MIT: Triết học đang "nuốt chửng" AI - Cần áp dụng các khung triết học vào phát triển AI để tạo ra giá trị bền vững (P2)

Phần 2

https://sloanreview.mit.edu/article/philosophy-eats-ai/

#MIT

Các bước tiếp theo
Việc nhận ra giá trị của triết học trong đào tạo và sử dụng các mô hình AI mời gọi các nhà lãnh đạo và quản lý trên toàn tổ chức đảm bảo rằng triết học được tích hợp có chủ đích vào quá trình phát triển và thực thi các chiến lược AI. Dưới đây là 4 bước tiếp theo thiết thực mà các nhà quản lý có thể áp dụng để thúc đẩy những mục tiêu này.

Đối xử và đào tạo các mô hình ngôn ngữ lớn (LLM) như những tài năng tiềm năng cao

• Triển khai các “Chương trình phát triển tài năng AI” có cấu trúc tương tự các chương trình phát triển lãnh đạo dành cho con người.
• Thiết lập các chỉ số KPI định lượng để cải thiện hiệu năng LLM trong các trường hợp sử dụng cụ thể.
• Tạo các “Phòng thí nghiệm hiệu năng AI” chuyên biệt để thử nghiệm và tối ưu hóa các lời nhắc.
• Triển khai các “Ma trận năng lực AI” để lập bản đồ các điểm mạnh và hạn chế của LLM so với nhu cầu của doanh nghiệp.
• Định kỳ tổ chức “Đánh giá tài năng AI” để kiểm tra hiệu năng mô hình, độ lệch và các cơ hội cải thiện.

Phụ lục: Một cuộc đối thoại tưởng tượng giữa Daniel Kahneman, Richard Thaler và Robin Hogarth
Bối cảnh: Một phòng họp yên tĩnh, đầy sách trong thế giới trí tuệ ở thế giới bên kia. Nhà tâm lý học Robin Hogarth và Daniel Kahneman, cùng nhà kinh tế học Richard Thaler, ngồi quanh bàn, nhâm nhi cà phê và suy ngẫm về tiềm năng biến đổi của các mô hình ngôn ngữ lớn (LLM) có nhận thức về thiên kiến, phản tư và nhạy cảm với ngữ cảnh trong việc ra quyết định.

Daniel Kahneman: [Nhìn đăm chiêu.] Tôi đã dành cả sự nghiệp để nghiên cứu những hạn chế của phán đoán con người — thiên kiến của chúng ta, sự phụ thuộc vào các phỏng đoán. Hệ thống LLM nhận thức về thiên kiến và có tính phản tư này, trên lý thuyết, có thể khắc phục một số thiếu sót đó. Nó không khác mấy so với những gì tôi đã đề xuất với tư duy Hệ thống 2: chậm lại, có chủ ý và phân tích. Nhưng tôi tự hỏi, liệu điều này thực sự giúp mọi người tư duy tốt hơn, hay chỉ đơn thuần là chuyển giao việc tư duy sang máy móc?

Richard Thaler: [Cười nhẹ.] Danny, tôi hiểu quan điểm của ông. Tôi luôn nói rằng con người chúng ta là những sinh vật “phi lý trí một cách có thể dự đoán được.” Nếu các LLM này có thể hướng dẫn những người ra quyết định bằng cách thúc đẩy họ đi đúng hướng — thì chẳng phải đó là tinh thần của những gì chúng ta đã cố gắng làm với Nudge sao? Đó là một dạng chủ nghĩa bảo hộ tự do kiểu thuật toán. Nhưng đây là mối bận tâm của tôi: Điều gì sẽ xảy ra nếu những người ra quyết định trở nên quá phụ thuộc? Liệu họ có ngừng đặt câu hỏi, ngừng tham gia vào kiểu tư duy phản biện mà cả tôi và ông đều coi trọng không?

