UMW's Rao Pitches Multi-Agent Debate to Replace Monolithic AI Models

AI and the Humanities: Through the Lens of AI Pluralism With University of Mary Washington — Ai4

Ai4

Structuring AI systems around competitive rhetorical debate eliminates black-box opacity and cultural bias while providing educators with an un-fakeable method for student assessment.

AI-assisted summary of AI and the Humanities: Through the Lens of AI Pluralism With University of Mary Washington — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published by Ai4 on YouTube on August 28, 2025, this presentation features Anand Rao, Director of the Center for AI and the Liberal Arts and Professor of Communication and Digital Studies at the University of Mary Washington. The session description highlights how debate and the framework of AI pluralism can transform classroom assessment methods while modeling transparent, aligned, and multi-perspective artificial intelligence.

## TL;DR Anand Rao argues that higher education must move past ineffective bans and probabilistic AI detectors by embracing oral communication and rhetorical debate as authentic assessment methods. Drawing parallels to historical media panics from Plato to the printing press, he demonstrates how competitive debate builds durable cognitive skills and student agency rather than outsourcing thought to generative tools. Rao critiques the tech industry's pursuit of monolithic AI models and proposes 'AI Pluralism,' using multi-agent debate architectures to ensure transparency, explainability, and contestability. Ultimately, the talk positions the humanities not as passive casualties of technological disruption, but as vital leaders in designing ethical and democratic AI systems.

## Atmosphere & style The presentation delivers a measured, intellectually reassuring, and optimistic academic narrative. Rao speaks with the grounded clarity of a veteran educator and collegiate debate coach, shifting the conversation away from modern educational panic through reasoned historical analogies and concrete classroom successes. The pacing builds smoothly from classical philosophical critiques to multi-agent technical frameworks, leaving listeners with an empowering, constructive vision for the future of liberal arts in an automated world.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=4s) Anxieties over generative AI mirror classical historical fears, such as Socrates warning in Plato's Phaedrus that writing would erode human memory and Abbot Trithemius arguing that the printing press would destroy the spiritual discipline of copying manuscripts by hand. - [00:04](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=4s) Douglas Adams's rule of technology illustrates why institutions resist rapid innovation: technology invented after age 35 is often perceived as unnatural, fueling defensive reactions among educators. - [03:05](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=185s) Competitive debate demonstrates how AI can augment cognitive agency rather than replace it, allowing novice debaters to rapidly research, brief, and strategize without delegating active thinking during live competition. - [05:00](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=300s) Institutional bans, AI-free zones, and automated AI detection tools fail because detectors are probabilistic, produce false positives, and disproportionately penalize non-native English speakers and developing writers. - [06:54](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=414s) Returning to the oral tradition and in-class debate provides authentic assessment that cannot be faked with AI, requiring students to synthesize research, understand both sides of an issue, and adapt to live rebuttals. - [10:36](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=636s) The prevailing industry drive toward 'monolithic AI' flattens human nuance through homogeneous training data and reward models that enforce an artificial consensus. - [12:08](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=728s) AI Pluralism offers an alternative framework grounded in the humanities, utilizing multi-agent debate pipelines—like affirmative, negative, and judge agents—to incorporate belief revision and legitimate perspective diversity. - [12:08](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=728s) Pluralistic AI architectures can be operationalized through three distinct models: Overton pluralistic, steerable pluralistic, and distributional pluralistic models. - [15:01](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=901s) Debate-driven multi-agent systems crack open black-box foundation models by forcing agents to state their reasoning (explainability) and providing procedural mechanisms for human operators to challenge model decisions (contestability). - [16:27](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=987s) The Center for AI and the Liberal Arts at the University of Mary Washington is developing open-source debate tools and pluralistic multi-agent reasoning architectures to bridge educational pedagogy with technical alignment. - [17:45](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=1065s) Rather than acting as defensive observers, the humanities must lead AI development by framing artificial intelligence as a cultural, ethical, and pluralistic project.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=4s) Historical Panics: From Plato's Phædrus to the Printing Press

Anand Rao introduces his background as a professor of communication and digital studies and director of the newly launched Center for AI in the liberal arts at the University of Mary Washington. Reflecting on his year-long research sabbatical, Rao opens his inquiry into generative artificial intelligence by grounding it in classical historical anxieties over emerging educational media. He cites Socrates in Plato's *Phaedrus*, who warned that the invention of writing would produce students filled "not with wisdom but with a conceit of wisdom" who would end up being a burden because they were given information without true instruction. Socrates feared that reliance on the written word would destroy the oral tradition, erode human memory, and substitute shallow information retrieval for genuine intellectual discipline.

Fast-forwarding to the late 15th century, Rao draws a direct parallel with Johannes Trithemius, a German abbot who issued dire warnings about Johannes Gutenberg's printing press. Trithemius declared: "He who ceases from zeal for writing because of printing is no true lover of the scriptures." Trithemius maintained that mechanical printing could never match the spiritual discipline and devotion cultivated by monks painstakingly copying religious manuscripts by hand. Rao highlights the inescapable irony embedded in both cases: the world only knows about Socrates' critique because Plato wrote it down, just as Trithemius's tract criticizing printed text was ultimately preserved and distributed as a printed book.

```mermaid flowchart TD accTitle: Historical Media Transitions vs AI subgraph S1["Historical Media Anxieties"] A["Oral Tradition"] -->|"Plato's Critique"| B["Written Word"] B -->|"Trithemius's Critique"| C["Printed Press"] end subgraph S2["Modern Paradigm Shift"] C -->|"Generative AI Disruption"| D["Interactive LLMs"] D -->|"Humanistic Framework"| E["AI Pluralism & Rhetorical Debate"] end ```

To explain the institutional anxiety gripping higher education today, Rao cites author Douglas Adams's observation on human reactions to technology: anything in the world when you are born is normal; anything invented between ages 15 and 35 is revolutionary and career-making; and anything invented after age 35 is seen as "against the natural order of things." While acknowledging that both Plato and Trithemius were partially right about the disruptive friction of new technologies, they were ultimately wrong about the final outcomes. Rao emphasizes that while generative AI introduces genuine pedagogical risks, higher education should avoid outright resistance and instead look for ways to preserve core humanistic values.

- **Key Historical Parallels**: - **Socrates / Plato (*Phaedrus*)**: Feared writing would destroy active memory and replace wisdom with the illusion of knowledge. - **Johannes Trithemius (15th Century)**: Argued the printing press would extinguish the spiritual discipline of handwritten manuscript copying. - **Douglas Adams's Rule of Technology**: Cultural resistance to technology often reflects the age and comfort zone of educators rather than inherent technological flaws.

---

[03:05](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=185s) Accelerating Student Agency: AI in Competitive Debate

Drawing on his background as a high school and college debater and ten years as a collegiate debate coach, Rao identifies competitive debate as a prime vehicle for showing how AI can augment rather than replace student learning. Rather than using AI to bypass critical thought, competitive students utilize large language models to rapidly scale steep learning curves without outsourcing their cognitive agency.

Rao shares a case study from the past year involving novice high school debaters on his daughter's debate team. With only a few weeks to prepare for their very first tournament on an unfamiliar topic, these students leveraged a stack of AI tools—including Google Gemini, ChatGPT, and Claude—to conduct research, draft argumentative briefs, brainstorm counter-responses, and formulate strategies. During the tournament, the students did not rely on AI during active rounds; instead, they used it between rounds (including Perplexity AI) to investigate unfamiliar arguments run by opponents and formulate tactical responses. The novice team advanced all the way to the tournament quarterfinals against significantly more experienced competitors, showcasing how AI can empower deep, self-directed learning.

---

[05:00](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=300s) The Failure of Bans, Detection Tools, and AI-Free Zones

Looking back at the widespread panic following the public release of ChatGPT in late 2022, Rao critiques the early administrative impulse to construct "AI-free zones" and rely on automated AI detection software. He notes that AI detectors are merely probabilistic and suffer from severe technical and ethical limitations, generating unacceptable rates of false negatives and false positives. Most harmfully, detection algorithms disproportionately flag writing from vulnerable student populations, including non-native English speakers (ESL) and students with developing foundational writing skills.

```mermaid flowchart LR accTitle: Pitfalls of Administrative Resistance to AI subgraph Ineffective["Failed Defenses"] A["AI Detectors"] -->|"Probabilistic Flaws"| B["False Positives for ESL Students"] C["AI-Free Zones"] -->|"Impracticality"| D["Classroom-World Disconnect"] end subgraph Constructive["Authentic Approach"] E["Oral Assessment & Debate"] -->|"Demands Direct Mastery"| F["Durable Cognitive Skills"] end ```

Beyond technical unreliability, Rao argues that attempting to isolate the classroom from generative AI fails students pedagogically. Higher education is traditionally defined by slow, methodical reflection, but the velocity of modern technological change requires proactive curricular adaptation. Shielding students behind artificial bans denies them the opportunity to cultivate the critical, ethical, and cognitive frameworks necessary to navigate a society saturated with automated systems.

---

[06:54](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=414s) Oral Tradition and Debate as Authentic Assessment

To solve the challenge of verifying student learning without relying on surveillance or flawed detection tools, Rao advocates returning to oral communication. Leaning into his 25 years of work with the University of Mary Washington's 30-year-old Speaking Intensive program and its Speaking Center, Rao argues that oral performance offers an un-fakeable window into student comprehension. While a standard presentation with a robust Q&A provides solid evidence of mastery, an in-class debate requires an even higher level of cognitive command: students must understand both sides of a complex issue, anticipate counter-arguments, collaborate in real time, and adapt to live rebuttals.

Rao links in-class debate to durable, humanistic capabilities that remain indispensable in the workforce. Referencing recent articles in *Security Management* and *Inc.* magazine regarding the resurgence of the liberal arts degree, he stresses that technology leaders and employers prioritize communication, critical inquiry, empathy, and leadership. Because the humanities are fundamentally pluralistic—exploring human culture, history, language, philosophy, and religion through dialogue and contrasting interpretations—humanities education is uniquely equipped to prepare students for an AI-driven economy.

