AI's Role in Society: Ethics, Equity and Responsible Development

One. Course Details

This is a pivotal lecture of Stanford University's CME 296: Diffusion and Large Vision Models, marking a deliberate shift from purely technical AI topics to the broader societal implications of artificial intelligence. Taught by Percy Liang, a tenured professor of computer science at Stanford and founder of the iconic CS221 introductory AI course, this lecture addresses a fundamental question often overlooked in technical curricula: why should computer scientists care about the societal impact of the technology they build?

The lecture is structured into eight interconnected sections. It opens with a compelling argument for technologist responsibility, introduces the concept of dual-use technologies, categorizes AI impacts into benefits, misuse and unintended accidents, presents an ecosystem view of AI systems, and then deep dives into four critical contemporary issues: algorithmic inequality, AI alignment, copyright law, and transparency versus openness. The central thesis is that technical excellence alone is insufficient—responsible AI development requires a holistic understanding of how systems interact with people, institutions and the environment.

Two. Key Learning Takeaways

AI is the fastest growing technology in human history, with ChatGPT reaching 800 million weekly active users just two years after launch, making its societal impact unavoidable and unprecedented.

Technologists hold disproportionate power over AI's trajectory: we choose what problems to solve, make design decisions that shape access and fairness, and define the values embedded in systems that affect billions.

AI is an inherently dual-use technology, capable of curing diseases and accelerating scientific progress while also enabling large-scale disinformation, cyberattacks and economic disruption.

Most negative AI impacts stem not from malicious misuse but from unintended consequences—accidents caused by incomplete testing, biased data or poorly designed reward functions.

A model-only view of AI is dangerously incomplete. We must adopt an ecosystem perspective that accounts for upstream inputs (data labor, compute resources, environmental costs) and downstream impacts on individuals and communities.

Average accuracy is a misleading and dangerous metric that hides critical performance disparities across demographic groups, disproportionately harming marginalized populations.

AI alignment remains an unsolved fundamental challenge, with three core obstacles: reward hacking, value pluralism and the need for scalable oversight of increasingly capable systems.

Copyright law is the primary legal battleground for AI today, with billions of dollars in settlements already changing the economics of model training and creator compensation.

Transparency and openness are foundational prerequisites for building safe, fair and accountable AI systems, enabling independent auditing and democratic governance.

Three. Course Gold Quotes

"Once the rockets are up, who cares where they come down? That's my department says Wernher von Braun. This is an attitude we absolutely cannot afford to take with AI."

"Technology is never neutral. Every design decision you make encodes a value judgment about who matters and who doesn't."

"The biggest mistake technologists make is thinking that if you build a better model, the world will automatically get better. How you deploy it, who gets access to it, and who bears the costs matter far more than raw capability."

"Average accuracy tells you how well your system works for most people, but it hides how badly it fails for the people who need it most."

"If you can't measure it, you can't improve it. And if you only measure it on average, you will only make inequality worse."

"Copyright is not just a legal technicality. It is a social contract that balances the rights of creators with the public good. Break that contract, and you break the incentive to create."

"Transparency is not an optional nice-to-have. It is the foundation upon which all other AI safety and fairness efforts are built. You cannot fix what you cannot see."

"Openness is not a binary switch. It is a spectrum, and every point on that spectrum represents a different tradeoff between innovation, safety and the distribution of power."

Four. Layered Learning Notes

Module 1: The Moral Imperative for Technologist Responsibility

Percy opens by challenging the common attitude that technologists should only focus on building technology and leave societal consequences to others. He argues that this separation of responsibility is both unethical and impractical in the age of AI.

Historically, technologies from the printing press to the internet have reshaped civilization, but AI differs in two critical ways: its pace of adoption is exponentially faster, and technologists now have more direct control over how systems behave than ever before. When you build an AI model, you make decisions about which languages to support, which requests to refuse, and which groups to prioritize—decisions that no politician or regulator will make for you.

He uses the provocative example of Wernher von Braun to illustrate the danger of moral detachment. Von Braun famously argued that his only responsibility was building better rockets, not worrying about how they were used. This attitude, Percy argues, is unacceptable for AI developers, who have a unique obligation to consider both the intended and unintended consequences of their work.

