BATCH Report
By Matthew Stone
Critical AI helped organize a workshop at ACL 2026, BATCH: Transdisciplinary conversations on human-AI futures, which was held on July 4 in San Diego, CA. Those who registered for the conference can view the workshop recordings on ACL’s underline page; others can take a look at the program. Partial support for the workshop was provided by our NSF planning grant on “Writing education through design-oriented AI,” in line with the project’s goal: to convene interdisciplinary meetings to advance responsible, human-centered frameworks for language technology. [1]
Alignment refers in the first instance to a technical practice that supports the development of AI systems that can follow instructions expressed in natural language. More generally, however, talk of “alignment” evokes a framing of the trajectory of AI research as leading to systems that make autonomous, consequential decisions, with potentially superhuman speed and scope. In this framing, alignment encompasses any effort to ensure that such systems remain responsive to the interests of their designers, their users, and the broader public.
Our first talk of the workshop was a critical survey of alignment in NLP presented by Professor Emily M. Bender representing collaborative work with several colleagues. Their work highlighted the diversity of approaches to the technical problems of alignment, by casting alignment as an “umbrella” that protects research teams while leaving other stakeholders (as it were) “exposed to the elements”. This perspective reveals a concerning gap between researchers’ technical approaches and their far-reaching, programmatic, and often speculative motivations, expressed through rhetorics of ethics, safety, and even superintelligence.
Our next session was a “fishbowl conversation” designed to demonstrate the potential of transdisciplinary approaches. The format built on ACL’s tradition of welcoming conversations across disciplines—conversations I first appreciated in my own early work on generating, understanding, and following instructions. NL action descriptions do not tell somebody what to do in anything like the way a computer program does. Rather (linguistics teaches us), meanings are abstract and underspecified, so (as psychology confirms), interlocutors count on one another’s capacities for planning, mutual understanding, and “theory of mind” to comply in intelligent, helpful, and expected ways.
You can see this in the different kinds of strategy that will be necessary across the instructions here:
Take the kids to school.
Take the kids to the beach.
Take the kids to ice cream.
Take the kids to dinner.
The instructions leave open the transportation modality, whether the addressee is expected to join the kids in the activities, or even accompany them. It takes effort to build automatic systems that recover such distinctions, as you’ll recognize if you saw one of this spring’s viral videos about the failures of chatbot advice. Apparently, state-of-the-art systems do not always infer that when you go to the car wash, you should probably take your car.
Nowadays, linguistics and psychology are only the beginning. As the field considers more powerful AI systems with more pervasive effects, work on alignment has come to overlap with the concerns of the social sciences more broadly, including economics, political science, and even anthropology. For example, when we imagine designing an AI agent so that it responds to the interests of diverse stakeholders, we must reflect on the inherent limitations captured, for example, by Kenneth Arrow’s impossibility theorem for social choice.
As Michael Morreau writes in the Stanford Encyclopedia of philosophy, Arrow’s result generalizes classical paradoxes of voting and
manifests a much wider problem with the very idea of collecting many individual preferences into one. On the face of it, anyway, Arrow’s theorem tells us that collective decisions cannot in general be made on the basis of a common preference, that assimilates the tastes and values of all the individual men and women who make up a society.
A sobering thought for a workshop on the 4th of July.
In the fishbowl conversation, we heard panel presentations by Malihe Alikhani, Lauren M.E. Goodlad and Michael McCarrin, sharing new interdisciplinary insights and challenges for NLP. We then opened the floor, hosting a wide-ranging conversation about the political economy and discursive frameworks surrounding AI and alignment, the ethical and ontological status of the anthropomorphization that confronts us when we speak of AI “choices” and AI “values,” and the reception for interdisciplinary work in AI communities and at AI conferences. This conversation highlighted what we at Rutgers think of as “critical AI” perspectives—a term in which “critical” never means a thumbs down or even a negative approach, but rather perspectives that emphasize judgment, knowledge, and empowerment.
The fishbowl grew out of a plan to launch our discussions with provocations from Stanford’s Michele Elam, but an unfortunate loss of a family member prevented her from joining us. As some may know, Professor Elam is a leading literary critic, Senior Vice Provost for Undergrad Education, and a senior fellow at Stanford’s Institute for Human-Centered AI. She is also the author of an award-winning essay, “Poetry Will Not Optimize: Or What Is Literature To AI?” In the next issue of Critical AI, Elam will take a turn as guest “humanist in the loop,” arguing for “Slow AI” as an alternative to “AI Fast,” and writing eloquently of the critical “need to place the arts, humanities, and social sciences dead center in conversations about the design and impact of emerging technologies on self, society, and the human condition.”
The arts represent an additional source of transdisciplinary visions for AI. At Rutgers, we are familiar with such visions through our collaborations with the Design Justice Network, whose work on community-led, reparative, and transformative technology and design practices have been a longstanding inspiration for our research, pedagogy, and outreach.