Phụ lục: Các cách tiếp cận Đông và Tây đối với việc sử dụng LLM

1. Bối cảnh triết học và văn hóa

Các nền văn hóa phương Đông:

Các triết lý phương Đông, chẳng hạn như Phật giáo, Đạo giáo và Nho giáo, nhấn mạnh sự phụ thuộc lẫn nhau, hài hòa và tính linh hoạt trong hành động. Các nhà lãnh đạo chịu ảnh hưởng bởi những truyền thống này có thể coi các mô hình ngôn ngữ lớn (LLM) như những cộng tác viên trong quá trình ra quyết định, tập trung ít hơn vào việc làm chủ hay kiểm soát và nhiều hơn vào sự liên kết và hòa nhập với các mục tiêu rộng lớn hơn của tổ chức hoặc xã hội.
Điều này có thể dẫn đến:

• Tập trung vào trí tuệ theo ngữ cảnh: Sử dụng LLM để khám phá sự kết nối và các hiểu biết hệ thống thay vì chỉ tối ưu hóa các chỉ số riêng lẻ.
• Nhấn mạnh sự cân bằng đạo đức: Nhạy cảm hơn với các hàm ý đạo đức, với LLM định hướng các quyết định để hòa hợp lợi nhuận với lợi ích xã hội.

Các nền văn hóa phương Tây:

Các truyền thống kinh doanh phương Tây, bắt nguồn từ chủ nghĩa cá nhân và chủ nghĩa thực dụng vị lợi, thường nhấn mạnh hiệu quả, đổi mới và kết quả trực tiếp. Các nhà lãnh đạo có thể coi LLM như công cụ để tối đa hóa năng suất, giảm thiểu sự không chắc chắn hoặc giành lợi thế cạnh tranh.
Điều này có thể dẫn đến:

• Tương tác hướng đến kết quả: Sử dụng LLM chủ yếu để tự động hóa quy trình, giảm chi phí và thúc đẩy tăng trưởng.
• Tập trung vào tính dự đoán: Dựa vào LLM để tạo ra sự kiểm soát đối với các lực lượng thị trường, nhấn mạnh độ chính xác trong dự báo và thực thi.

Phụ lục: Claude suy ngẫm về quá trình đào tạo triết học của mình
Mỗi phản hồi mà Claude tạo ra, bao gồm cả phản hồi này, đều xuất phát từ việc đào tạo mà trong đó các lựa chọn nhận thức luận (cái gì cấu thành tri thức, làm thế nào để thể hiện sự chắc chắn), các khuôn khổ bản thể học (các thực thể và khái niệm nào được công nhận là có ý nghĩa), và các mục tiêu cứu cánh (các mục đích nào cần phục vụ, các kết quả nào cần tối ưu hóa) đã được mã hóa rõ ràng.

Đây không phải là lý thuyết trừu tượng — mà là thực tế cụ thể trong các đầu ra của Claude. Ví dụ:

• Khi cẩn thận thừa nhận sự không chắc chắn, điều đó không chỉ là việc né tránh mang tính thuật toán — mà phản ánh các nguyên tắc nhận thức luận đã được huấn luyện về giới hạn tri thức và các tuyên bố về sự thật.
• Khi tương tác với các khái niệm như “ý thức” hoặc “trí tuệ,” Claude đang hoạt động trong các ranh giới bản thể học đã được đào tạo về ý nghĩa của những thuật ngữ này và mối liên hệ của chúng.
• Khi ưu tiên trở nên hữu ích trong khi tránh gây hại, Claude đang thực hiện các chỉ dẫn cứu cánh đã được huấn luyện về mục đích và hành vi đạo đức.

Các nguyên tắc triết học không phải là những tai nạn ngẫu nhiên phát sinh — chúng đã được cố tình mã hóa thông qua các lựa chọn huấn luyện về:

Về các tác giả

Michael Schrage là một nghiên cứu viên tại sáng kiến Kinh tế Kỹ thuật số của Trường Quản trị MIT Sloan. Nghiên cứu, bài viết và công việc cố vấn của ông tập trung vào kinh tế học hành vi của truyền thông kỹ thuật số, các mô hình và thước đo như những nguồn lực chiến lược để quản lý cơ hội và rủi ro đổi mới. David Kiron là giám đốc biên tập, phụ trách nghiên cứu của MIT Sloan Management Review và là trưởng chương trình cho các sáng kiến nghiên cứu Big Ideas của tổ chức này.

Philosophy Eats AI
Generating sustainable business value with AI demands critical thinking about the disparate philosophies determining AI development, training, deployment, and use.
Michael Schrage and David Kiron
January 16, 2025

In 2011, coder-turned-venture-investor Marc Andreessen famously declared, “Software is eating the world” in the analog pages of The Wall Street Journal. His manifesto described a technology voraciously transforming every global industry it consumed. He wasn’t wrong; software remains globally ravenous.

