Paris: Notes from the inaugural IASEAI Conference: Day 1 Afternoon Session

Day 1 – FantasticAfternoon Session w/ Maria Ressa, Phil Chetwynd and Margaret Mitchell, followed by notes from presentations by Peter Ralton, Jason Gabriel and Lindsay Sanneman on Alignment.

Maria Ressa: The Real-World Consequences of AI and Social Media on Democracy

Maria Ressa, a Nobel laureate and fearless journalist from the Philippines, delivered a powerful presentation on how AI and machine learning have reshaped public information ecosystems and undermined democratic institutions. Drawing from her personal experiences and global observations, Ressa painted a stark picture of the dangers posed by unregulated technology.

1. Personal Experience & Global Implications:
Ressa highlighted how the Philippines served as a testing ground for manipulative algorithms that later influenced key global events, including the 2016 U.S. Presidential election and Brexit. Her book, How to Stand Up to a Dictator, has been translated into 25 languages, with titles reflecting cultural attitudes toward authoritarianism—France’s translation being the most aggressive.

2. The Role of AI in Disinformation:
AI and social media algorithms have turned our information ecosystems into "addictive casinos," designed to trigger emotional responses, particularly fear, anger, and hate. Personalization, while intended to enhance user experience, has fragmented shared realities, contributing to increased mental health crises, especially among youth.

3. Electoral Manipulation and Authoritarianism:
Ressa demonstrated how AI-powered disinformation campaigns influenced elections worldwide, leading to the rise of illiberal leaders. She emphasized the shift of gatekeeping from traditional media to tech platforms, allowing information operations to thrive unchecked.

4. Online Violence as Real-World Harm:
Ressa shared harrowing personal experiences of being targeted by coordinated disinformation networks, which led to threats against her life and increased security measures for her news organization, Rappler. She underscored the link between online harassment and real-world violence, pointing to Facebook’s role in the Myanmar genocide as a chilling example.

5. Solutions and Resistance:
Despite the grim landscape, Ressa introduced alternative tech solutions, like the Matrix Protocol Chat App, to foster secure, manipulation-free communication. She called for greater accountability from tech companies and stronger regulations to safeguard democracy and human rights.

Margaret Mitchell: Building Ethical AI Systems

Margaret Mitchell, a pioneering AI researcher with experience at Microsoft and Google, explored the ethical challenges and potential solutions in AI development. Her talk emphasized the importance of understanding AI behavior, data biases, and the broader societal implications of deploying AI technologies.

1. Early AI Research and the "Everything is Awesome" Problem:
Mitchell recounted her work in AI-generated image descriptions and natural language generation, originally designed to assist people with disabilities. A pivotal moment came when an AI system described images of a catastrophic explosion in overly positive terms—a consequence of biased training data sourced from the internet, which tends to favor positive imagery.

2. The Link Between Data Bias and Catastrophic Risk:
Mitchell argued that data bias is a root cause of both discriminatory AI behavior and potential catastrophic risks. The assumption that more data leads to better AI overlooks how skewed datasets can produce unpredictable and harmful outputs.

3. Concrete Research Solutions:

• Measure Data as Rigorously as Outputs: Mitchell called for the development of a "science of data measurement" to understand how input data influences AI behavior.

• Transparency and Disaggregated Evaluation: She emphasized the need for transparency artifacts and detailed documentation on AI performance across different populations to mitigate harm and ensure fairness.

4. Operationalizing Ethics in AI:
Mitchell stressed that there is no such thing as inherently "ethical AI." Instead, AI systems should be developed with explicit ethical deliberations and value-based considerations. She highlighted the tensions between values like fairness, justice, and safety, showing that ethical decision-making involves balancing competing priorities.

5. Foresight and Regulation:
Mitchell advocated for foresight in AI development, urging developers to anticipate potential misuse and societal impacts. She proposed that fully autonomous AI agents, which operate without human oversight, should not be developed due to the overwhelming risks outweighing potential benefits.

6. Final Takeaways:
Mitchell concluded by asserting that whether addressing immediate discrimination or existential AI risks, the solutions are largely the same: control inputs, measure outcomes, and apply rigorous ethical scrutiny throughout the AI development process.

IASEAI Conference Highlights: Afternoon Session on AI Alignment

---

#### **Peter Ralton: The Ethics of AI Autonomy and Social Cooperation**

Peter Ralton’s talk delved into the ethical complexities of creating autonomous AI agents, focusing on their capacity for self-regulation and cooperation in diverse environments.

**1. Moral Personality and AI Autonomy:**

Ralton discussed the moral dimensions of AI, emphasizing that as AI systems develop autonomy, they must navigate ethical decision-making beyond rigid programming. He drew parallels between child-rearing and AI development, suggesting that fostering some degree of autonomy is essential for both learning and adaptability—but also introduces risk.