Design Justice practitioners often cite the utopian, Afrofuturist fiction of Octavia Butler as a guiding vision. It took me a long time to appreciate where they were coming from. Butler’s most famous writings, such as Parable of the Sower, are set in the ruins of our own civilization, which gives them a superficially dystopian cast. But actually, of course, the most frightening dystopias are those that imagine the continuation of today’s most disturbing trends—I think of the technological “jackpot” of Willian Gibson’s The Peripheral, in which humanity realizes a future of abundance and stewardship too late, when only the most misanthropic oligarchs are left to enjoy it, or the apocalypse-by-business-model of Charlie Stross’s Accelerando.
When they aren’t reminding us that the market can stay irrational longer than we can remain solvent, economists are also quick to point out that if something cannot go on forever, it will stop. Marginalized people across the world already expect to have to rebuild the world on our wreckage. If the only way out is through, Butler’s narratives do limn a world we could want on the other side.
At some point as researchers, we must get down to brass tacks. In this workshop, it was after lunch. We started with a long session getting more precise about the technical challenges surrounding current practices of alignment. Su Lin Blodgett presented collaborative work designing chatbots that defuse users’ intuitive anthropomorphizing assumptions. It’s important work on alignment because it shows how different users are in their conceptions, reactions, and preferences around AI—and because it shows telling limitations of existing LLM alignment methods.
Idris Abdulmumim then sketched work-in-progress towards adapting LLM alignment techniques to the languages and cultures of Africa. The NLP community is increasingly aware of the challenges of applying the data-profligate technologies of the English-speaking world to settings where such resources are scarce and even sharing them can exacerbate inequalities of power and self-determination. Alignment raises the further challenge of resisting the Western idea of an autonomous, self-interested individual in favor of indigenous frameworks of collective reciprocity, as embodied, for example, in the Bantu philosophy of Ubuntu.
Finally, we were honored to have Prof David Leslie of Queen Mary University London give an invited keynote entitled “Intelligence just ain’t in the head: On misrecognition and mismeasurement in AGI discourse.” David’s talk challenged the idea that mere decision-making is a pathway to human-like (or more than human-like) intelligence—arguing, in his words, that
this narrow and speculative picture systematically erases the embodied, enactive, distributed, social, cultural, and ethical dimensions of intelligence and cognitive agency that more coherently describe these terms under their biological and socially enacted auspices.
After the break, we returned to take stock and plan next steps. We shared experiences about the challenges of nurturing academic contributions in an age of giant corporate labs, and of the setbacks, the grind, and the more-than-offsetting rewards of community-engaged research. Participants saw an opportunity to broaden alignment tracks at ACL conferences to more clearly welcome critical and interdisciplinary research.
Participants also debated the possibility of shifting the field’s tasks and metrics away from sweeping but vague assessments of agency, towards the evaluation of specific linguistic abilities—contextualization, disambiguation, clarification, paraphrase—that needn’t position the system as a human-like interlocutor or assistant. There was controversy: Datasets that inspire more “aligned” machine learning research would be exciting, but the nuance of applications and social needs tends to be lost in benchmarking, often leading to a sense among ML researchers that some tasks are “solved” even when performance is not good enough for methods to be used in practice.
I close on a personal note. I’ve just taken on the role of Dean of Mathematical and Physical Sciences at the School of Arts and Sciences at Rutgers. Obviously, my time for research going forward will be limited. I have little choice but to try to encourage others to do the work I’d like to see. You might think that my interest in convening meetings like BATCH is purely directed to developing better AI. In fact, across my career, I have probably had more success in incorporating insights from AI into cognitive sciences than the other way around. Alignment is, I think, an area ripe for such mutual influence.
To prepare for the workshop I read Agnes Callard’s masterful philosophy monograph Aspiration. Callard is trying to do justice to the everyday intention (going back to Aristotle) that the good life demands the cultivation of virtue, which centers, above all, on the disciplined pursuit of richer, more capacious, and more compelling apprehensions of the good. This is learning, she argues, but not learning on the model of scientific inquiry. It is the opening of oneself to the “transformative experience” through which one finally becomes one’s best self.
It sounds great… but I don’t think anyone knows what it really means. I’m not one to typically acknowledge the prospect of superintelligence, but I can tell you this: if humanity does someday develop superintelligence, I want it not just to be superhumanly smart and capable. I want it to be superhumanly good. If its values are to be as more-than-human as its problem-solving, we can’t build it merely to respect what we today think is right (as the most facile understanding of “alignment” seems to imply). That would foreclose the possibility of its—and our—moral progress. A sympathetic reading of Callard suggests that we should design systems not just to “align” with those values we can currently operationalize, but also to assist and accompany us in our efforts of aspiring for more.
I have no idea if we can do such a thing, but I am sure that trying will tell us as much about humanity as it tells us about AI—if only because it will force us to be precise and reflective about what we’re really doing when we sensitize ourselves to a richer range of values, and when we cultivate selves that can uphold those values in relationship with others with commitment and integrity.
Notes
[1] Any opinions, findings, and conclusions or recommendations expressed in this blog post are of course those of the author and do not necessarily reflect the views of the National Science Foundation.