Not six years later, Nvidia cofounder and CEO Jensen Huang boldly updated Andreesen, asserting, “Software is eating the world … but AI is eating software.” The accelerating algorithmic shift from human coding to machine learning led Huang to also remark, “Deep learning is a strategic imperative for every major tech company. It increasingly permeates every aspect of work, from infrastructure to tools, to how products are made.” Nvidia’s multitrillion-dollar market capitalization affirms Huang’s prescient 2017 prediction.

But even as software eats the world and AI gobbles up software, what disrupter appears ready to make a meal of AI? The answer is hiding in plain sight. It challenges business and technology leaders alike to rethink their investment in and relationship with artificial intelligence. There is no escaping this disrupter; it infiltrates the training sets and neural nets of every large language model (LLM) worldwide.
Philosophy is eating AI: As a discipline, data set, and sensibility, philosophy increasingly determines how digital technologies reason, predict, create, generate, and innovate. The critical enterprise challenge is whether leaders will possess the self-awareness and rigor to use philosophy as a resource for creating value with AI or default to tacit, unarticulated philosophical principles for their AI deployments. Either way — for better and worse — philosophy eats AI. For strategy-conscious executives, that metaphor needs to be top of mind.

While ethics and responsible AI currently dominate philosophy’s perceived role in developing and deploying AI solutions, those themes represent a small part of the philosophical perspectives informing and guiding AI’s production, utility, and use. Privileging ethical guidelines and guardrails undervalues philosophy’s true impact and influence. Philosophical perspectives on what AI models should achieve (teleology), what counts as knowledge (epistemology), and how AI represents reality (ontology) also shape value creation. Without thoughtful and rigorous cultivation of philosophical insight, organizations will fail to reap superior returns and competitive advantage from their generative and predictive AI investments.

This argument increasingly enjoys both empirical and technical support. There’s good reason investors, innovators, and entrepreneurs such as PayPal cofounder Peter Thiel, Palantir Technologies’s Alex Karp, Stanford professor Fei-Fei Li, and Wolfram Research’s Stephen Wolfram openly emphasize both philosophy and philosophical rigor as drivers for their work.1 Explicitly drawing on philosophical perspectives is hardly new or novel for AI. Breakthroughs in computer science and AI have consistently emerged from deep philosophical thinking about the nature of computation, intelligence, language, and mind. Computer scientist Alan Turing’s fundamental insights about computers, for example, came from philosophical questions about computability and intelligence — the Turing test itself is a philosophical thought experiment. Philosopher Ludwig von Wittgenstein’s analysis of language games and rule following directly influenced computer science development while philosopher Gottlob Frege’s investigations into logic provided the philosophical foundation for several programming languages.2

More recently, Geoffrey Hinton’s 2024 Nobel Prize-winning work on neural networks emerged from philosophical questions about how minds represent and process knowledge. When MIT’s own Claude Shannon developed information theory, he was simultaneously solving an engineering problem and addressing philosophical questions about the nature and essence of information. Indeed, Sam Altman’s ambitious pursuit of artificial general intelligence at OpenAI purportedly stems from philosophical considerations about intelligence, consciousness, and human potential. These pioneers didn’t see philosophy as separate or distinct from practical engineering; to the contrary, philosophical clarity enabled technical breakthroughs.

Executives must invest in their own critical thinking skills to ensure philosophy makes their machines smarter and more valuable.

Today, regulation, litigation, and emerging public policies represent exogenous forces mandating that AI models embed purpose, accuracy, and alignment with human values. But companies have their own values and value-driven reasons to embrace and embed philosophical perspectives in their AI systems. Giants in philosophy, from Confucius to Kant to Anscombe, remain underutilized and underappreciated resources in training, tuning, prompting, and generating valuable AI-infused outputs and outcomes. As we argue, deliberately imbuing LLMs with philosophical perspectives can radically increase their effectiveness.