- **Benefits of Debate-Centered Assessment**: - **Authentic Verification**: Impossible to delegate to an automated text generator in real time. - **Comprehensive Subject Mastery**: Requires grasping the merits and flaws of opposing positions. - **Development of Durable Skills**: Fosters critical reasoning, collaborative problem-solving, rapid research synthesis, and verbal agility. - **UMW Speaking Intensive Model**: Backed by three decades of faculty development, learning communities, and open-source curricular rubrics.

---

[10:36](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=636s) The Danger of Monolithic AI Systems

Rao transitions from classroom pedagogy to the architectural direction of artificial intelligence itself, arguing that humanities methodologies can guide future AI design. He critiques the prevailing industry drive toward "monolithic AI"—the pursuit of "one model to rule them all" where competing labs race toward a single supreme model rather than creating ecosystems that support multiple coexisting viewpoints.

```mermaid flowchart TD accTitle: Monolithic AI vs AI Pluralism subgraph Monolithic["Monolithic AI (Current Paradigm)"] M1["Homogeneous Training Data<br>(Western, English, College-Educated)"] --> M2["Global Alignment Loss Function"] M2 --> M3["Flattened Monolithic Output<br>(False Consensus)"] end subgraph Pluralistic["AI Pluralism (Proposed Framework)"] P1["Diverse Multi-Agent Architecture"] --> P2["Rhetorical Argumentation & Debate"] P2 --> P3["Transparent, Contestable Decision-Making"] end ```

This monolithic paradigm enforces artificial homogeneity through two primary mechanisms: 1. **Training Data Bias**: Models are disproportionately trained on Western, college-educated, English-language text, creating systematic cultural blind spots. 2. **Alignment Objectives**: Reward models tuned for broad acceptability wash out distinct subcategories of perspective and nuanced values in pursuit of universal inoffensiveness.

---

[12:08](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=728s) Operationalizing AI Pluralism and Multi-Agent Debate

As an alternative to monolithic design, Rao introduces the concept of **AI Pluralism**, which adapts the pluralistic traditions of the humanities to represent legitimate diversity rather than ungrounded moral relativism. To illustrate this in practice, Rao highlights DebbHub, an NSF-funded platform developed by John Hines and computer scientist Devon Garnier to support urban debate leagues through AI scaffolding.

Garnier constructed an experimental multi-agent debate architecture where distinct instances of AI models are assigned discrete roles: affirmative debaters, negative debaters, and an impartial judge. Within this pipeline, the models independently research evidence, construct argumentative strategies, execute cross-examinations, deliver rebuttal speeches, and receive an evaluative decision from the judge agent. Crucially, the system incorporates "belief revision" and memory mechanisms, allowing agents to evaluate competing perspectives, revise their internal representations based on sound argumentation, and carry those nuanced beliefs forward into subsequent decision-making contexts.

```mermaid flowchart LR accTitle: Multi-Agent Debate Architecture subgraph DebateRound["Multi-Agent Debate Pipeline"] Aff["Affirmative AI Agent<br>(Constructive & Rebuttal)"] <-->|"Cross-Examination & Clash"| Neg["Negative AI Agent<br>(Constructive & Rebuttal)"] Aff --> Judge["AI Judge Agent<br>(Evaluation & Adjudication)"] Neg --> Judge end Judge --> Mem["Belief Revision & Memory Update"] Mem --> Next["Downstream Alignment & Decision Tasks"] ```

Rao argues that integrating **rhetorical argumentation**—which models audience adaptation, context, and legitimate disagreement—provides a necessary complement to traditional **computational argumentation**, which often defaults back to monolithic assumptions. He outlines three primary models for deploying pluralistic AI:

| AI Pluralism Model | Core Mechanism & Behavior | |:--- |:--- | | **Overton Pluralistic Model** | Surfaces a spectrum of acceptable and reasonable viewpoints across a defined debate window. | | **Steerable Pluralistic Model** | Explicitly conditions outputs on specific transparent values, personas, or cultural perspectives. | | **Distributional Pluralistic Model** | Proportionately reflects the statistical distribution of opinions found across a diverse human population. |

Rao concludes by pointing out that these debate-driven pluralistic architectures offer one of the most promising frontiers for AI alignment, replacing centralized, top-down guardrails with transparent, contestable deliberation.

[15:01](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=901s) AI Alignment, Explainability, and Contestability Through Debate

Addressing the broader AI community, the conversation shifts to the urgent challenge of AI alignment: ensuring that as artificial intelligence advances in capabilities, it aligns with human values rather than running counter to them. Current machine learning architectures make this task exceptionally difficult because foundation models operate largely as "black boxes." Engineers and researchers understand how the architectures are set up and how training pipelines are constructed, but the internal latent representations and mechanisms that yield any specific output remain opaque.