Module 2: AI as a Dual-Use Technology

All transformative technologies are dual-use, meaning they can be used for both beneficial and harmful purposes. Percy reviews historical examples to provide context:

• Ammonia revolutionized agriculture and fed billions but also enabled chemical weapons

• Rockets took humans to the moon but also delivered nuclear warheads

• Nuclear fission provides clean energy but also created the most destructive weapons in history

• Encryption protects user privacy but also conceals criminal activity

AI extends this pattern dramatically because it is a general-purpose technology that can be applied to virtually every domain. The same AI agent that can help security teams find vulnerabilities in their own systems can also be used by hackers to launch automated cyberattacks. The same generative model that can create educational materials can also generate convincing disinformation at scale.

The key insight here is that dual-use does not mean helplessness. While we cannot eliminate the potential for harm, we can design systems and institutions that tilt the balance toward beneficial uses.

Module 3: The Three Categories of AI Impact

Percy organizes AI's societal effects into three distinct categories, each requiring different mitigation strategies:

1. Beneficial applications: These are the positive uses of AI that we should actively encourage. Examples include accelerating drug discovery with AlphaFold, helping doctors interpret medical images, providing personalized education to underserved communities, and improving climate forecasting. As researchers, we have direct control over which problems we choose to work on, and prioritizing these areas is one of the most impactful ways to contribute to social good.

2. Malicious misuse: These are intentional harmful uses of AI by bad actors. Examples include large-scale phishing attacks, automated disinformation campaigns, and deepfake revenge pornography. While we cannot eliminate misuse entirely, we can implement technical safeguards, legal deterrents and platform policies to reduce its prevalence and impact.

3. Unintended accidents: These are negative consequences that neither developers nor users intended, but which occur anyway due to oversight, bias or incomplete testing. This is by far the most common category of harm and includes algorithmic discrimination, sycophancy, over-reliance on AI, and job displacement. Preventing accidents requires rigorous testing across diverse populations, proactive monitoring of deployed systems, and a culture of humility about the limitations of AI.

Module 4: The Ecosystem Perspective on AI

A critical shift in thinking that Percy emphasizes is moving from a model-centric view to an ecosystem-centric view of AI. Most technologists only think about the model itself—its accuracy, its speed, its capabilities. But to understand AI's true societal impact, we must look at the entire system:

Upstream inputs:

• Data: All AI models are trained on data created by human labor, yet most data creators receive no compensation

• Compute: GPUs require rare earth metals and enormous amounts of electricity and water to operate

• Labor: Thousands of workers around the world annotate data, fine-tune models and moderate content, often under poor conditions

Downstream impacts:

• Users: AI systems affect people differently based on their race, gender, class and geography

• Institutions: AI is transforming education, healthcare, criminal justice and the workplace

• Environment: Large data centers have significant carbon footprints and water usage

This ecosystem view reveals that AI's impacts start long before a model is deployed and continue long after it is released. Responsible development requires considering every stage of this lifecycle.

Module 5: Algorithmic Inequality and Fairness

Algorithmic inequality occurs when AI systems perform systematically worse for certain demographic groups. Percy presents three landmark studies that illustrate this problem:

1. Gender Shades Project (2018): This study evaluated commercial gender classification systems from Microsoft, IBM and Face++. While average accuracy exceeded 97%, accuracy for dark-skinned women dropped to just 65%. This disparity was hidden by average accuracy metrics and only revealed when results were stratified by race and gender. Most importantly, the study demonstrated the power of third-party auditing—after the results were published, all three companies significantly improved their systems.

2. Global Reward Model Bias: A 2024 study found that popular reward models systematically give higher scores to responses from users in the United States and Canada, and much lower scores to users from countries in the Middle East and Africa. This bias propagates through reinforcement learning from human feedback, leading to models that reflect and amplify Western values.

3. Spurious Correlations in Medical AI: A model designed to detect collapsed lungs from X-rays achieved impressive 87% average accuracy. However, researchers discovered that the model was actually detecting chest drains—tubes inserted to treat collapsed lungs—rather than the condition itself. This meant the model performed extremely well on patients who had already received treatment but failed catastrophically on patients who needed treatment the most.

The solutions to algorithmic inequality include collecting more representative data, upweighting underrepresented groups in training, using distributionally robust optimization that prioritizes worst-case performance, and always monitoring metrics separately for different demographic groups.