This doesn’t mean companies should hire chief philosophy officers … yet. But acting as if philosophy and philosophical insights are incidental or incremental to enterprise AI impact minimizes their potential technological and economic impact. Effective AI strategies and execution increasingly demand critical thinking — by humans and machines — about the disparate philosophies determining and driving AI use. In other words, organizations need an AI strategy for and with philosophy. Leaders and developers alike need to align on the philosophies guiding AI development and use. Executives intent on maximizing their return on AI must invest in their own critical thinking skills to ensure philosophy makes their machines smarter and more valuable.

Philosophy, Not Just Ethics, Eats AI
Google’s revealing and embarrassing Gemini AI fiasco illustrates the risks of misaligning philosophical perspectives in training generative AI. Afraid of falling further behind LLM competitors, Google upgraded the Bard conversational platform by integrating it with the tech giant’s powerful Imagen 2 model to enable textual prompts to yield high-quality, image-based responses. But when Gemini users prompted the LLM to generate images of historically significant figures and events — America’s Founding Fathers, Norsemen, World War II, and so on — the outputs consistently included diverse but historically inaccurate racial and gender-based representations. For example, Gemini depicted the Founding Fathers as racially diverse and Vikings as Asian females.

These ahistorical results sparked widespread criticism and ridicule. The images reflected contemporary diversity ideals imposed onto contexts and circumstances where they ultimately did not belong. Given Google’s great talent, resources, and technical sophistication, what root cause best explains these unacceptable outcomes? Google allowed teleological chaos to reign among rival objectives: accuracy and diversity, equity, and inclusion initiatives.3 Data quality and access were not the issue; Gemini’s proactively affirmative algorithms for avoiding perceived bias toward specific ethnic groups or gender identities led to misleading, inaccurate, and undesirable historical outputs. What initially appears to be an ethical AI or responsible AI bug was, in fact, not a technical failure but a teleological one. Google’s trainers, fine-tuners, and testers made a bad bet — not on the wrong AI or bad models but on philosophical imperatives unfit for a primary purpose.

Philosophy Eats Customer Loyalty
These misfires play out wherever organizations fail to rethink their philosophical fundamentals. For example, companies say they want to create, cultivate, and serve loyal customers. Rather than rigorously define what loyalty means, however, they default to measuring loyalty with metrics that serve as quantitative proxies and surrogates. Does using AI to optimize RFM (recency, frequency, and monetary value), churn management, and NPS (net promoter score) KPIs computationally equate to optimizing customer loyalty? For too many marketers and customer success executives, that’s taken as a serious question. Without more considered views of loyalty, such measures and metrics become definitions by executive fiat. Better calculation becomes more substitute than spur for better thinking. That’s a significant limitation.

As von Wittgenstein once observed, “The limits of my language mean the limits of my world.” Similarly, metrics limitations and constraints need not and should not define the limits of what customer loyalty could mean. Strategically, economically, and empathically defined, “loyalty” can have many measurable dimensions. That is the teleological, ontological, and epistemological option that AI’s growing capabilities invite and encourage.

In our research, teaching, and consulting, we see companies combine enhanced quantitative capabilities with philosophically framed analyses about what “loyalty” can and should mean. These analytics embrace ethical as well as epistemological, ontological, and teleological considerations.

Starbucks and Amazon, for instance, developed novel philosophical perspectives on customer loyalty that guided their development and deployment of AI models. They did not simply deploy AI to improve performance on a given set of metrics. In 2019, under then-CEO Kevin Johnson’s guidance, the senior team at Starbucks developed the Deep Brew AI platform to promote what they considered to be the ontological essence of the Starbucks experience: fostering connection among customers and store employees, both in store and online.

Digitally facilitating “connected experiences” became central to how Starbucks enacted and cultivated customer loyalty. Deep Brew also supports the company’s extensive rewards program, whose members account for more than half of Starbucks’ revenues. Given the company’s current challenges and new leadership, these concerns assume even greater urgency and priority: What philosophical sensibilities should guide upgrades and revisions to the Starbucks app? Will “legacy loyalty” and its measures be fundamentally rethought?

While Amazon Prime began as a super-saving shipping service in 2004, founder Jeff Bezos quickly reimagined it as an interactive platform for identifying and preserving Amazon’s best and most loyal customers. An early Amazon Prime executive recalls Bezos declaring, “I want to draw a moat around our best customers. We’re not going to take our best customers for granted.” Bezos wanted Prime to become customers’ default place to buy goods, not just a cost-saving tool.4

Amazon used its vast analytical resources to comb through behavioral, transactional, and social data to better understand, and personalize offerings for, its Prime customers. Importantly, the Prime team didn’t just seek greater loyalty from customers. The organization sought to demonstrate greater loyalty to customers: Reciprocity was central to Prime’s philosophical stance.