A debate-centered model grounded in AI pluralism offers two vital mechanisms to crack open this black box: **explainability** and **contestability**. When systems are structured around debate rather than single-model monolithic outputs, individual AI agents are forced to communicate with one another, share perspectives, and expose their line of reasoning. This dynamic surfaces the underlying arguments and decision-making pathways, granting stakeholders much greater transparency. Second, it establishes contestability. Beyond simply knowing how a machine reached a conclusion, human operators and communities must have the procedural means to challenge a decision when it conflicts with local human values. Within a pluralistic, multi-agent setup where models negotiate and debate rather than impose a single monolithic decree, both transparency and contestability become structural prerequisites rather than afterthoughts.

```mermaid flowchart TD accTitle: AI Pluralism Alignment Model accDescr: Diagram showing how a pluralistic debate architecture yields explainability and contestability for value alignment.

subgraph Problem["Black Box Challenge"] A["Opaque AI Models"] --> B["Uncertain Value Alignment"] end

subgraph Pluralism["Debate-Centered Multi-Agent System"] C["Agent Perspective A"] <-->|"Deliberation & Debate"| D["Agent Perspective B"] end

subgraph Mechanisms["Alignment Mechanisms"] E["Explainability<br/>(Visible Arguments & Reasoning)"] F["Contestability<br/>(Actionable Challenge by Humans)"] end

B -.-> C Pluralism --> E Pluralism --> F E & F --> G["Transparent, Community-Aligned AI"] ```

- **Black box opacity**: Developers know the training regimens and parameters, yet cannot cleanly trace why a model generated a particular output. - **Explainability**: Debate requires agents to state reasons and critique counterarguments, generating an inherent paper trail of the rationale. - **Contestability**: Pluralism provides a deliberate procedural opening for stakeholders to counter, audit, and push back against model choices.

---

[16:27](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=987s) Next Steps: Publications and the Center for AI and the Liberal Arts

Turning to concrete initiatives and scholarly roadmaps, two key publications highlight this approach. The first, released in the spring of 2023, is a co-edited collection addressing the authentication crisis in education, gathering educators together to rethink assessments in the face of generative tools. The second volume, slated for release in the fall, is co-edited by Rao and titled *AI Pluralism: Fostering Diversity and Dialogue in Artificial Intelligence*, which sets out the theoretical and technical framework for pluralistic models.

On an institutional level, this work forms the backbone of the University of Mary Washington's newly founded Center for AI and the Liberal Arts, which was officially announced earlier that week. The center is launching with an active schedule of programming and events, operating on a public, open-source model to disseminate research, pedagogic frameworks, and technical tooling. Its objectives bridge both short-term pedagogical needs and long-term computational engineering:

- **Debate facilitation tools**: Developing and releasing near-term tools that students and faculty can implement directly in classroom activities. - **Multi-agent demonstrations**: Building and testing practical multi-agent debate pipelines to validate pluralism in decision-making contexts. - **Embedded pluralistic reasoning**: Designing long-term architectures that incorporate multi-perspective, contestable deliberation directly into future AI systems.

---

[17:45](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=1065s) The Humanities as Leaders in AI Development

Reflecting on technological transitions throughout history, the discussion emphasizes that societal fears over disruptive technologies routinely overlook rare moments for systemic renewal. The humanities hold a crucial vantage point because their traditional discipline is centered on valuing pluralism, evaluating divergent perspectives, and navigating conflicting interpretations. Applying humanistic models directly to technical development equips computer science with the frameworks necessary to represent complex, diverse human values instead of flattening them into singular, centralized outputs.

Rhetorical debate acts as the connective tissue between authentic, un-cheatable student learning on the one hand, and explainable, responsible artificial intelligence on the other. Rather than treating AI strictly as an engineering phenomenon, society must approach it as an expansive "cultural and ethical project." Handled with this mindset, the humanities are not merely reacting and adapting defensively to software breakthroughs, but actively directing the future trajectory of AI toward transparent, democratic ends. The session wraps with an invitation to connect via LinkedIn or email, an offer to answer audience questions afterward, and gratitude to attendees of the conference's education track.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=4s) not with wisdom but with a conceit of wisdom — Anand Rao

> [00:04](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=4s) He who ceases from zeal for writing because of printing is no true lover of the scriptures. — Johannes Trithemius

> [00:04](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=4s) against the natural order of things. — Douglas Adams

> [05:00](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=300s) AI-free zones — Anand Rao

> [10:36](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=636s) one model to rule them all — Anand Rao

> [12:08](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=728s) AI Pluralism — Anand Rao

> [15:01](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=901s) black boxes — Anand Rao

> [17:45](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=1065s) cultural and ethical project — Anand Rao

## Mentioned Anand Rao, Center for AI and the Liberal Arts, University of Mary Washington, Plato, Phaedrus, Socrates, Johannes Trithemius, Johannes Gutenberg, Douglas Adams, Google Gemini, ChatGPT, Claude, Perplexity AI, Security Management, Inc., DebbHub, NSF, John Hines, Devon Garnier, AI Pluralism: Fostering Diversity and Dialogue in Artificial Intelligence

## Video description Anand Rao, Director of the Center for AI and the Liberal Arts and Professor of Communication and Digital Studies, University of Mary Washington

Abstracts

What if debate could simultaneously transform our classrooms and our AI systems—creating both authentic assessment methods for today's educators and models for more transparent, aligned artificial intelligence in the future? I will explore AI pluralism as a framework for integrating artificial intelligence with humanities education while preserving core humanistic values of critical thinking, ethical reasoning, and diverse viewpoints. I will demonstrate how rhetorical argumentation and debate among AI agents can create more transparent, accountable systems that enhance rather than diminish learning environments. Drawing from my research on AI explainability and contestability, I will showcase how debate emerges as one of the most effective approaches for authenticating student learning in an era where traditional assessments are increasingly vulnerable to AI-generated content. Attendees will learn practical techniques—such as implementing AI-facilitated debate activities—that humanities educators can immediately apply to foster authentic assessment and critical thinking skills. Join this presentation to discover how to bridge technological innovation with humanistic inquiry in ways that create more inclusive, ethically-grounded educational experiences.

[Source](https://www.youtube.com/watch?v=QWWnNU-QYvY)

中文简报

UMW学者Rao提议用多智能体辩论替代单体AI模型 围绕竞技性修辞辩论构建AI系统不仅能消除黑箱不透明性与文化偏见，还能为教育工作者提供一种无法作伪的学生评估方法。 Anand Rao认为，高等教育必须告别无效的禁令和基于概率的AI检测工具，转而将口语交流与修辞辩论作为真实的评估手段。通过类比从柏拉图到印刷机问世的历史性媒介恐慌，他阐明了竞技性辩论如何培养持久的认知能力和学生的主体性，而不是将思考外包给生成式工具。Rao批评了科技界对单体AI模型的追求，并提出“AI多元主义”，主张利用多智能体辩论架构来确保透明度、可解释性与可争辩性。最终，这场演讲将人文学科定位为设计伦理与民主化AI系统的关键引领者，而非技术颠覆下的被动受害者。 对生成式 AI 的焦虑呼应了历史上对新媒介的恐慌（如苏格拉底担忧文字削弱人类记忆、特里特米乌斯指责印刷术破坏抄经修道纪律），而封禁策略与自动化检测工具也因高误报率及对非英语母语者的系统性偏见而走向失败。 (0:04) 回归口头表达传统与课堂辩论为教育界提供了无法被 AI 伪造的真实评估方式，迫使学生深入理解议题的正反双方、实时整合研究成果并进行敏捷的反驳应对。 (6:54) 针对当前主流工业界追求“单体 AI”（Monolithic AI）而抹杀人类认知细微差异的现状，“AI 多元主义”（AI Pluralism）框架提出通过正方、反方与裁判组成的多智能体辩论流水线，实现观点的合法多样性与信念修正。 (12:08) 多智能体辩论机制为不透明的黑箱模型提供了“可解释性”与“可质辩性”，不仅展现了决策背后的论证路径，还赋予利益相关者依据自身价值观质疑与审查 AI 结论的程序化途径。 (15:01)

Reading summary