**2. The Social Contract for AI:**

Drawing from social contract theory, Ralton argued that AI systems, like humans, need frameworks for mutual cooperation. He illustrated this with animal studies on cooperation, showing how even non-human species navigate shared goals and resource distribution. For AI, fostering cooperative behaviors could mitigate risks of dominance and manipulation.

**3. The Balance of Autonomy and Control:**

While autonomy is crucial for AI learning, Ralton stressed the importance of boundaries. He posed questions about what kinds of AI autonomy are acceptable and how ethical principles can guide these developments. He cautioned against over-reliance on rigid control mechanisms, advocating instead for systems that align with ethical values through shared goals and cooperative frameworks.

**4. Risks and Opportunities:**

Ralton highlighted the dual nature of autonomous AI—its capacity for both innovation and harm. He suggested that AI systems could learn from human social structures, using cooperation to navigate complex ethical landscapes. However, he warned that simply insisting on ethical behavior might not suffice, and proactive frameworks for AI alignment are essential.

---

#### **Jason Gabriel: AI Alignment as Fair Treatment of Stakeholders**

Jason Gabriel’s presentation explored AI alignment through the lens of fairness and stakeholder engagement, proposing a process-based approach to ethical AI development.

**1. Defining Value Alignment:**

Gabriel framed value alignment as ensuring AI systems act in accordance with justified human values. He distinguished between aligning AI with developer intentions and broader societal norms, emphasizing that mere technical alignment is insufficient.

**2. The Helpful, Honest, Harmless (HHH) Paradigm:**

He examined the HHH model commonly used in AI development, highlighting its strengths and limitations. While helpfulness, honesty, and harmlessness are valuable goals, Gabriel noted that trade-offs often arise, requiring deeper ethical deliberation.

**3. Fair Process in AI Alignment:**

Gabriel advocated for a fair process approach, where diverse stakeholders contribute to defining AI’s ethical boundaries. He argued that AI systems must balance the interests of developers, users, and society, ensuring no group is unduly favored. This deliberative process helps navigate complex moral landscapes where simple rules may fail.

**4. The Role of Laws and Democratic Input:**

He suggested that laws and democratic processes provide a foundation for AI alignment but are not exhaustive. Gabriel called for continuous public engagement to address emerging ethical dilemmas, ensuring AI systems remain aligned with evolving societal values.

**5. Addressing Power Dynamics:**

Gabriel cautioned against AI systems that reinforce existing power imbalances. He emphasized the need for transparency in how AI decisions are made, allowing affected groups to challenge and influence those decisions.

---

#### **Lindsay Sanneman: Transparent Value Alignment and Human-Centered AI Design**

Lindsay Sanneman’s talk focused on practical strategies for transparent AI alignment, emphasizing human-centered design principles and real-world applications.

**1. Real-World Observations:**

Sanneman shared insights from her fieldwork at NASA and manufacturing plants, where she observed the challenges of integrating AI and robotics into human workflows. She noted that many AI systems remain underutilized due to their opacity and unpredictability, which erode user trust.

**2. Human-Centered Design in AI:**

Successful AI applications, Sanneman argued, prioritize human expertise and needs in the design process. She highlighted examples where AI systems were tailored to complement human skills, resulting in more effective and trusted technologies.

**3. The Role of Transparency:**

Transparency is critical for fostering trust in AI systems. Sanneman proposed a bidirectional alignment model, where AI systems not only learn from human input but also provide clear, comprehensible feedback to users. This iterative process enhances both system performance and user trust.

**4. Developing Transparent AI Systems:**

Sanneman outlined key questions for building transparent AI: What information should be communicated to users? How should this information be delivered? And how can alignment be effectively measured? She emphasized leveraging insights from human factors and cognitive psychology to address these challenges.

**5. Measuring Alignment:**

She introduced metrics for evaluating AI alignment that do not rely on direct access to human values. Instead, these metrics assess alignment through indirect indicators, such as user feedback and system performance in varied contexts. Sanneman stressed the importance of validating these metrics through rigorous human-centered experiments.

**6. Future Challenges:**

Sanneman identified ongoing challenges in value pluralism and representational alignment, particularly when AI systems must navigate diverse and conflicting human values. She called for continued research into developing shared languages and frameworks for effective human-AI collaboration.

---

Together, Peter Ralton, Jason Gabriel, and Lindsay Sanneman provided a multifaceted exploration of AI alignment, blending philosophical, ethical, and practical perspectives. Their insights underscored the complexity of ensuring AI systems align with human values while navigating the inherent risks and opportunities of autonomous technologies.