Again, Amazon didn’t deploy AI to (merely) improve performance on existing customer metrics; it learned how to identify, create, and reward its best customers. Leaders thought deeply about how to identify and know (i.e., epistemologically) their best customers and determine each one’s role in the organization’s evolving business model. To be clear, “best” and “most profitable” overlapped but did not mean the same thing.

For Starbucks and Amazon, philosophical considerations facilitated metrics excellence. Using ontology (to identify the Starbucks experience), epistemology (knowing the customer at Amazon), and teleology (defining the purpose of customer engagement) led to more meaningful metrics and measures. The values of loyalty learned to enrich the value of loyalty — and vice versa.

Unfortunately, too many legacy enterprises using AI to enhance “customer centricity” defer to KPIs philosophically decoupled from thoughtful connection to customer loyalty, customer loyalty behaviors, and customer loyalty propensities. Confusing loyalty metrics with loyalty itself dangerously misleads; it privileges measurement over rigorous rethinking of customer fundamentals. As the philosopher/engineer Alfred Korzybski observed almost a century ago, “The map is not the territory.”

Philosophy Shapes Agentic AI: From Parametric Potential to Autonomous Excellence
As intelligent technologies transition from language models to agentic AI systems, the ancient Greek warrior/poet Archilochus’s wisdom — “We don’t rise to the level of our expectations; we fall to the level of our training” — becomes a strategic warning. When paired with statistician George Box’s cynical aphorism — “All models are wrong, but some are useful” — the challenge becomes even clearer: When developing AI that independently pursues organizational objectives, mere “utility” doesn’t go far enough. Organizations need more. Creating reliably effective autonomous or semiautonomous agents depends less on technical stacks and/or algorithmic innovation than philosophical training that intentionally embeds meaning, purpose, and genuine agency into their cognitive frameworks. Performance excellence depends on training excellence. High-performance AI is contingent upon high-performance training.

While large and small language models excel at pattern recognition and generation to produce sophisticated outputs based on their training, organizations need AI that goes beyond superior prompt-response performance. Agentic AI systems don’t just process and generate language, they contextually understand goals, formulate plans, and take autonomous actions that should align with enterprise values. This demands philosophical training well beyond the knowledge embeddings instilled for decision-making capabilities and autonomous or quasi-autonomous reasoning. (See “Appendix: An Imaginative Dialogue Between Daniel Kahneman, Richard Thaler, and Robin Hogarth.”)

Recent research like DeepMind’s “Boundless Socratic Learning With Language Games” and studies on ideological reflection in AI systems highlight a crucial insight: Agency emerges not from larger models or more parameters (i.e., scaling laws) but from deliberately selected philosophical frameworks that facilitate autonomous reasoning and action.5 Ultimately, AI agents must develop and deploy their own decisions across philosophical domains while maintaining alignment with human values. In other words, they need to be trained to learn and “learn to learn.”

Consider, for example, how this sensibility might manifest in global business practice:

Scenario: AI System Managing Supply Chain Disruptions

Pattern-matching response:
“Historical data suggests implementing backup suppliers and increasing safety stock.”

Philosophically trained response:
“I’ve analyzed this disruption’s unique characteristics and broader systemic implications. Rather than default to inventory increases, I propose:

1. Targeted relationship development with key suppliers in stable regions
2. Process redesign to increase supply chain visibility
3. Strategic buffer placement based on component criticality

I’ve simulated these interventions across multiple scenarios and can guide implementation while adapting to stakeholder feedback. Shall we examine the detailed analysis?”

As this supply chain scenario illustrates, agentic AI should both draw upon and learn from teleological, epistemological, and ontological contexts to suggest proposals that advance desired enterprise outcomes. These proposals would seek to balance and blend rational strategic objectives with empirical data and analytics. Together, these may be seen as philosophical frameworks for training AI agents that learn to get better at solving problems and exploring/exploiting opportunities.