Anand Rao, director of the Center for AI in the Liberal Arts at the University of Mary Washington, argues that higher education and computer science should adopt humanistic traditions of debate and pluralism rather than defensive bans or surveillance. Rao posits that organizing pedagogy and model architecture around multi-perspective argumentation addresses both the academic assessment crisis and the risks of opaque, monolithic artificial intelligence. Historical Panics and the Failure of AI Bans (0:04) Drawing parallels to Socrates’s warning in Plato's Phaedrus that the written word would erode memory and Johannes Trithemius’s 15th-century claim that printing would ruin monastic scribal discipline, Rao notes that resistance to new educational media follows a long historical pattern. In both instances, critics paradoxically relied on the technologies they condemned to preserve their arguments. Rao cites author Douglas Adams to explain that institutional backlash often reflects generational discomfort rather than fundamental pedagogical flaws. In the wake of ChatGPT's launch, institutions attempted to enforce 'AI-free zones' and deploy automated detection software. Rao critiques these measures as technically unreliable and pedagogically damaging. Because AI detectors are probabilistic, they generate severe false positives that disproportionately harm non-native English speakers and students with foundational writing challenges. Rao argues that attempting to isolate classrooms from generative systems denies students the chance to cultivate the critical and ethical frameworks required in an automated society. Oral Debate as Authentic Assessment and Learning Accelerator (3:05) To evaluate student mastery without automated surveillance, Rao recommends turning to oral communication and debate, drawing on three decades of practice in the University of Mary Washington's Speaking Intensive program. Live debate requires students to comprehend multiple facets of complex issues, anticipate rebuttals, and adapt in real time—skills that cannot be outsourced to a large language model during a live exchange. Rao references reports in Security Management and Inc. highlighting an employer resurgence in demand for liberal arts competencies like communication, empathy, and critical inquiry. Rao illustrates how AI can augment rather than replace student learning through a case study of novice high school debaters. Using an AI stack—including Google Gemini, ChatGPT, Claude, and Perplexity AI—students conducted rapid research, drafted briefs, and investigated unfamiliar arguments between rounds. The novice team advanced to tournament quarterfinals against experienced competitors, demonstrating that structured AI use can accelerate steep learning curves while keeping cognitive agency in student hands. Replacing Monolithic AI with Multi-Agent Pluralism (10:36) Beyond pedagogy, Rao argues that the humanities can correct the prevailing computational trend toward 'monolithic AI'—the pursuit of a single dominant model. He asserts that monolithic systems enforce artificial homogeneity through training data skewed toward Western, English-speaking, college-educated demographics, coupled with global alignment tuning that suppresses nuanced cultural perspectives to produce bland inoffensiveness. As a counterweight, Rao introduces 'AI Pluralism,' which incorporates diverse perspectives via multi-agent architectures. He highlights DebbHub, an NSF-funded platform created by John Hines and Devon Garnier that arranges distinct AI agents into affirmative debaters, negative debaters, and an evaluative judge. Crucially, the system uses belief revision and memory updates so agents can evolve their stances based on argument quality. Rao outlines three pluralistic operational models: the Overton model (presenting a defined window of acceptable views), the Steerable model (conditioning responses on transparent personas or values), and the Distributional model (reflecting statistical population opinion distributions). Achieving Explainability and Alignment Through Deliberation (15:01) Rao addresses the broader AI alignment challenge, arguing that foundation models remain dangerous 'black boxes' whose internal latent operations cannot be cleanly traced by developers. Structuring AI systems around multi-agent debate addresses this opacity by enforcing two vital alignment mechanisms: explainability and contestability. Deliberation forces competing models to articulate reasoning and critique opposing evidence, naturally exposing their decision-making logic. Furthermore, contestability guarantees that human operators and communities retain the procedural means to challenge outputs that conflict with local values. These principles anchor the University of Mary Washington's newly founded Center for AI and the Liberal Arts, alongside Rao's forthcoming co-edited book, AI Pluralism: Fostering Diversity and Dialogue in Artificial Intelligence. The open-source center focuses on developing classroom debate facilitation tools, building multi-agent demonstration testbeds, and engineering long-term architectures that embed rhetorical deliberation into downstream AI decision systems. Rao concludes that treating artificial intelligence as an expansive cultural and ethical project allows the humanities to guide technical development rather than react defensively to software disruption. However, successfully implementing this framework depends on moving multi-agent debate pipelines from experimental prototypes into scalable, contestable systems capable of handling real-world institutional decisions.

精炼阅读摘要

玛丽华盛顿大学（University of Mary Washington）教授阿南德·拉奥（Anand Rao）指出，高等教育应对生成式人工智能不应诉诸禁令或不可靠的检测工具，而应依托人文学科的多元对话传统，通过口语辩论重塑教学评估，并将“AI多元主义”（AI Pluralism）与多智能体辩论机制引入模型架构，以化解单体大模型的文化偏见与黑箱对齐困境。 历史技术焦虑与AI禁令检测的局限 (0:04) 拉奥在考察当下教育界对生成式人工智能的恐慌时，首先将其置于媒介演进的历史语境中。柏拉图在《斐德罗篇》（Phaedrus）中记载，苏格拉底曾指责文字的发明会摧毁人类记忆力，让学生自负博学却缺乏真正智慧；15世纪末，德国修道院长约翰内斯·特里特米乌斯（Johannes Trithemius）亦撰文警示古腾堡印刷机会破坏僧侣手抄经文的精神修持。拉奥指出这两起历史公案均充满讽刺意味：苏格拉底的批评正是依赖文字记录才流传至今，而特里特米乌斯批判印刷的书卷最终也是经由活字印刷得以保存。正如作家道格拉斯·亚当斯（Douglas Adams）对技术接纳规律的调侃，教育界当前对AI破坏教学秩序的抵触，很大程度上折射了学者和教育者的代际心理防线。 针对2022年底ChatGPT普及后高校构筑“无AI区”及使用自动化AI检测工具的做法，拉奥明确指出了其技术和伦理缺陷。自动化检测软件本质上是概率推断，存在极高的假阳性与假阴性误差，尤其对非英语母语学生（ESL）和处于基础写作发展阶段的学生造成系统性偏见与误判。拉奥主张，在自动化系统广泛渗透的社会背景下，高校试图通过行政禁令将课堂与AI隔绝不仅行不通，而且剥夺了学生建立技术批判思维与伦理驾驭能力的关键机会。 竞技辩论赋能学习与口语真实评估 (3:05) 结合自身中学、大学辩论经历及十年的大学辩论执教生涯，拉奥强调辩论能够证明AI如何增强而非剥夺学生的主体认知能力。拉奥以其女儿所在的高中新手辩论队为例，这批学生在备赛首场陌生议题比赛时只有数周时间，通过组合使用Google Gemini、ChatGPT和Claude等工具开展文献研究、拟定论点提纲与头脑风暴，并在多轮比赛的间歇利用Perplexity AI快速剖析对手论点。学生们在正赛发言环节完全脱离AI，仅在场外将工具作为认知杠杆，最终在经验更丰富的赛场上打入八强（quarterfinals），证实了AI对深度自主学习的加速价值。 为了绕过失效的AI检测软件并真实检验学习成效，拉奥提倡重返口语传播体系。依托玛丽华盛顿大学推行30年的“强化口语表达”（Speaking Intensive）项目与口语中心经验，拉奥认为口语互动构成了无法被自动化工具实时伪造的考核窗口。与常规陈述相比，课堂对抗性辩论要求学生必须熟知对立立场、预判反驳并进行实时协作应对。拉奥援引《Security Management》与《Inc.》关于文科学位复兴的报道指出，当前用人单位与技术领袖高度看重沟通、批判性探究、同理心与领导力等核心素养，而人文学科的辩论训练恰恰能够培养这些不可替代的持久能力。 单体AI的偏见与多元智能体辩论架构 (10:36) 拉奥将讨论从教学策略延伸至人工智能的基础架构设计，批评当前科技行业盲目追求“统领一切的单体模型”（Monolithic AI）。这种竞相打造超级单一模型的范式导致了严重的人为同质化：其一是训练数据严重偏向受过高等教育的西方英语文本，形成内生性文化盲区；其二是追求全域中立与通用合格的对齐奖励机制，往往抹平了不同群体的细分观点与特定价值取向，制造出虚假的群体共识。 作为替代路径，拉奥提出了植根于人文学科多元思想传统的“AI多元主义”（AI Pluralism）。他以约翰·海因斯（John Hines）与计算机科学家德文·加尼尔（Devon Garnier）开发的美国国家科学基金会（NSF）资助项目DebbHub为例，展示了多智能体辩论架构的可行性。该系统将不同的AI实例分别设定为正方辩手、反方辩手以及独立裁判智能体，各方自主检索证据、展开交叉质询与反驳，并在裁判裁决后触发“信念修正”（belief revision）与记忆更新机制。拉奥进一步将AI多元主义归纳为三类落地模式：呈现特定议题合理争论光谱的“奥弗顿多元模型”（Overton Pluralistic Model）、依据明确价值观显式定制输出视角的“可控多元模型”（Steerable Pluralistic Model），以及按真实人口比例映射多元观点的“分布多元模型”（Distributional Pluralistic Model）。 博弈破局黑箱对齐与人文学科的引领 (15:01) 在解决AI价值对齐难题方面，拉奥认为多智能体辩论机制能有效破解基础模型底层潜变量不透明的“黑箱”困境，并带来“可解释性”（Explainability）与“可争议性”（Contestability）。在多智能体辩论模型中，系统不再直接输出单一裁决，而是迫使各个AI智能体互相辩驳并公开其论证推演过程，从而形成清晰完整的推理链条；同时，该机制为人类运营者和社区提供了程序化质疑窗口，当机器结论违背本地价值标准时，利益相关者能够发起实质性抗辩与审计。 为了将该框架付诸工程与教学实践，玛丽华盛顿大学正式启动了“文理AI中心”（Center for AI and the Liberal Arts），通过开源模式开发课堂辩论工具并搭建多智能体辩论验证平台。拉奥联合主编的专著《AI多元主义：在人工智能中促进多样性与对话》亦计划于秋季出版。拉奥强调，人工智能不应单纯被视为一项计算机工程课题，而是一场广泛的文化与伦理工程；人文学科不应被动防御，而应当凭借其对多样性、语境和分歧的处理经验，积极引领未来AI技术走向透明与民主对齐。 拉奥的论述阐明了人文学科从防御性规避转向主动定义AI发展路径的理论可能，修辞辩论不仅为学术真实性提供了无需监控的评估解法，也为打破单体AI的数据专制开辟了技术架构。然而，多智能体博弈系统在面对极端虚假证据时的纠偏效率、在算力成本上的现实可行性，以及如何与现行主流商业闭源系统对接，仍是有待进一步工程检验的深层课题。

Detailed notes — zh

## Background 本视频由 Ai4 官方频道于 2025 年 8 月 28 日在 YouTube 平台发布，记录了玛丽华盛顿大学传播与数字研究教授、文科人工智能中心主任阿南德·拉奥（Anand Rao）的专题演讲。内容围绕“AI 多元主义”（AI Pluralism）展开，探讨如何将人工智能与通识人文教育深度融合，并借助辩论机制重塑教学评估与推动透明对齐的 AI 系统构建。 ## TL;DR 拉奥教授首先回顾了文字与印刷术引发的历史性技术焦虑，指出高校试图通过检测工具或设立无 AI 禁区应对技术冲击完全是徒劳的。他提出重回口语论辩传统，将课堂辩论作为检验学生真实理解、团队协作与批判性思维的终极评估形式。进一步地，他批判了当前科技巨头争夺“单一垄断型模型”的同质化倾向，主张依托人文学科构建多智能体辩论系统（Multi-Agent Debate System），利用修辞论证为 AI 提供可解释性与可争议性。最终，他呼吁人文学科应从被动应激转向主动引领，将 AI 视作一项文化与伦理工程来重塑其发展轨迹。 ## Atmosphere & style 演讲整体流露出深厚的人文底蕴与前瞻的技术视野，语调风趣幽默、论证严密且充满建设性。讲者巧妙穿插苏格拉底的历史讽刺与指导新手辩手逆袭的鲜活案例，有力化解了教育界对 AI 认知的恐慌与应激排斥。整场分享给人一种极具启发性的文科通识大师课氛围，在数千年古典修辞学传统与前沿多智能体系统之间搭建起一座充满思辨张力与实用价值的桥梁。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=4s) 人类对新兴技术的焦虑贯穿历史，从古希腊苏格拉底担忧文字削弱记忆与智慧，到文艺复兴时期特里特米乌斯抵制印刷术，历史反讽在于这些担忧最终都依赖被批判的新技术才得以流传。 - [00:04](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=4s) 生成式 AI 对教育的冲击无法用过去的经验简单等同，但破局的关键在于重拾具有深厚人文传统的“辩论”机制。 - [03:15](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=195s) 高中辩论队的新手辩手实践表明，将 AI 工具（Gemini、ChatGPT、Perplexity）作为赛前推演和调研的认知放大器而非全盘代劳，能极速提升跨学科学习曲线并深化逻辑理解。 - [05:05](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=305s) 依赖 AI 检测工具是一条死路，其概率模型不仅漏报率高，且对 ESL 国际生和写作薄弱群体存在系统性偏见，高校试图设立“纯净无 AI 隔离区”既不现实也违背培养现实生存能力的初衷。 - [06:54](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=414s) 在 AI 时代实现真实性评估（Authentic Assessment）的最佳途径是回归口语论辩传统；课堂辩论强制要求学生透彻理解正反双向观点，能高维检验批判性思维与持久核心能力。 - [08:59](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=539s) 雇主长期看重的人际沟通与批判思维正是人文学科的核心；人文学科本质上是一种多元主义传统，其对话与同理心素养是释放大语言模型潜力的关键底座。 - [10:56](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=656s) 当前 AI 研发陷入“单一垄断型 AI”（Monolithic AI）困境，受限于西方英语数据偏见和泛化对齐目标；“AI 多元主义”（AI Pluralism）提出以多视角交锋代替单一模型霸权。 - [12:26](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=746s) 多智能体辩论系统（Multi-Agent Debate System）通过正方、反方与裁判 AI 的全流程对抗与信念修正机制，引入“修辞论证”以弥补传统“计算论证”在多元视角调适上的不足。 - [14:54](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=894s) AI 多元主义涵盖奥弗顿多元模型、可操纵多元模型和分布多元模型，为 AI 对齐提供了“可解释性”（透明推导过程）与“可争议性”（赋予人类质疑与协商决策的权力）两大支柱。 - [17:49](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=1069s) AI 不仅是一项技术突破，更是一项深层的文化与伦理工程；人文学科应从被动应对转向主动引领，以辩论为桥梁连接课堂真实性学习与更对齐、可解释的 AI 架构。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=4s) 技术恐慌的千年回响：从苏格拉底、特里特米乌斯到生成式 AI 来自玛丽华盛顿大学（University of Mary Washington）的传播与数字研究教授、文科人工智能中心（Center for AI in the Liberal Arts）主任阿南德·拉奥（Anand Rao）在结束为期一年的学术休假后，分享了他关于高等教育如何适应人工智能时代的研究思考。他首先借由教育技术史上的两次重大焦虑拉开序幕： 早在古希腊时期，苏格拉底在柏拉图的《斐德罗篇》（*Phaedrus*）中就曾对文字这一新兴技术的诞生提出严厉批判，认为文字传授大量信息却不进行真正的教学，会让学生“看似无所不知，实则一无所知”，最终让学生充满虚妄的智力自负而非真正的智慧，成为社会的负担。快进到15世纪末，德国本笃会修道院院长约翰内斯·特里特米乌斯（Johannes Trithemius）在《赞抄写员》（*In Praise of Scribes*）中对印刷术发出了极为相似的警告，断言因印刷术而放弃手抄热情的人算不上真正的经文热爱者，认为印刷书籍永远无法取代手抄本，修道士也将因此丧失在逐字抄写中获得的崇高精神纪律。 拉奥幽默地指出，历史充满了绝妙的讽刺：我们之所以能知晓柏拉图对文字的担忧，恰恰是因为文字被记录了下来；而特里特米乌斯对印刷术的抗议，最终也是通过印刷书籍才得以流传后世。这呼应了作家道格拉斯·亚当斯（Douglas Adams）关于技术认知的经典论述： - 任何在你出生时就已存在的技术，都是自然而平常的； - 任何在你 15 到 35 岁之间发明出来的技术，都是令人振奋的革命性突破，你甚至能以此开创事业； - 任何在你 35 岁之后发明出来的技术，都违背了自然的正常法则。 拉奥调侃道，高校中许多学者正处于这一年龄段，这很大程度上解释了高等教育界对 AI 的本能抵触。虽然柏拉图与特里特米乌斯都担心技术会用“便利”取代“人类的努力”，且他们对风险的预判部分属实、对最终结局的判断完全失误，但拉奥强调，生成式 AI 对现有教育模式的冲击绝不能被简单等同于过去的口语对文字或印刷术之争。然而，破局的希望依然存在——这个希望就是“辩论”（Debate）。 ```mermaid flowchart TD accTitle: 技术演进中的人类认知焦虑流变 subgraph S1["古代：口语转向文字"] A["苏格拉底 / 柏拉图"] -->|"担忧"| B["丧失记忆与真知<br/>滋生虚妄的智力自负"] end subgraph S2["文艺复兴：手抄转向印刷"] C["约翰内斯·特里特米乌斯"] -->|"担忧"| D["丧失抄写精神纪律<br/>印刷品不如手抄本纯粹"] end subgraph S3["当代：传统写作转向生成式 AI"] E["高等教育界与教育者"] -->|"担忧"| F["认知外包与学术作弊<br/>冲击传统学习与评估体系"] end B -.->|"历史反讽：靠文字流传"| G["技术最终融入文明底层"] D -.->|"历史反讽：靠印刷品传播"| G F ==>|"破局路径"| H["回归辩论与口语论辩传统"] ``` --- ### [03:15](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=195s) 辩论赛场上的 AI 实战：新手辩手的跃升范式 拉奥拥有深厚的辩论背景：高中和大学时期曾是辩手，学术生涯的前十年担任专职辩论教练，此后多年持续指导初高中辩论队。他观察到，当充满竞争内驱力的辩手接触到前沿教育技术时，往往能在保持深入理解核心内容的同时，完成极为惊人的进阶。 在刚过去的这一学年中，拉奥指导了女儿高中辩论队里的几位完全没有经验的新手女辩手。面对首场锦标赛，她们只有短短几周时间准备一个完全陌生的辩题。这批极具求胜欲的学生综合运用了 Google Gemini、ChatGPT 和 Claude 等 AI 工具来展开前期调研、起草论辩摘要（briefs）、头脑风暴攻防论点并制定策略。关键在于，她们并没有将认知能力“全盘外包”，而是利用 AI 极速跨越了学习曲线，在消化大量复杂议题的同时深刻理解了辩论逻辑。 在第一场正式比赛中，这支新手队伍一举杀入了四分之一决赛（Quarterfinals），击败了多支经验丰富、胜率更高的老牌队伍。在整场赛事中： - 她们并未在正式辩论发言环节使用 AI，而是利用 AI 进行了高强度的赛前推演与准备； - 走出赛场复盘时，一旦遇到不熟悉的论点，学生会立即使用 Perplexity 等检索工具实时查证背景、拆解论点结构并迅速生成回应方案； - 这种将 AI 作为认知放大器而非替代品的自主实践，正是拉奥希望在大学课堂中推广的理想学习范式。 --- ### [05:05](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=305s) 传统考核的溃败与“AI 无禁区”的现实 回顾 2022 年底 ChatGPT 横空出世的冲击，拉奥坦言自己与绝大多数高校教师一样感到猝不及防。面对学生迅速采纳 AI 的现实，学术界最初的应激反应往往是试图将 AI 拒之门外，并寄希望于“AI 检测工具”（AI Detectors）。但拉奥指出，依赖检测工具完全是一条死路： - **概率性且不可靠**：检测软件本质上只是概率模型，漏报（False Negatives）极为普遍； - **危害极大的误判（False Positives）**：检测工具对写作基础薄弱的学生、以及将英语作为第二语言（ESL）的国际生等弱势群体存在系统性偏见，极易将他们的真实创作错误判定为 AI 生成； - **机构适应周期的断层**：高等教育体系素以“缓慢、条理化反思再到渐进式适应”见长，但生成式 AI 的迭代速度已彻底打破了这一节奏； - **“无 AI 隔离区”的不切实际与不公**：在课堂上设立所谓“纯净无 AI 气泡”不仅技术上无法实现，更是对学生未来的不负责任。大学的核心职责是培养学生的认知能力与智力深度，帮助他们为现实世界的生活做好准备，假装社会中如此庞大的技术力量不存在无异于掩耳盗铃。 --- ### [06:54](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=414s) 重回口语传统：辩论作为终极的真实性评估 为了在 AI 时代实现真实的学生学习成果评估（Authentic Assessment），拉奥提出回归自己的专业原点——传播学与口语论辩传统。既然书面文本极易被 AI 代劳，那么柏拉图推崇的口头表达便成为了最佳解药。 在玛丽华盛顿大学，全校推行“高强度口头表达”（Speaking Intensive）课程要求已有约 30 年历史，并设有专门的演讲中心（Speaking Center）和开源教师学习社区，帮助生物学、商学、化学、哲学等各学科教师在课程中融入演讲与口头互动。 拉奥指出，虽然带有即时问答（Q&A）环节的课堂展示（Presentation）已能较好检验学生对内容的真实掌握，但**课堂辩论（In-class Debate）是更为卓越的评估形式**： - **双向透彻理解的刚性要求**：辩手必须时刻准备接招拆招并捍卫自身立场，哪怕是一场简短的辩论，也强制要求学生必须同时深度理解议题的“正反双方”，彻底杜绝了依靠 AI 机械复述单一观点的可能； - **综合素养的高维展现**：辩论不仅考核内容掌握，更在高压互动中全面激活并考察批判性思维、信息检索、团队协作以及即时领导力； - **历久弥新的核心能力（Durable Skills）**：辩论所锻造的沟通与认知抗辩能力，正是数千年来乃至近几十年职场雇主始终最为看重的持久技能。 ### [08:59](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=539s) 辩论与人文学科在 AI 时代的核心价值 每一项针对雇主的调查都在询问他们看重候选人的哪些素质，而雇主们列出的核心能力惊人地一致——辩论训练在培养和展现这些能力方面，比几乎任何其他活动都更为有效。近期《安全管理》（Security Management）杂志发表了一篇题为《AI 时代的辩论之道》的文章，明确指出：“辩论技能数世纪以来一直是社会进步的基石，受过辩论训练的专业人士往往表现卓越；通过磨砺这些技能，组织机构能够最大限度地释放大语言模型与人工智能的潜力。” ```mermaid flowchart TD accTitle: 人文学科与辩论赋能 AI 的闭环逻辑 subgraph S1["人文学科与辩论核心素养"] A["批判性思维与同理心"] B["对话、辩论与多重视角"] end subgraph S2["职场与技术应用"] C["雇主核心需求与文科复兴"] D["最大化释放 LLM 与 AI 潜力"] end subgraph S3["AI 发展新范式"] E["破除单一垄断型 AI (Monolithic AI)"] F["构建 AI 多元主义 (AI Pluralism)"] end S1 --> S2 S2 --> S3 ``` 这正是通识人文教育（Liberal Arts Education）致力于帮助学生塑造的核心素养。《Inc.》杂志在 7 月份也报道了“文科学位的复兴”，指出当前市场对具备此类技能的学生需求日益高涨。甚至科技行业的领军人物也在公开强调，未来真正能够让学生充分用好 AI 工具的，恰恰是这些人文学科的底层素养。 人文学科的核心在于通过历史、语言、艺术、哲学和宗教的研究，深入探索人类境况、文化与价值观。没有人文学科能够脱离“辩论、对话与同理心”而独立存在。人文学科本质上是一种多元主义传统，它正视并拥抱人类经验与表达的内在多样性，容纳多元视角、不同方法论以及多样化的阐释。面对生成式 AI 带来的冲击，文科与人文学科不仅能够积极适应挑战，更拥有成熟的思想模型来引领未来 AI 走向更负责任的研发道路。 --- ### [10:56](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=656s) 单一垄断型 AI 的困境与 AI 多元主义的提出 当前 AI 发展面临着一个鲜少被公开讨论的严峻挑战——“单一垄断型 AI”（Monolithic AI）。大多数科技公司目前的开发逻辑都是试图打造“统治一切的唯一模型”（the one model to rule them all），全力争夺单一模型的绝对霸权，而非促成不同模型之间的共存与互动。 这种追求“单一最优模型”的研发范式，加上多项工程实践，导致了 AI 模型及其部署的高度同质化： - **训练数据偏差**：用于训练大语言模型的数据主要来源于西方、受过大学教育、使用英语的人群，这使得模型底座的数据底色极度均质化。 - **对齐目标局限**：奖励模型通常仅针对“广泛可接受性”（broad acceptability）进行优化，其结果是模型试图迎合所有人，却抹平了特定个体或细分群体的独特观点与差异化诉求。 作为对抗单一垄断的替代方案，“AI 多元主义”（AI Pluralism）应运而生。人文学科为 AI 系统表达合理的多样性（而非陷入相对主义）提供了蓝图。AI 多元主义所追求的绝非仅仅是“形式上的品种繁多”，而是不同视角之间积极的对话、交锋与互动。 --- ### [12:26](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=746s) 多智能体辩论系统与修辞论证范式 为了将 AI 多元主义落地，研究团队开展了一系列实践探索。团队成员创立了 DebHub 平台，利用美国国家科学基金会（NSF）的资助，为城市辩论联盟的学生提供支持工具。研究团队成员（包括 John Hines 与 Devon Gonier）持续开发科研工具并发表成果，以支撑 AI 多元主义模型的构建。 计算机科学家兼前辩论学者 Devon Gonier 在一年半前搭建了一套**多智能体辩论系统（Multi-Agent Debate System）**： - **角色分配**：实例化多个 AI 模型，分别训练并设定为辩论中的正方辩友、反方辩友以及评审裁判（Judge AI）。 - **全流程对抗**：AI 智能体自主完成资料检索、论点构建与辩论策略设计，完整进行立论演讲、交叉质询（Cross-examination）及总结驳论。 - **裁判评估与信念修正**：裁判 AI 对整场辩论进行最终裁决。系统进一步引入了“信念修正”（Belief Revision）与长期记忆机制，使 AI 不仅能代入特定视角并评估其他模型的论点，还能将辩论中形成的认知沉淀应用到未来的决策中。 ```mermaid flowchart LR accTitle: 多智能体辩论与对齐系统架构 subgraph Debate["多智能体辩论交互"] direction TB Aff["正方 AI 智能体<br/>(Affirmative)"] Neg["反方 AI 智能体<br/>(Negative)"] Aff <-->|"演讲 / 质询 / 驳论"| Neg end subgraph Evaluation["评审与沉淀"] direction TB Judge["裁判 AI (Judge)"] Memory["信念修正与记忆<br/>(Belief Revision)"] Judge --> Memory end Debate -->|"提交辩论记录"| Judge ``` 在理论层面，研究团队提出引入“修辞论证”（Rhetorical Argumentation）来补充目前 AI 学界普遍依赖的“计算论证”（Computational Argumentation）。计算论证往往假定或导向单一垄断视角；而修辞论证的核心在于理解不同立场之间的差异，并学会针对这些差异进行策略性调适，这正是构建多智能体高效沟通的核心。 在具体部署形态上，AI 多元主义涵盖多种模型架构： - **奥弗顿多元模型（Overton Pluralistic Model）**：在合理的奥弗顿窗口光谱内提供多维度的回应。 - **可操纵多元模型（Steerable Pluralistic Model）**：可根据特定的价值观或文化视角进行条件化引导。 - **分布多元模型（Distributional Pluralistic Model）**：忠实反映某一特定人群内部的真实意见分布概率。 --- ### [14:54](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=894s) AI 对齐中的可解释性与可争议性及未来展望 AI 多元主义最具前景的应用领域之一是 **AI 对齐（AI Alignment）**。当前 AI 模型如同“黑盒子”，开发者清楚训练方法与架构设计，却难以解释模型产生特定输出的具体内部机理。基于辩论机制的 AI 多元主义为对齐问题提供了两大核心支柱： | 对齐机制 | 核心内涵与辩论赋能 | |:--- |:--- | | **可解释性 (Explainability)** | 在多元辩论模型中，AI 智能体必须互相沟通、阐明自身立场并展开交锋。这迫使模型展示其论点演进与决策推导过程，带来更透明的内部机理。 | | **可争议性 (Contestability)** | 理解决策过程只是第一步，更重要的是赋予人类“质疑与挑战决策”的权力。若决策偏离了所服务社区的伦理价值，辩论系统天然内置的协商机制可直接支持对模型输出提出异议。 | 关于后续的研究推进与学术产出： - **学术出版**：2023 年春季曾联合多位教育学者编著出版了聚焦 AI 时代“真实性评估与认证”（Authentication）的专著；最新联合主编的著作《AI 多元主义：在人工智能中促进多样性与对话》（*AI Pluralism: Fostering Diversity and Dialogue in Artificial Intelligence*）将于今年秋季正式面世。 - **玛丽华盛顿大学“AI 与通识文科中心”**：该中心已于本周正式宣布成立，后续将秉持开源精神，向公众开放大量研究资源与工具包。 - **发展规划**：近期目标是为师生提供辩论辅助与教学评估工具，中期持续展示多智能体系统在多元主义实践中的效用，远期目标则是将多元主义推理范式彻底嵌入到未来通用 AI 系统的底层架构之中。历史反复证明，面对新兴技术的恐惧往往会让人错失真正的时代机遇。 ### [17:49](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=1069s) 人文学科对 AI 多元主义的引领与总结致谢 演讲在对人文学科与人工智能未来关系的深刻总结中落下帷幕。讲者指出，人文学科的核心价值在于教会人们珍视“多元主义”（Pluralism）。将人文学科的多元视角引入 AI 开发，能够为系统提供兼顾并表达人类多元价值观的技术模型。在此之中，“辩论”（Debate）不仅是一种教学形式，更是一座关键的桥梁，它连接了课堂中的真实性学习（Authentic Learning）与技术层面更高效、更具可解释性的 AI 架构。 实现这一转变的第一步，是重新审视 AI 的本质——它绝非单纯的技术现象，而是一个深层的“文化与伦理项目”（Cultural and Ethical Project）。讲者坚信，面对人工智能浪潮，人文学科绝不应仅仅被动适应，而是完全有能力、也应当主动站出来，引领 AI 技术走向更加健康、高效且富有成效的发展道路。在演讲最后，讲者分享了自己的联系方式，欢迎现场观众通过 LinkedIn 或电子邮件进行后续交流探讨，并在会后留下来答疑，同时对参与本次教育专题分论坛的与会者表达了诚挚的谢意。 ```mermaid flowchart TD accTitle: 人文学科与AI多元主义结合路径 subgraph CoreValues["人文学科核心价值"] A["人文学科的多元主义 (Pluralism)"] --> B["文化与伦理维度重塑"] end subgraph Bridge["连接机制"] C["辩论机制 (Debate)"] end subgraph Outcomes["未来赋能方向"] D["教学端：真实性学习与评估"] E["技术端：更具可解释性与对齐的 AI"] end CoreValues --> C C --> D C --> E ``` - **重塑认知**：必须将 AI 视为一项文化与伦理工程，而非单纯的算法与算力演进。 - **人文学科的角色定位**：从“被动应对技术冲击”转向“主动引领技术开发方向”，使 AI 能够更好地代表人类多元价值。 - **互动与交流**：讲者鼓励大家通过 LinkedIn 及邮件保持沟通，并在教育专题会场后继续深入讨论。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=4s) 看似无所不知，实则一无所知 — 阿南德·拉奥（引述苏格拉底） > [00:04](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=4s) 任何在你出生时就已存在的技术，都是自然而平常的；任何在你 15 到 35 岁之间发明出来的技术，都是令人振奋的革命性突破，你甚至能以此开创事业；任何在你 35 岁之后发明出来的技术，都违背了自然的正常法则。 — 阿南德·拉奥（引述道格拉斯·亚当斯） > [05:05](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=305s) 在课堂上设立所谓“纯净无 AI 气泡”不仅技术上无法实现，更是对学生未来的不负责任。 — 阿南德·拉奥 > [08:59](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=539s) 辩论技能数世纪以来一直是社会进步的基石，受过辩论训练的专业人士往往表现卓越；通过磨砺这些技能，组织机构能够最大限度地释放大语言模型与人工智能的潜力。 — 《安全管理》（Security Management） > [10:56](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=656s) 统治一切的唯一模型 — 阿南德·拉奥 > [10:56](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=656s) AI 多元主义所追求的绝非仅仅是“形式上的品种繁多”，而是不同视角之间积极的对话、交锋与互动。 — 阿南德·拉奥 > [14:54](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=894s) 理解决策过程只是第一步，更重要的是赋予人类“质疑与挑战决策”的权力。 — 阿南德·拉奥 > [17:49](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=1069s) AI 绝非单纯的技术现象，而是一个深层的“文化与伦理项目”。 — 阿南德·拉奥 > [17:49](https://www.youtube.com/watch?v=QWWnNU-QYvY&t=1069s) 人文学科绝不应仅仅被动适应，而是完全有能力、也应当主动站出来，引领 AI 技术走向更加健康、高效且富有成效的发展道路。 — 阿南德·拉奥 ## Mentioned 阿南德·拉奥 (Anand Rao), 玛丽华盛顿大学 (University of Mary Washington), 文科人工智能中心 (Center for AI in the Liberal Arts), 苏格拉底 (Socrates), 柏拉图 (Plato), 《斐德罗篇》 (Phaedrus), 约翰内斯·特里特米乌斯 (Johannes Trithemius), 《赞抄写员》 (In Praise of Scribes), 道格拉斯·亚当斯 (Douglas Adams), Google Gemini, ChatGPT, Claude, Perplexity, 《安全管理》 (Security Management), 《Inc.》杂志, DebHub, 美国国家科学基金会 (NSF), John Hines, Devon Gonier, 《AI 多元主义：在人工智能中促进多样性与对话》 (AI Pluralism: Fostering Diversity and Dialogue in Artificial Intelligence) ## Video description Anand Rao, Director of the Center for AI and the Liberal Arts and Professor of Communication and Digital Studies, University of Mary Washington Abstracts What if debate could simultaneously transform our classrooms and our AI systems—creating both authentic assessment methods for today's educators and models for more transparent, aligned artificial intelligence in the future? I will explore AI pluralism as a framework for integrating artificial intelligence with humanities education while preserving core humanistic values of critical thinking, ethical reasoning, and diverse viewpoints. I will demonstrate how rhetorical argumentation and debate among AI agents can create more transparent, accountable systems that enhance rather than diminish learning environments. Drawing from my research on AI explainability and contestability, I will showcase how debate emerges as one of the most effective approaches for authenticating student learning in an era where traditional assessments are increasingly vulnerable to AI-generated content. Attendees will learn practical techniques—such as implementing AI-facilitated debate activities—that humanities educators can immediately apply to foster authentic assessment and critical thinking skills. Join this presentation to discover how to bridge technological innovation with humanistic inquiry in ways that create more inclusive, ethically-grounded educational experiences. [Source](https://www.youtube.com/watch?v=QWWnNU-QYvY)

Source quotes

[0:04] 阿南德·拉奥（引述苏格拉底）: 看似无所不知，实则一无所知 [0:04] 阿南德·拉奥（引述道格拉斯·亚当斯）: 任何在你出生时就已存在的技术，都是自然而平常的；任何在你 15 到 35 岁之间发明出来的技术，都是令人振奋的革命性突破，你甚至能以此开创事业；任何在你 35 岁之后发明出来的技术，都违背了自然的正常法则。 [5:05] 阿南德·拉奥: 在课堂上设立所谓“纯净无 AI 气泡”不仅技术上无法实现，更是对学生未来的不负责任。 [8:59] 《安全管理》（Security Management）: 辩论技能数世纪以来一直是社会进步的基石，受过辩论训练的专业人士往往表现卓越；通过磨砺这些技能，组织机构能够最大限度地释放大语言模型与人工智能的潜力。 [10:56] 阿南德·拉奥: 统治一切的唯一模型 [10:56] 阿南德·拉奥: AI 多元主义所追求的绝非仅仅是“形式上的品种繁多”，而是不同视角之间积极的对话、交锋与互动。 [14:54] 阿南德·拉奥: 理解决策过程只是第一步，更重要的是赋予人类“质疑与挑战决策”的权力。 [17:49] 阿南德·拉奥: AI 绝非单纯的技术现象，而是一个深层的“文化与伦理项目”。 [17:49] 阿南德·拉奥: 人文学科绝不应仅仅被动适应，而是完全有能力、也应当主动站出来，引领 AI 技术走向更加健康、高效且富有成效的发展道路。

Source quotes

[0:04] Anand Rao: not with wisdom but with a conceit of wisdom [0:04] Johannes Trithemius: He who ceases from zeal for writing because of printing is no true lover of the scriptures. [0:04] Douglas Adams: against the natural order of things. [5:00] Anand Rao: AI-free zones [10:36] Anand Rao: one model to rule them all [12:08] Anand Rao: AI Pluralism [15:01] Anand Rao: black boxes [17:45] Anand Rao: cultural and ethical project

Original transcript

[0:04] Well, good morning everybody. My name is [0:06] Anandra. I'm a professor of [0:07] communication and digital studies at the [0:10] University of Mary Washington where I'm [0:12] also director of our new center for AI [0:14] in the liberal arts. Um, I'm finishing a [0:17] year-long research sobbatical and I'm [0:18] excited to share with you some of the [0:20] results of the work that I've been doing [0:22] as I prepare to launch this center and [0:24] to explore some of the ways that we can [0:25] adapt in higher education to the age of [0:28] AI. Now you when faced with a new [0:31] educational technology there was a very [0:33] well-known and influential scholar who [0:35] said by telling them of many things [0:37] without teaching them you'll make them [0:39] seem to know well for the most part they [0:42] know nothing [0:44] and as men filled not with wisdom but [0:46] with a conceit of not wisdom they will [0:48] be a burden now this of course was [0:51] Socrates and Plato's faded the [0:53] technology he feared was writing and he [0:56] he was arguing for the oral tradition [0:59] Now, Socrates was worried that students [1:01] would become forgetful that they would [1:03] mistake information for wisdom and they [1:05] would lose the discipline of true [1:07] intellectual engagement. Now, if we fast [1:10] forward to late 15th century, a German [1:13] abbott named Johannes Trmus issued very [1:16] similar warnings about the printing [1:17] press when he said, "Only time will [1:19] tell. He who ceases from zeal for [1:22] writing um for writing because of [1:25] printing is no true lover of the [1:27] scriptures. Now Trius argued that [1:30] printed books would never be the [1:31] equivalent of handwritten texts. He also [1:33] said that monks would lose the spiritual [1:36] discipline that came by the painstaking [1:38] copying and handwriting of the [1:40] scriptures. Now both Plato and Trius [1:44] were wrestling with the same fundamental [1:45] question that we as educators are faced [1:47] with AI today. when transformative [1:50] technology arrives, how do we preserve [1:53] what's most valuable about human [1:55] learning? Now, it also isn't lost on me [1:58] the irony that the reason we know about [2:00] Plato's concerns is because they were [2:01] written down and Trius's concerns about [2:04] the printing press were of course [2:06] printed in a book um and subsequent [2:08] books after that. These examples also [2:10] remind me of one of my favorite quotes [2:12] from Douglas Adams who talked about how [2:13] we view technology. And part of this he [2:16] says anything that is in the world when [2:18] you're born is normal and ordinary. [2:20] Anything that's invented when you're [2:21] between 15 and 35 is new, revolutionary, [2:25] and you'll probably make a career of it. [2:27] And anything invented after you're 35 is [2:30] against the natural order of things. And [2:32] I think that's where most of my [2:33] colleagues and I probably fall into that [2:35] age group. And that probably explains [2:36] some of the resistance that we found in [2:38] higher ed. Now, both Plato and Trius [2:41] feared the technology would replace [2:43] human effort with convenience. Sound [2:45] familiar, right? They were each [2:47] partially right about the challenges, [2:49] but they were ultimately wrong about the [2:50] outcomes. Now, I'm not so quick to to [2:54] quickly dismiss the challenges faced by [2:56] generative AI um and to suggest that [2:58] it's the same as the fight of orality [3:01] over writing or the concerns about the [3:03] printing press over writing. Um, I do [3:05] think the generative AI presents a [3:07] unique challenge to our educational [3:09] model. Um, but I do have a hope for a [3:12] way forward. And for me, that hope comes [3:15] in the form of debate. Debate's the [3:17] reason I'm in academia. I was a debater [3:19] in high school and in college. Spent the [3:20] first 10 years of my career as a debate [3:22] coach. And uh, ever since I've been [3:24] coaching my kids high school debate [3:26] teams, middle school and high tools, [3:27] high school debate teams for a number of [3:29] years. And what I found is that the in [3:32] the the ability to incorporate new [3:33] educational technologies into debate has [3:36] allowed students in debate given their [3:38] competitive drive to do even more [3:40] advanced and more astonishing things as [3:43] they're still learning the content and [3:45] as they're able to work through it. And [3:46] one of the best examples that I [3:48] experienced was just this past year [3:49] working with some students, some novice [3:51] debaters on my daughter's high school [3:53] debate team. They were brand new to [3:54] debate. They weren't typical students. [3:56] They're very highly motivated, very [3:58] competitive young women, but they were [4:00] getting ready for their first [4:01] tournament. They only had a couple of [4:02] weeks to get ready for this this topic [4:04] and they didn't know anything about it. [4:05] But they were able to use a number of AI [4:07] tools from Gemini and ChatBT and Claude [4:10] to do the research to prepare briefs to [4:12] brainstorm about responses and [4:14] developing argumentative strategies. [4:16] They didn't outsource the work all in [4:19] terms of their cognitive abilities. They [4:21] were grappling with all of these issues [4:23] and learning very quickly about that [4:25] topic and how to debate, but they were [4:27] able to jump very fast up that learning [4:28] curve. When they got to that first [4:30] tournament, they made it to [4:31] quarterfinals against many debaters that [4:33] were far more experienced, had more [4:35] years of experience, um, and more wins [4:37] under their belt. But this team was able [4:39] to use AI effectively to prepare. They [4:41] didn't use AI in the debates. They used [4:43] it to make sure they were ready for the [4:44] debates. And when they came out