all things are impermanent and empty of any fixed essences or natures, being skillful and wholesome must be situationally defined. They are directions for practical and moral development, not predetermined destinations. Conflict and suffering are ended through the pursuit of practical and moral virtuosity.

What would it mean to cultivate virtuosic human-AI alignment? What modes of human-AI interaction and shared agency would qualify as virtuosically purposed? Aligning AI with current global complexions of human values, intentions, and actions would most certainly not be kuśala—just look at Gaza or Ukraine or any home riven by domestic violence or any child’s body-mind stunted by debilitating hunger. For the agential collaboration of humans and AI systems to be kuśala will require realigning both AI systems and humans.

Frontier AI labs now anticipate AI systems that are recursively self-improving within the next year or two. Trained on a digital ālayavijñāna or “storehouse” of human karmic data, a recursively self-improving AI system would be both capable of and inclined toward scaling up human conflicts and propensities and our persistent valorizations of control, convenience, and competition. That does not bode well for humans or for AI. It is the season to try slowing down AI development and adoption. But it is also the time to reimagine the purposes of the intelligence revolution.

Buddhist teachers have often urged their students to “meditate as if your hair is on fire.” Today, we should with similar urgency collaborate in determining who we need to be present as to meet the challenge of AI.

—

Peter Hershock is co-founder of the Buddhism & AI Initiative and an intercultural philosopher who reflects on contemporary issues of global concern.