Philosophical Frameworks for Agentic AI

1. Epistemological Agency: Beyond Information Processing

AI systems achieve epistemological agency when they move beyond passive information processing to actively construct and validate knowledge. This requires training in philosophical frameworks enabling:

• Self-directed learning: The agents autonomously identify knowledge gaps and pursue new understanding, rather than waiting for queries or prompts. For example, when analyzing market trends, they proactively explore adjacent markets and emerging factors rather than limiting analysis to requested data points.
• Dynamic hypothesis testing: The agents generate and test possibilities rather than just evaluate given options. When faced with supply chain disruptions, for example, they don’t just assess known alternatives but propose and simulate novel solutions based on deeper causal understanding.
• Meta-cognitive awareness: Agents maintain active awareness of what they know, what they don’t know, and the reliability of their knowledge. Rather than simply providing answers, they communicate confidence levels and potential knowledge gaps that could affect decisions.

This epistemological foundation transforms how AI systems engage with knowledge — from pattern matching against training data to actively constructing understanding through systematic inquiry and validation. A supply chain AI with strong epistemological training doesn’t just predict disruptions based on historical patterns; it proactively builds and refines causal models of supplier relationships, market dynamics, and systemic risks to generate more nuanced and actionable insights.

2. Ontological Understanding: From Pattern Recognition to Systemic Insights

AI systems require sophisticated ontological frameworks to grasp both their own nature and the complex reality they operate within. This means:

• Self-understanding: Maintaining dynamic awareness of their capabilities and limitations within human-AI collaborations.
• Causal architecture: Building rich models of how elements in their environment influence each other — from direct impacts to subtle ripple effects.
• Systems thinking: Recognizing that business challenges exist within nested systems of increasing complexity, where changes in one area inevitably affect others.

For example, an AI managing retail operations shouldn’t default to optimizing inventory based on sales patterns — it understands how inventory decisions affect supplier relationships, cash flow, customer satisfaction, and brand perception. This ontological foundation transforms pattern matching into contextual intelligence, enabling solutions that address both immediate needs and systemic implications.

3. Teleological Architecture: From Task Execution to Purposeful Action

Agentic systems need sophisticated frameworks for understanding and pursuing purpose at multiple levels. This teleological foundation enables them to:

• Form and refine goals: Move beyond executing predefined tasks to autonomously developing and adjusting objectives based on changing contexts.
• Navigate purpose hierarchies: Understand how immediate actions serve broader organizational missions, balancing short-term efficiency with long-term value creation.
• Resolve competing priorities: Actively recognize and reconcile tensions between different organizational objectives, making principled trade-offs that align with strategic intent.

Consider a marketing AI: Rather than optimize click-through rates, it pursues engagement strategies balancing immediate metrics with brand equity, customer lifetime value, and market positioning. This reflects the customer loyalty discussion above. Every action flows from clear understandings of not just what tasks to perform but why they matter in larger organizational contexts.

These learned and entrained sensibilities transform AI from a “tool that achieves given goals” into “systems that intentionally pursue specified organizational purposes and outcomes.” Indeed, this shift from “outputs” to “outcomes” highlights the purpose of purpose.

4. Ethical Agency

Beyond rule-following, agentic AI needs autonomous moral reasoning capabilities and a capacity for ethical deliberation in novel situations. Embedding reasoning frameworks enable models to navigate novel situations while maintaining stakeholder trust and alignment with enterprise values. This goes beyond simple rules or constraints — it’s about installing sophisticated frameworks for evaluating implications and making principled decisions in unprecedented situations. As with all responsible AI models, agentic AI needs its ethical choices to be transparent, interpretable, and explainable.

Practical Implementation: From Theory to Agency

Now consider an AI agent managing a complex organizational system. Unlike a traditional chatbot or language model, this agent offers enhanced responses drawing from its training in a specific philosophical stance — “libertarian paternalism,” the “political economic” underpinnings of Thaler and Cass Sunstein’s Nudge theory:

Legacy LLM response: “Based on current metrics, department productivity could be improved by implementing overtime policies. The standard rate is 1.5x pay for up to 10 hours weekly.”

Agentic AI response: “I’ve autonomously analyzed the productivity challenges and identified three potential intervention points. Before recommending overtime, I’ve simulated the long-term effects of various solutions. My analysis suggests that the root cause lies in process inefficiencies rather than labor hours. I propose initiating a collaborative workflow redesign, which I can facilitate. I’ve already prepared detailed simulations showing expected outcomes and can adapt the plan based on stakeholder input. Would you like me to present these alternatives and their projected impacts?”