of a [4:46] debate and they're like, "Anan, I don't [4:47] know this one argument. We got to look [4:48] it and they would go and find perplexity [4:50] and they would look up the argument, [4:51] figure out some ways they could develop [4:52] answers to with they were adopting and [4:55] deploying these tools in ways that I [4:57] want all of our students to be able to [4:58] deploy these AI tools in our classrooms. [5:00] And I think it really gives me a lot of [5:01] hope. Now, jumping back of close to [5:05] three years now, when we think about the [5:07] end of the fall semester in 2022, [5:09] I like many other educators were really [5:11] knocked back on our heels with the [5:13] introduction of Chat GPT. I've been [5:15] teaching and working within our digital [5:16] studies program to talk about AI, but it [5:19] was still astonishing to me to see the [5:21] capabilities of chat GPT and the way our [5:23] students were able to take to chat GPT. [5:26] And you know, it's very understandable [5:28] that at first the reaction for many is [5:31] we have to keep this out of the [5:32] classroom. And one knee-jerk reaction is [5:34] we need to find AI detectors. I think we [5:36] all know the detectors don't work. [5:38] They're only probabilistic at best. [5:40] They're not reliable. They have a lot of [5:41] false negatives. And more concerningly [5:44] to me is that they have a lot of false [5:46] positives. They tend to identify writing [5:48] as generated by AI from students that [5:50] don't have the same foundational writing [5:52] skills or students for whom English is a [5:55] second language. Some of the most [5:56] vulnerable students we have are [5:58] identified as using AI for their work [6:00] when they really didn't. So we we can't [6:02] rely on that. There's also a concern [6:04] that the pace of technological change is [6:07] really outrunning our ability to adapt [6:10] in higher ed. We're known for slow, [6:12] methodical uh reflection and eventual [6:15] adaptation. Um we need to move a little [6:17] bit faster when we're when we're dealing [6:19] with something like generative AI. Now, [6:21] it was natural for many educators to [6:23] think that maybe what we could do is [6:24] create AI free zones in our classrooms, [6:26] create a bubble. That's just not [6:28] realistic. Uh there's no way that you [6:30] can keep all of the technologies out. [6:31] And I think more importantly, it's [6:33] unfair to our students. Our job is to [6:35] make sure we're preparing them to [6:37] develop their cognitive abilities to [6:39] work with them intellectually to prepare [6:41] them for a life well-lived. And they [6:43] aren't going to be able to do that if we [6:45] pretend that there's a whole sector of [6:46] society that doesn't exist. We need to [6:48] make sure that we're incorporating it in [6:50] responsible ways. And I think there are [6:51] ways we can go about doing that. [6:54] Now, the solution for me was to lean [6:56] into my communication background, of [6:58] course. How can we authenticate student [7:00] work and better assess their learning? [7:03] Well, Plato would love it. I think it's [7:04] the return to the oral tradition, right? [7:06] It's public speaking. I spent my first [7:08] 25 years in academia coaching speech and [7:10] debate and then also working with [7:12] colleagues through our speaking [7:14] intensive program to help them integrate [7:16] oral communication activities into their [7:18] biology courses, business courses, [7:20] chemistry, um, philosophy. Um, they were [7:23] integrating these because it was a great [7:25] way for students to develop those skills [7:27] as well as to better assess their true [7:29] understanding of of some of that [7:31] content. Now, a student presentation, [7:34] preferably with a robust Q&A, can really [7:37] demonstrate their understanding of the [7:38] content. And I think that's something [7:40] that many of us should be exploring if [7:42] we haven't already to figure out how we [7:43] can use that for assessment. But really, [7:46] even better is an in-class debate. [7:48] Successful participation in a debate can [7:50] demonstrate mastery of content better [7:53] than most any other form of assessment. [7:55] um to be ready to grapple with those [7:56] arguments and be able to defend your [7:58] position well means that you understand [8:00] both sides of that issue at the very [8:02] least even in a short superficial [8:04] debate. Um it also allows students to [8:06] develop and exhibit those skills that [8:08] are really foundational communication [8:10] skills, critical thinking, research, [8:12] collaboration, even leadership skills as [8:15] they're working together. Now my [8:17] university has a long history of [8:19] supporting oral communication [8:21] activities. We've had this speaking [8:22] intensive requirement for I think about [8:25] 30 years. We have a speaking center that [8:28] I directed for a number of years as well [8:30] that's provides support to students and [8:31] to faculty. Um I've also run a lot of [8:34] faculty development, professional [8:35] development opportunities including you [8:37] know learning communities where we have [8:39] open- source materials that uh faculty [8:42] have put together resources, [8:44] assignments, rubrics and so forth for a [8:46] lot of the in-class work that they've [8:48] been doing. This is even pre-generative [8:49] AI. [8:52] What we're really talking about are the [8:53] durable skills that employers have been [8:55] seeking out for really millennia and [8:58] certainly for the last several decades. [8:59] Every survey of employers ask them what [9:02] they're looking for and they talk about [9:03] many of these same skills. And debate [9:05] does more to develop and exhibit those [9:07] skills than really any other activity. [9:09] Um yesterday there was a publication [9:11] security management had an article that [9:13] talked about the case for debate in the [9:15] age of AI. Here's a snippet of it. They [9:17] say debate skills have been foundational [9:20] to societal progress for centuries. [9:22] Debat trained professionals excel. By [9:25] honing these skills, organizations can [9:27] maximize the potential of large language [9:29] models and AI. [9:32] These are the skills that liberal arts [9:33] educations help student develop. And [9:36] this is why we see more and more stories [9:37] like this one from Inc. magazine in July [9:40] talking about kind of the rebirth of the [9:42] liberal arts degree and the the desire [9:44] for students to develop those skills. [9:46] We're also hearing this even from tech [9:48] leaders that the important skills are [9:50] those skill humanistic skills that are [9:52] going to allow students to make the most [9:53] of AI tools in the future. [9:56] But these are also the skills that are [9:58] at the core of the humanities. These are [9:59] the disciplines that explore the human [10:01] condition, culture and values through [10:04] the study of history, language, arts, [10:06] philosophy, religion. Um really I would [10:09] argue that you can't have the humanities [10:11] without debate, dialogue and empathy. [10:13] These are important skills within our [10:15] own intellectual work as well as for the [10:17] foundational skills the students are [10:19] going to need as they leave university. [10:21] Humanities really are a pluralistic [10:23] tradition. They embrace multiple [10:25] perspectives, methodologies and inherent [10:27] and different interpretations. And they [10:29] do that by recognizing the inherent [10:31] diversity of human experiences and [10:34] expression. [10:36] What I've realized over the last two and [10:37] a half years or so is that not only can [10:40] the liberal arts and the humanities [10:41] adapt to the challenges that are [10:43] presented by generative AI, but I think [10:45] more importantly, I would make the case [10:47] that the humanities and the liberal arts [10:49] can develop a model and have the model [10:51] to help lead AI development in more [10:53] responsible ways going forward. Let me [10:56] first start by talking about one [10:58] challenge that we're facing with AI that [10:59] I don't know that is really talked about [11:01] very often and that is monolithic AI. [11:04] the sense that what is being developed [11:06] and really the models for development by [11:07] most companies is to develop the one [11:10] model to rule them all. It's the the [11:12] really racing towards supremacy of one [11:15] model at the expense of another not [11:17] really for them to coexist or to be able [11:18] to engage with one another. um there's [11:20] an emphasis on building that one best [11:22] model model and in addition to that [11:24] there are a number of practices that we [11:26] find in in our research that really [11:28] result in kind of the homogeneity of AI [11:31] models and the deployment of AI [11:33] including things like training data [11:34] biases you know the data trained that's [11:37] used to train large language models [11:38] tends to come from western sources [11:41] college educated typically in English [11:43] and that tends to homogenize a lot of [11:45] the data that the models are then [11:46] working from there are also concerns [11:48] about alignment objectives ives reward [11:50] models are frequently optimized for [11:53] broad acceptability. And when you do [11:55] that, that just means that you make sure [11:56] that they are appealing to everybody [11:58] without really being able to address any [12:01] individual or subcategories of [12:03] perspectives that we might have. And [12:05] those are real concerns. Now, the [12:08] alternative to this is AI pluralism. U [12:11] the humanities offer a blueprint for [12:13] representing legitimate diversity and [12:14] not relativism in AI systems. And this [12:18] isn't really just about diversity as [12:19] mere variety, but really pluralism [12:22] offering active engagement and dialogue. [12:26] Let me say a little bit about some of [12:27] the work that we've been doing in this. [12:28] You know, two colleagues of mine founded [12:30] DebbHub. This is a platform that is used [12:32] to develop AI tools to support students [12:35] and debaters. Uh they have an NSF grant [12:37] that they've been using with urban [12:38] debate leagues to be able to provide [12:40] students with some of these tools. Um [12:42] and a group of us including um the [12:44] founders at the top um John Hines and [12:46] Devon Gonier are part of a research [12:48] group and we've been developing a number [12:50] of research tools uh as well as [12:52] publications to be able to support the [12:53] development of this AI pluralism model. [12:56] Another component of this is something [12:58] that Devon produced. He's a computer [13:00] scientist and happens to have a [13:01] background in debate. And a little over [13:03] a year and a half ago, he put together a [13:05] multi- aent debate system where he was [13:07] able to take multiple instances of AI [13:10] models, train them as particular members [13:12] or participants within a debate. So [13:14] there were affirmative side AI models, [13:17] there were negative side. He also [13:18] trained models to be able to serve as [13:20] the judge within the debate. Um the [13:23] models did the research, they developed [13:24] the arguments, um had argumentative [13:26] strategies for the debate and then went [13:28] back and forth from full speeches, [13:30] cross-examination, rebuttal speeches, [13:32] and then the judge AI was able to [13:34] evaluate the debate at the end. He's [13:36] also building this out so that you could [13:38] have kind of memory built in, belief [13:41] revision as he's referring to it, so [13:42] that you have the capability of AI [13:44] models, not only assuming different [13:46] perspectives and evaluating the [13:48] perspectives of other AI models, but [13:50] being able to carry that forward with [13:51] other decision-m in the future, which I [13:53] think is a really important component as [13:55] we move forward. [13:57] Now, another element of what we're [13:58] proposing within AI pluralism is to [14:00] incorporate rhetorical argumentation to [14:02] complement the computational [14:04] argumentation that most AI researchers [14:06] really rely upon. Computational [14:08] argumentation is more likely to presume [14:11] or allow for a monolithic AI [14:13] perspective. Whereas a rhetorical [14:15] perspective is really one about [14:17] differences of perspectives and learning [14:19] how to adapt to those differences and [14:21] perspectives. something that'll be [14:22] really core component when we build out [14:24] an AI pluralism model for those agents [14:26] to be able to work effectively to appeal [14:29] to one another. [14:31] Now there are a lot of v a v variety of [14:33] models that could be deployed within a [14:35] pluralism in an AI system. Um including [14:37] something like an overtism overton [14:39] pluralistic model which provides a [14:41] spectrum of reasonable requests or [14:42] responses. Steerable pluralistic models [14:45] can be conditioned on specific values or [14:47] perspectives or even a distributional [14:49] pluralistic model that reflects the [14:51] distribution of opinions within a [14:53] population. [14:54] Now one of the most promising [14:56] applications for AI pluralism is in the [14:58] field of AI alignment. And this is [15:00] something that a lot of researchers of [15:01] course um very importantly are doing [15:04] work in this area. It's the the concern [15:06] about how do we make sure that as AI [15:08] develops and it's um working in in [15:10] further capabilities that it's able to [15:12] align with human values and not work [15:14] counter to human values. This is proving [15:16] to be very difficult already as the AI [15:18] models are typically kind of like black [15:20] boxes. They know how they train them. [15:22] They know how they set it up. but they [15:23] aren't really sure what happens inside [15:25] as they get their output and they they [15:27] try to figure out how they were able to [15:28] get that specific output. Now, the use [15:31] of a debate centered model of AI [15:32] pluralism can provide two important [15:34] mechanisms for alignment. Um these [15:36] include explanability where u because of [15:39] the lack of transparency. If you have a [15:41] debate model sense of of pluralism, the [15:44] models then have to be able to [15:45] communicate with each other and to share [15:47] their perspectives. That means they're [15:48] offering greater transparency the [15:50] arguments and the development of the [15:52] decision-making that they're actually [15:53] going through. And then secondly, [15:55] contestability. It's one thing to [15:56] understand how they came up with the [15:58] decision that they had, but it's quite [16:00] another to make sure that you can [16:01] actually challenge that decision if [16:02] there's something that you think isn't [16:04] aligned with the values of the community [16:06] maybe that you're serving. um these are [16:08] really hardwired within a pluralistic [16:10] system because if you have AI models [16:12] that aren't trying to be monolithic but [16:14] trying to work together to be able to [16:16] negotiate have dialogue and debate about [16:19] applications and decision-m then they [16:22] naturally have to include some [16:24] transparency and also some [16:25] contestability within that model and [16:27] that system. So where do we go from here [16:29] and what what are the next steps? Well [16:31] the first is I wanted to point to a [16:33] couple of publications that I think are [16:35] are useful. The first one was uh from a [16:37] little while ago. It was from the spring [16:38] of 2023 really focusing on the [16:40] authentication issue. I co-edited and [16:43] contributed this volume as well of a [16:44] number of other educators to think about [16:46] how do we address that assessment [16:47] concern. The second is coming out this [16:49] fall and this is a volume that I've [16:51] co-edited on AI pluralism titled AI [16:54] pluralism fostering diversity and [16:56] dialogue and artificial intelligence. [17:00] Now, the other aspect of this is that [17:02] I'll be working to continue this work [17:03] with my university's new center for AI [17:05] and the liberal arts. As I mentioned, um [17:07] we've just announced it officially this [17:09] week. Uh and we already have a slate of [17:11] different activities and events. Um the [17:13] goal is to make sure that we're out [17:14] public open sourcing um and making [17:16] available a lot of the resources and the [17:18] research so that others can be involved [17:20] in the discussion of not only how do we [17:22] help our students adapt, but also how we [17:24] can inform the development of current [17:25] and future AI models going forward. I [17:28] think in the near term it's about [17:29] providing debate facilitation tools for [17:31] students and faculty. We also want to [17:33] make sure that we continue to [17:34] demonstrate the use of multi- aent [17:36] systems to be able to provide those [17:37] models for pluralism and the long term [17:40] to hopefully embed some pluralistic [17:41] reasoning into AI systems going forward. [17:45] History shows that fears about new [17:47] technologies often miss the opportunity [17:49] for renewal. I think the humanities [17:52] teach us to value pluralism. Applying [17:54] this model to AI development offers [17:56] technical models to represent diverse [17:58] human values. Debate can serve as a [18:01] bridge between authentic learning and [18:03] more effective and explainable AI. The [18:06] first step is to rethink AI not just as [18:09] a technical phenomenon but also as a [18:11] cultural and ethical project. And in [18:14] this way the humanities I would argue [18:16] will not just adapt but can also help [18:18] lead AI development in more productive [18:20] ways going forward. My contact [18:22] information is available here. I've [18:24] encouraged you to reach out. I'm on [18:25] LinkedIn of course uh and always happy [18:27] to answer an email. I will be around [18:29] afterwards to answer any questions you [18:30] might have. Um thank you for being part [18:32] of this discussion uh and being here for [18:34] the education track uh as well as for [18:36] attending this session. I appreciate it.