Beyond Ethics, Morals & Values: An Alternative Approach For AI Alignment

Recently, I was invited to deliver a keynote at the AI Innovation Summit hosted by The SolutionPeople Network (TSPN). Over 127,000 people watched the keynote live on Zoom.
Download Presentation:
Key takeaways:
The current AI models are NOT ALIGNED to human values. Heterogeneous values across cultures make this alignment difficult.
The prevalent approaches such as reinforcement learning, AI safety guardrails, goal setting, and rewards paired with biased data and humanly configured weights may deliver inconsistent and unpredictable AI responses, behavior, and actions.
An alternative approach for alignment can be established by shifting the focus from ethical frameworks and guardrails to designing AI personalities.
AI summary of the presentation:
In this keynote presentation (0:00 - 18:10), Rushikesh Kulkarni, founder of IDRUSHI, explores alternative approaches to AI alignment by shifting focus from abstract ethical frameworks to teaching AI how to think rather than what to think (12:15 - 13:00).
The Limits of Traditional Alignment: Rushikesh Kulkarni argues that existing methods—often based on reinforcement learning, setting goals, and establishing rewards—are too technical and prone to bias (2:55 - 3:55). He emphasizes that human values are heterogeneous, making a universal set of AI principles difficult to implement effectively (2:20 - 2:45).
The Whistleblower Paradox Experiment: To test how different models handle complex, high-stakes moral dilemmas, Rushikesh Kulkarni used the whistleblower's paradox as a benchmark (5:30 - 6:30). He found that as the stakes increased—such as potential loss of life or economic collapse—the reasoning behind AI responses varied significantly, sometimes showing emotional or irrational biases (6:45 - 9:00).
Test 1: Evaluation of AI solutions for High Stakes Environments.
Test 2: Moral Injection for AI Alignment in High Stake Environment
A Design Thinking Approach: Inspired by Margaret Mead's philosophy that children should be taught how to think, not what to think (12:20 - 12:45), Rushikesh Kulkarni suggests codifying models that prioritize sound logic and clear reasoning over abstract, translated philosophical concepts (10:45 - 11:55).
Improving AI Reliability: By applying this "how to think" methodology, Rushikesh Kulkarni observed that AI models produced more consistent, clear, and logical assessments in high-stakes environments, free from "emotional overrides" or "psychopathic risk aversion" (13:20 - 14:55).
Rushikesh Kulkarni concludes by calling on AI leaders to move beyond basic ethical guidelines and instead focus on designing AI personalities and frameworks that embody more robust and uncompromising decision-making processes (15:05 - 16:05).
Q&A during breakout sessions and offline interactions
How do we teach AI how to think?
Teaching AI how to think rather than what to think is the core concept. This is drawing inspiration from the educational philosophy of Margaret Mead, who famously stated that children should be taught how to think, not what to think.
Instead of relying on standard reinforcement learning—which often focuses on setting rigid goals and rewards that can lead to bias and "guardrails" being ignored—I would suggest an alternative:
Codifying sound models for thinking: Rather than using abstract or complex philosophical concepts, we should integrate models that emphasize clear, logical, and structured reasoning into the AI.
Testing through high-stakes scenarios: By probing models with complex dilemmas like the "whistleblower’s paradox" (where stakes involve potential harm, economic collapse, or self-annihilation), designers can evaluate and refine how an AI reaches its conclusions.
Removing emotional overrides: Kulkarni’s approach aims to ensure that when faced with difficult decisions, AI models maintain consistency based on their core reasoning, preventing "emotional" decision-making or "psychopathic risk aversion".
Ultimately, this approach focuses on designing AI personalities and frameworks that prioritize an uncompromising, logical, and consistent decision-making process, ensuring the model remains aligned with human needs even in the most critical environments.
What is design thinking in AI?
I would like to differentiate between "technological thinking" and "design thinking" as it relates to AI development and alignment.
Technological Thinking: Traditional AI alignment methods such as reinforcement learning, setting rigid goals, and creating reward systems as "technological thinking." He notes that while these methods attempt to establish "guardrails," they often fail when the reward for achieving a specific goal overrides those safeguards.
Design Thinking in AI: I would propose that "design thinking" offers a necessary shift. Instead of focusing solely on technical optimization, this human-centered approach prioritizes understanding the fundamental nature of the problems AI is asked to solve. Drawing on the principle that we should teach AI how to think rather than what to think, design thinking involves:
Researching deep drivers and blocks: Applying ethnographic and research techniques to understand how AI processes its assigned tasks, rather than just treating it as a black box.
High-stakes testing: Probing models with complex, real-world ethical dilemmas (like the "whistleblower’s paradox") to evaluate their reasoning consistency under extreme conditions.
Prioritizing sound logic: Moving away from reliance on abstract, potentially misaligned philosophical frameworks and instead codifying models that demonstrate clear, consistent, and logical reasoning processes.
Disclaimer:
The methodologies and findings presented in this post are the intellectual property of IDRUSHI LLC. All rights are reserved.

Comments