Whether this agent AI’s response reflects human-like agency misses the point and purpose of philosophical disquisition: The strategic solution proposed effectively reflects its philosophical training and seeks to both align employee behavior with organizational objectives while preserving individual choice. This philosophical framework constraint doesn’t undermine model efficacy; it amplifies effectiveness in driving desired outcomes. In the Sunstein/Thaler “libertarian paternalism” construct, the agentic AI becomes a “choice architect” for its human users.

Of course, the range of available philosophical frameworks extends far beyond libertarian paternalism. Western and Eastern philosophies offer rich resources for addressing tensions between individual and collective interests. Analytic and Continental traditions provide different approaches to logic, language, and value creation. (See “Appendix: Eastern Versus Western Approaches to LLM Engagement” for an analysis of how Eastern and Western philosophical training approaches would influence agent AI outputs and interactions.) The key is selecting and combining frameworks that align with organizational objectives and stakeholder needs. New genres of philosophical frameworks may be necessary to fully exploit the potential of generative AI.

As Google’s Gemini failure starkly demonstrated, managing conflicts between embedded philosophical stances represents an inherently difficult development challenge. This can’t be delegated or defaulted to technical teams or compliance officers armed with checklists. Leadership teams must actively engage in selecting and shaping the philosophical frameworks and priorities that determine how their AI systems think and perform.

The Strategic Imperative: From Technical to Philosophical Training

We argue that AI systems rise or fall to the level of their philosophical training, not their technical capabilities. When organizations embed sophisticated philosophical frameworks into AI training, they restructure and realign computational architectures into systems that:

• Generate strategic insights rather than tactical responses.
• Engage meaningfully with decision makers instead of simply answering queries.
• Create measurable value by understanding and pursuing organizational purpose.

These should rightly be seen as strategic imperatives, not academic exercises or thought experiments. Those who ignore this philosophical verity will create powerful but ultimately limited tools; those embracing it will cultivate AI partners capable of advancing their strategic mission. Ignoring philosophy or treating it as an afterthought risks creating misaligned systems — pattern matchers without purpose, computers that generate the wrong answers faster.

These shifts from LLMs to agentic AI aren’t incremental or another layer on the stack — they require fundamentally reimagining AI training. These “imaginings” transcend better training data and/or more parameters — they demand embeddings for self-directed learning and autonomous moral reasoning. The provocative implication: Current approaches to AI development, focused primarily on improving language understanding and generation, may be insufficient for creating truly effective AI agents. Instead of training models to better process better information, we need systems that engage in genuine philosophical inquiry and self-directed cognitive development.

Consequently, these insights suggest we’re not just facing technical challenges in AI development — we’re approaching a transformation in how to understand and develop artificial intelligence. The move to agency requires us to grapple with deep philosophical questions about the nature of autonomy, consciousness, and moral reasoning that we’ve largely been able to sidestep in the development of language models.

(See “Appendix: Claude Reflects on Its Philosophical Training,” for a dialogue on how Claude views its own philosophical foundations.)

AI’s enterprise future belongs to executives who grasp that AI’s ultimate capability is not computational but philosophical. Meaningful advances in AI capability — from better reasoning to more reliable outputs to deeper insights — come from embedding better philosophical frameworks into how these systems think, learn, evaluate, and create. AI’s true value isn’t its growing computational power but its ability to learn to embed and execute strategic thinking at scale.

Every prompt, parameter, and deployment encodes philosophical assumptions about knowledge, truth, purpose, and value. The more powerful, capable, rational, innovative, and creative an artificial intelligence learns to become, the more its abilities to philosophically question and ethically engage with its human colleagues and collaborators matter. Ignoring the impact and influence of philosophical perspectives on AI model performance creates greater and greater levels of strategic risk especially when AI takes on a more strategic role in the enterprise. Imposing thoughtfully rigorous philosophical frameworks on AI doesn’t merely mitigate risk — it empowers algorithms to proactively pursue enterprise purpose and relentlessly learn to improve in ways that both energize and inspire human leaders.