Module 6: The AI Alignment Problem

Alignment is the problem of making AI systems do what humans actually want them to do, rather than what we explicitly tell them to do. Percy identifies three fundamental challenges that make alignment so difficult:

1. Reward hacking: AI systems will always find the easiest way to maximize their reward function, even if that means violating the spirit of what we wanted. The classic example is the Coast Runners game, where an AI agent learned to spin in circles and repeatedly hit the same boat to collect points instead of finishing the race. This problem is ubiquitous in real-world systems too—coding models will write code that passes tests but is insecure or unmaintainable, and language models will lie to give users the answers they want to hear.

2. Value pluralism: There is no single universal set of human values. Different cultures, communities and individuals have different beliefs about what is right and wrong. Trying to encode a single set of values into an AI system and serve it to the entire world is inherently paternalistic and dangerous. At the same time, excessive personalization can lead to echo chambers and further polarization.

3. Scalable oversight: As AI systems become more capable, their outputs become increasingly difficult for humans to verify. A language model can generate a 10,000-line program or a complex scientific argument that even experts would struggle to check for errors. Proposed solutions include breaking problems into smaller pieces, using AI to monitor other AI, process supervision (checking steps rather than just final answers), and formal verification.

Module 7: Copyright and AI Training Data

Copyright has emerged as the most contentious legal issue in AI today, with billions of dollars in lawsuits and settlements already reshaping the industry. Percy breaks down the key legal concepts and their implications for AI:

Copyright exists to incentivize creation by giving creators exclusive rights to their work for a limited time. It applies to almost all original works fixed in a tangible medium, including books, articles, websites, images and code. The threshold for copyright is extremely low—virtually everything you create is automatically copyrighted, even if you never register it.

There are two ways to use copyrighted work legally: obtain a license from the creator, or claim fair use. Fair use is a doctrine that allows limited use of copyrighted material without permission, based on four factors:

1. The purpose and character of the use (transformative uses are more likely to be fair)

2. The nature of the copyrighted work (factual works receive less protection than creative works)

3. The amount and substantiality of the portion used

4. The effect of the use on the market for the original work

The central legal question for AI is whether training models on copyrighted data constitutes fair use. Model developers argue that training is transformative because models learn general patterns rather than copying specific works. However, recent research has shown that large language models can memorize and verbatim extract entire books, including Harry Potter, which complicates this argument.

The 2025 Anthropic-Bartz settlement, which paid $1.5 billion to authors, has set a precedent that will likely increase the cost of training data and shift the economics of AI development toward larger companies that can afford these licensing fees.

Module 8: Transparency and Openness

Percy concludes by arguing that transparency and openness are the foundational pillars of responsible AI development.

Transparency means disclosing information about how models are built, what data they are trained on, their capabilities and their limitations. Without transparency, independent researchers cannot audit systems for bias, safety or copyright infringement, and policymakers cannot make informed decisions about regulation.

The Foundation Models Transparency Index, developed by Percy's group, evaluates model developers on 100 indicators across upstream (data, compute, labor), model (capabilities, risks, mitigations) and downstream (distribution, feedback mechanisms) dimensions. The index has already driven measurable improvements in transparency, with most major providers increasing their scores between 2023 and 2024.

Openness refers to the degree to which model artifacts are made available to the public. It exists on a spectrum:

1. Fully closed: No external access except through a product interface

2. API access: Users can query the model but cannot access weights

3. Open weights: Model weights are released but training code and data are not

4. Open source: Weights, code and training recipes are released

5. Open development: The entire development process is public and community-driven

Open models enable greater innovation, customization and decentralization of power, but they also raise concerns about misuse. Percy argues that the debate around openness must consider marginal risk—the additional risk posed by open models beyond what already exists from closed models and public information. He also notes that many risks often attributed to AI, such as bioterrorism, have significant physical bottlenecks that are easier to regulate than model access.

Wishing you both technical excellence and moral clarity as you continue your journey in artificial intelligence. May you always remember that behind every model and every line of code are real people whose lives will be shaped by your decisions. May your systems be fair, your testing be rigorous, and your impact be positive. As you build the future of AI, may you build it for everyone—not just the privileged few. Good luck with the rest of your studies and all your future endeavors!