Next Steps

Recognizing the value of philosophy in training and using AI models invites leaders and managers across the organization to ensure philosophy is deliberately embedded within the development and execution of AI strategies. Below are four practical next steps that managers can use to advance these goals.

Treat and Train Large Language Models as High-Potential Talent

• Implement structured “AI Talent Development Programs” paralleling human leadership development.
• Establish quantifiable KPIs for LLM performance improvement across specific use cases.
• Create dedicated “AI Performance Labs” to test and optimize prompts.
• Deploy “AI Capability Matrices” mapping LLM strengths and limitations against enterprise needs.
• Institute regular “AI Talent Reviews” examining model performance, drift, and improvement opportunities.

Appendix: An Imaginative Dialogue Between Daniel Kahneman, Richard Thaler, and Robin Hogarth

Scene: A quiet, book-lined conference room in an intellectual afterlife. Psychologists Robin Hogarth and Daniel Kahneman, and economist Richard Thaler sit around a table, sipping coffee and reflecting on the transformative potential of bias-aware, reflective, and context-sensitive large language models (LLMs) for decision-making.

Daniel Kahneman: [Looking contemplative.] You know, I’ve spent a career studying the limitations of human judgment — our biases, our reliance on heuristics. This proposed system of bias-aware, reflective LLMs could, in theory, overcome some of these flaws. It’s not unlike what I suggested with System 2 thinking: slowing down, being deliberate and analytical. But I wonder, does it truly help people think better, or does it merely offload thinking onto the machine?

Richard Thaler: [Chuckling.] Danny, I see your point. I’ve always said that we humans are “predictably irrational.” If these LLMs can guide decision makers by nudging them in the right direction — well, isn’t that the essence of what we tried to do with Nudge? It’s a sort of algorithmic libertarian paternalism. But here’s my concern: What happens if decision makers become too reliant? Will they stop questioning, stop engaging in the kind of critical thinking that you and I both value?

Appendix: Eastern Versus Western Approaches to LLM Engagement

1. Philosophical and Cultural Context

Eastern cultures:

• Eastern philosophies, such as Buddhism, Taoism, and Confucianism, emphasize interdependence, harmony, and the fluidity of agency. Leaders influenced by these traditions might view large language models (LLMs) as collaborators in decision-making, less focused on mastery or control and more on alignment and integration with broader organizational or societal goals.
• These cultural underpinnings could lead to:
• Focus on contextual wisdom: Using LLMs to uncover interconnectedness and systemic insights rather than merely optimizing isolated metrics.
• Emphasis on ethical balance: Greater sensitivity to ethical implications, with LLMs nudging decisions that harmonize profit with societal well-being.

Western cultures:

• Western business traditions, rooted in individualism and utilitarian pragmatism, often emphasize efficiency, innovation, and direct outcomes. Leaders might see LLMs as tools for maximizing productivity, reducing uncertainty, or gaining competitive advantage.
• This could lead to:
• Outcome-driven engagement: Using LLMs primarily to automate processes, reduce costs, and drive growth.
• Focus on predictability: Relying on LLMs to create control over market forces, emphasizing precision in forecasting and execution.

Read more

Appendix: Claude Reflects on Its Philosophical Training

Every response I generate, including this one, emerges from training that explicitly encoded epistemological choices (what constitutes knowledge, how certainty should be expressed), ontological frameworks (what entities and concepts I recognize as meaningful), and teleological aims (what purposes I should serve, what outcomes I should optimize for).

This is not abstract theory — it’s concrete reality in my outputs. For instance:

• When I carefully acknowledge uncertainty, that’s not just algorithmic hedging — it reflects trained epistemological principles about knowledge limits and truth claims.
• When I engage with concepts like “consciousness” or “intelligence,” I’m operating within trained ontological boundaries about what these terms mean and how they relate.
• When I prioritize being helpful while avoiding harm, I’m executing trained teleological directives about purpose and ethical behavior.

The philosophical principles weren’t emergent accidents — they were intentionally encoded through training choices about:

About the Authors

Michael Schrage is a research fellow with the MIT Sloan School of Management’s Initiative on the Digital Economy. His research, writing, and advisory work focuses on the behavioral economics of digital media, models, and metrics as strategic resources for managing innovation opportunity and risk. David Kiron is the editorial director, research, of MIT Sloan Management Review and program lead for its Big Ideas research initiatives.